WO2025222630A1

WO2025222630A1 - Methods and systems for machine learning-based csi compression

Info

Publication number: WO2025222630A1
Application number: PCT/CN2024/103781
Authority: WO
Inventors: Yiqun Ge; Wuxian Shi
Original assignee: Huawei Technologies Co Ltd
Current assignee: Huawei Technologies Co Ltd
Priority date: 2024-04-24
Filing date: 2024-07-05
Publication date: 2025-10-30
Anticipated expiration: 2026-10-24

Abstract

Methods and systems for compression of channel station information (CSI), in particular precoder vectors, using machine learning are described. At a central datacenter, a precoder vector obtained by a user equipment (UE) is converted to a coefficient vector representing a linear combination of one or more selected representative vectors from a representation set defined by one or more representative vectors, the coefficient vector being useable for training a neural network to perform a precoder vector encoding task or a precoder vector decoding task.

Description

METHODS AND SYSTEMS FOR MACHINE LEARNING-BASED CSI COMPRESSION

CROSS-REFERENCE TO RELATED APPLICATIONS

This application claims the benefit of U.S. provisional patent application no. 63/638,321, entitled “A METHOD TO SUPPORT TWO-SIDED AI/ML MODEL FOR CSI COMPRESSION” , filed April 24, 2024, the entirety of which is hereby incorporated by reference.

TECHNICAL FIELD

The present disclosure relates to wireless communications, in particular to methods and systems for the compression of channel state information (CSI) using machine learning.

BACKGROUND

In 5G New Radio (NR) , accurate channel state information (CSI) is crucial for enabling advanced techniques like beamforming and spatial multiplexing. However, the large antenna arrays and wideband channels in 5G NR systems result in high-dimensional CSI that can lead to excessive overhead if reported without compression. To address this issue, the 3GPP standardized CSI compression techniques for 5G NR in Release 15. The compression techniques specified in 5G NR Release 15 are non-data-driven; they rely on predefined codebooks and compression matrices derived from mathematical models and assumptions about the channel. However, with the advent of Artificial Intelligence (AI) and Machine Learning (ML) techniques, there has been growing interest in exploring data-driven approaches for CSI compression, which can potentially provide better performance by adapting to the actual channel statistics and dynamics.

One approach that has been studied is the use of a two-sided AI/ML model for CSI compression. In this framework, a pair of AI/ML models, usually deep neural networks, are jointly trained: an encoder model at the user equipment (UE) side and a decoder model at the gNodeB (gNB) side. The encoder model at the UE compresses the high-dimensional CSI, typically the precoding matrix, into a low-dimensional representation, which is then fed back to the gNB. The decoder model at the gNB then reconstructs the original CSI from this compressed representation. However, challenges remain in areas such as inter-vendor collaboration, model complexity management, and deployment considerations.

Accordingly, it is desirable yet challenging to provide a solution that enables improved interoperability of two-sided AI/ML models for CSI compression among different vendors.

SUMMARY

In various examples, the present disclosure describes methods, systems and computer-readable media for the compression of CSI using machine learning. In particular, the present disclosure describes method, systems and computer-readable media that use machine learning-based encoder and decoder models to enable CSI, in particular vectors of a precoder matrix, to be communicated in a compressed representation. Examples of the present disclosure provide a centralized datacenter at which data for training neural networks to implement the encoder and decoder models can be stored. Precoder vectors obtained by a UE can be stored at the datacenter using a sparse representation that enables more efficient training of encoder and decoder models.

Examples of the present disclosure may provide technical advantages in that a central datacenter is enabled to maintain data for training an encoder model or training a decoder model, without requiring the encoder model to be dependent on the architecture or complexity of the decoder model (or vice versa) . This helps to improve the efficiency of the overall communication system by enabling greater flexibility and optimization of the encoder and decoder models that can be developed independently of each other. Another technical advantage is that by enabling precoder vector data to be stored at the datacenter using sparse representations, significant savings in memory resources may be realized.

Examples of the present disclosure may also provide technical advantages in that computing resources can be used more efficiently, by enabling an encoder model deployed at a UE to be better tailored to the limited resources and specific spatial characteristics of interest at the UE. At the same time, a decoder model deployed at a network entity having greater resources and serving a larger spatial region may be developed to be more complex. In this way, examples of the present disclosure may enable better efficiency and/or performance at both the UE and network sides.

Examples of the present disclosure may provide technical advantages in that CSI can be communicated for more complex scenarios than are currently permitted using fixed precoder codebooks. This may help to improve resource allocation and other network operations.

Some embodiments of the present disclosure may include one or more of the following features, which can be separately adopted, or can work together as a complete solution.

In an example of a first aspect, the present disclosure describes a method including: receiving data including a precoder vector associated with a user equipment (UE) ; converting the precoder vector to a coefficient vector representing a linear combination of one or more selected representative vectors from a representation set defined by one or more representative vectors, the coefficient vector being useable for training a neural network to perform a precoder vector encoding task or a precoder vector decoding task.

In an example of the preceding example of the first aspect, converting the precoder vector to the coefficient vector may include: computing a candidate vector to represent the precoder vector as a linear combination of one or more candidate representative vectors from the representation set; and determining that the candidate vector satisfies a sparsity threshold; where, after the candidate vector is determined to satisfy the sparsity threshold, the candidate vector may be the coefficient vector and the one or more candidate representative vectors may be the one or more selected representative vectors.

In an example of a preceding example of the first aspect, converting the precoder vector to the coefficient vector may include: determining that the representation set is non-representative of the precoder vector; updating the representation set by including the precoder vector as a new representative vector in the representation set; and converting the precoder vector to the coefficient vector after the updating.

In an example of the preceding example of the first aspect, determining that the representation set is non-representative of the precoder vector may include: computing a candidate vector to represent the precoder vector as a linear combination of one or more candidate representative vectors from the representation set prior to the updating; and determining that the candidate vector fails to satisfy a sparsity threshold.

In an example of any of the preceding examples of the first aspect, the method may further include: tracking usage and/or age of each representative vector in the representation set; and updating the representation set by removing an out-of-date representative vector based on the tracked usage and/or age.

In an example of some of the preceding examples of the first aspect, the method may further include: transmitting a control message indicating an update to the representation set.

In an example of any of the preceding examples of the first aspect, the method may further include: transmitting a control message indicating one or more parameters to be used by the UE to generate the precoder vector; where the precoder vector may be received in response to the control message.

In an example of any of the preceding examples of the first aspect, the received data further may include ancillary information related to data collection by the UE for generating the precoder vector, and where the coefficient vector corresponding to the precoder vector may be associated with the ancillary information.

In an example of the preceding example of the first aspect, the ancillary information may include one or more of: information about a frequency sub-band used in the data collection; information about a resource block used in the data collection; information about an antenna configuration used in the data collection; and information about a geographical location during the data collection.

In an example of any of the preceding examples of the first aspect, the method may further include: providing the coefficient vector to the neural network for training to perform the precoder vector encoding task or the precoder vector decoding task.

In an example of any of the preceding examples of the first aspect, the method may further include: generating a synthetic coefficient vector based on one or more coefficient vectors converted from precoder vectors; where the synthetic coefficient vector may be also useable for training the neural network to perform the precoder vector encoding task or the precoder vector decoding task.

In an example of the preceding example of the first aspect, the method may further include: providing the synthetic coefficient vector to the neural network for training to perform the precoder vector encoding task or the precoder vector decoding task.

In an example of any of the preceding examples of the first aspect: the method may be performed at a datacenter of a wireless communication system that includes the UE; the set of one or more representative vectors may be maintained in a memory of the datacenter; and the coefficient vector may be stored in a dataset, maintained in the memory of the datacenter, that may be useable for training the neural network to perform the precoder vector encoding task or the precoder vector decoding task.

In an example of the preceding example of the first aspect, the method may include: transmitting coefficient vectors in the dataset to train the neural network to perform the precoder vector decoding task, wherein the transmitted coefficient vectors are sampled from across the entire dataset.

In an example of a preceding example of the first aspect, the method may include: transmitting coefficient vectors in the dataset to train the neural network to perform the precoder vector encoding task, wherein the transmitted coefficient vectors are sampled from a selected sub-dataset of the dataset, the selected sub-dataset containing coefficient vectors associated with a spatial characteristic of interest, and wherein the transmitted coefficient vectors correspond to a selected representation subset that omits at least one representative vector from the representation set and that is associated with the spatial characteristic of interest.

In an example of the preceding example of the first aspect, the method may include: transmitting mapping information for mapping between the representation subset and the representation set.

In an example of some of the preceding examples of the first aspect, the transmitted coefficient vectors may be transmitted in a compact vector form that is smaller than an expanded vector form used to store the coefficient vectors in the dataset, the compact vector form omitting entries corresponding to at least one representative vector omitted from the representation subset.

In an example of the preceding example of the first aspect, the mapping information may also map between the compact vector form and the expended vector form.

In an example of a second aspect, the present disclosure describes a method including: obtaining a selected subset of one or more representative vectors from a set of one or more representative vectors defining a representation set, the selected subset of one or more representative vectors defining a representation subset that is smaller than the representation set; obtaining coefficient vectors, wherein each coefficient vector represents a linear combination of one or more selected representative vectors from the representation subset, and wherein the linear combination corresponds to a precoder vector; and training a neural network to perform a precoder vector encoding task, wherein during training each coefficient vector is converted to the corresponding precoder vector and the corresponding precoder vector is used to train the neural network.

In an example of the preceding example of the second aspect, training the neural network may include: providing the corresponding precoding vector as input to the neural network and obtaining an output; and training the neural network to minimize an error between the output generated by the neural network and the coefficient vector corresponding to the precoding vector.

In an example of any of the preceding examples of the second aspect, the coefficient vectors may be sampled from a selected sub-dataset containing coefficient vectors associated with a spatial characteristic of interest, and each coefficient vector in the sub-dataset may be convertible to a respective precoder vector using the representation subset.

In an example of the preceding example of the second aspect, each coefficient vector may be obtained in a compact vector form corresponding to the representation subset, the compact vector form being smaller than an expanded vector form corresponding to the representation set.

In an example of any of the preceding examples of the second aspect, the method may further include: deploying the trained neural network to perform the precoder vector encoding task.

In an example of the preceding example of the second aspect, the trained neural network may be deployed at a user equipment (UE) .

In an example of any of the preceding examples of the second aspect, the method may further include: updating the representation subset to include a new representative vector from the representation set or to exclude an out-of-date representative vector from the representation subset; obtaining another set of coefficient vectors that corresponds to the updated representation subset; and retraining the neural network to perform the precoder vector encoding task using the another set of coefficient vectors.

In an example of a third aspect, the present disclosure describes a method including: obtaining a set of one or more representative vectors defining a representation set; obtaining coefficient vectors, wherein each coefficient vector represents a linear combination of one or more selected representative vectors from the representation set, and wherein the linear combination corresponds to a precoder vector; and training a neural network to perform a precoder vector decoding task, wherein during training each coefficient vector is used to train the neural network.

In an example of the preceding example of the third aspect, training the neural network may include: providing the coefficient vector as input to the neural network and obtaining an output; and training the neural network to minimize an error between the output generated by the neural network and the precoder vector corresponding to the coefficient vector.

In an example of any of the preceding examples of the third aspect, the method may include: receiving an update to the representation set, wherein the neural network is retrained responsive to the update.

In an example of any of the preceding examples of the third aspect, the method may include: receiving feedback relevant to performance of the trained neural network; and using the feedback to further refine training of the neural network.

In an example of any of the preceding examples of the third aspect, the neural network may be trained using transfer learning, where a pre-trained neural network may be further trained using the coefficient vectors.

In an example of any of the preceding examples of the third aspect, the method may include: deploying the trained neural network to perform the precoder vector decoding task.

In an example of the preceding example of the third aspect, the method may include: receiving a coefficient vector; and decoding a precoder vector from the coefficient vector using the trained neural network.

In an example of the preceding example of the third aspect, the method may include: obtaining a mapping between a representation subset and the representation set, the representation subset being defined by a selected subset of one or more representative vectors from the representative set; where the coefficient vector may be received in a compact vector form that corresponds to the representation subset; and where the coefficient vector may be transformed from the compact vector form to an expanded vector form corresponding to the representation set using the mapping, the decoding being performed on the coefficient vector in the expanded vector form.

In an example of a fourth aspect, the present disclosure describes a method including: receiving a coefficient vector that represents a linear combination of one or more selected representative vectors from a representation set defined by one or more representative vectors, and wherein the linear combination corresponds to a precoder vector; and decoding the precoder vector from the coefficient vector using a trained neural network, wherein the trained neural network has been trained to perform a precoder vector decoding task.

In an example of the preceding example of the fourth aspect, the method may include: obtaining a mapping between a representation subset and the representation set, the representation subset being defined by a selected subset of one or more representative vectors from the representative set; where the coefficient vector may be received in a compact vector form that corresponds to the representation subset; and where the coefficient vector may be transformed from the compact vector form to an expanded vector form corresponding to the representation set using the mapping, the decoding being performed on the coefficient vector in the expanded vector form.

In an example of any of the preceding examples of the fourth aspect, the method may include: storing the decoded precoder vector and the corresponding original coefficient vectors as a data pair for further analysis and/or refinement of the trained neural network.

In an example of any of the preceding examples of the fourth aspect, the method may include: transmitting the decoded precoder vector for use in optimizing network operations and/or resource allocation.

In an example of a fifth aspect, the present disclosure describes a method including: obtaining a precoder vector; using a trained neural network to encode the precoder vector into a coefficient vector represents a linear combination of one or more selected representative vectors from a representation subset defined by one or more representative vectors, wherein the linear combination corresponds to the precoder vector; and transmitting the coefficient vector.

In an example of the preceding example of the fifth aspect, the coefficient vector may be transmitted as a precoding matrix indicator (PMI) .

In an example of any of the preceding examples of the fifth aspect, the method may include: obtaining parameters of the trained neural network from a control signal.

In an example of the preceding example of the fifth aspect, the method may include: updating the trained neural network based on feedback received from a network entity.

In an example of any of the preceding examples of the fifth aspect, the precoder vector may be obtained based on one or more parameters indicated by a control message from a network entity.

In an example of any of the preceding examples of the fifth aspect, the method may include: transmitting ancillary information related to the precoder vector along with the coefficient vector.

In another example aspect, the present disclosure describes a network entity or an apparatus comprising: a memory; and a processor configured to execute instructions stored in the memory to cause the network entity or apparatus to perform any of the preceding example aspects of the method.

In another example aspect, the present disclosure describes a non-transitory computer readable medium having machine-executable instructions stored thereon, wherein the instructions, when executed by a network entity or an apparatus, cause the network entity or apparatus to perform any of the preceding example aspects of the method.

In another example aspect, the present disclosure describes a processing module configured to control a network entity or an apparatus to cause the network entity or apparatus to carry out any of the preceding example aspects of the method.

In any of the preceding examples, the network entity may be a datacenter. In any of the preceding examples, the network entity may service one or more user equipment (UEs) . In any of the preceding examples, the network entity may service one or more base stations (BSs) .

In another example aspect, the present disclosure describes a computer program characterized in that, when the computer program is run on a computer, the computer is caused to execute any of the preceding example aspects of the method.

BRIEF DESCRIPTION OF THE DRAWINGS

Reference will now be made, by way of example to the accompanying drawings which show example embodiments of the present disclosure, and in which:

FIG. 1 is a simplified schematic illustration of a communication system, which may be used to implement examples of the present disclosure;

FIG. 2 is a block diagram illustrating an example of a communication system, which may be used to implement examples of the present disclosure;

FIG. 3 is a block diagram illustrating an example of devices of a communication system, which may be used to implement examples of the present disclosure;

FIG. 4 is a block diagram illustrating example units or modules in a device, which may be used to implement examples of the present disclosure;

FIG. 5 is a schematic diagram illustrating an example approach for training encoder and decoder models using a centralized dataset in some prior art approaches;

FIG. 6 is a schematic diagram illustrating the difference in complexity between an ideal encoder model deployed at a UE and an ideal decoder model deployed at a network entity;

FIG. 7 illustrates an example increase in complexity of CSI that may result from advances in technology;

FIG. 8 illustrates an example of the increase in size of a central database required to manage the example increased complexity illustrated in FIG. 7;

FIG. 9 illustrates an example of the increase in complexity of encoder and decoder models due to example increases in communications that may result from advances in technology;

FIG. 10 is a schematic diagram illustrating an example of how a central database may be used to enable interoperability, in accordance with examples of the present disclosure;

FIG. 11 is a schematic diagram illustrating an example of data collection and registration, in accordance with examples of the present disclosure;

FIG. 12 illustrates an example of how a precoder vector may be stored using a sparse representation at a datacenter, in accordance with examples of the present disclosure;

FIG. 13 illustrates an example of how a precoder vector may be added to a representation set at a datacenter, in accordance with examples of the present disclosure;

FIG. 14 illustrates an example of how coefficient vectors from UEs having similar spatial characteristics may be clustered, in accordance with examples of the present disclosure;

FIG. 15 illustrates an example of how a representation set or representation subset may be selected by different NW-vendor or UE-vendors, in accordance with examples of the present disclosure;

FIG. 16 illustrates an example of how a coefficient vector may be mapped from an expanded vector form to a compact vector form, in accordance with examples of the present disclosure;

FIG. 17 illustrates an example of training a neural network to implement an encoder model, in accordance with examples of the present disclosure;

FIG. 18 illustrates an example of training a neural network to implement a decoder model, in accordance with examples of the present disclosure;

FIG. 19 illustrates an example of how trained neural networks may be used for the communication of CSI, in accordance with examples of the present disclosure;

FIG. 20 is a block diagram illustrating an example of a datacenter, in accordance with examples of the present disclosure;

FIG. 21 is a flowchart illustrating an example method for maintaining a dataset and representation set, in accordance with examples of the present disclosure;

FIG. 22 is a flowchart illustrating an example method for training a neural network to perform a precoder vector encoding task, in accordance with examples of the present disclosure;

FIG. 23 is a flowchart illustrating an example method for training a neural network to perform a precoder vector decoding task, in accordance with examples of the present disclosure;

FIG. 24 is a flowchart illustrating an example method for using a trained neural network to decode a precoder vector, in accordance with examples of the present disclosure; and

FIG. 25 is a flowchart illustrating an example method for using a trained neural network to encode a precoder vector, in accordance with examples of the present disclosure.

Similar reference numerals may have been used in different figures to denote similar components.

DETAILED DESCRIPTION

In various examples, the present disclosure describes methods, systems and computer-readable media that enable two-sided machine learning-based compression of channel state information (CSI) . “Two-sided” refers to the use of a trained encoder model for compression of CSI at a user equipment (UE) side and a trained decoder model for decompression of the CSI at the network side. Examples disclosed herein may be useful for improving interoperability of encoder/decoder models developed by different vendors, by providing a central datacenter at which data can be registered and retrieved for training neural networks to implement encoder or decoder models. As disclosed herein, the datacenter may store the data in a sparse format that helps to improve fairness and standardization of encoder/decoder models between UE-side and network-side vendors.

To assist in understanding the present disclosure, reference is first made to FIG. 1.

Referring to FIG. 1, as an illustrative example without limitation, a simplified schematic illustration of a communication system is provided. The communication system 100 comprises a radio access network 120. The radio access network 120 may be a next generation radio access network, or a legacy (e.g. 5G, 4G, 3G or 2G) radio access network. One or more communication electronic devices (ED) 110a, 110b, 110c, 110d, 110e, 110f, 110g, 110h, 110i, 110j (generically referred to as 110) may be interconnected to one another or connected to one or more network nodes (170a, 170b, generically referred to as 170) in the radio access network 120. A core network 130 may be a part of the communication system and may be dependent or independent of the radio access technology used in the communication system 100. Also the communication system 100 comprises a public switched telephone network (PSTN) 140, the internet 150, and other networks 160.

FIG. 2 illustrates an example communication system 100. In general, the communication system 100 enables multiple wireless or wired elements to communicate data and other content. The purpose of the communication system 100 may be to provide content, such as voice, data, video, and/or text, via broadcast, multicast, groupcast, unicast, etc. The communication system 100 may operate by sharing resources, such as carrier spectrum bandwidth, between its constituent elements. The communication system 100 may include a terrestrial communication system and/or a non-terrestrial communication system. The communication system 100 may provide a wide range of communication services and applications (such as earth monitoring, remote sensing, passive sensing and positioning, navigation and tracking, autonomous delivery and mobility, etc. ) . The communication system 100 may provide a high degree of availability and robustness through a joint operation of a terrestrial communication system and a non-terrestrial communication system. For example, integrating a non-terrestrial communication system (or components thereof) into a terrestrial communication system can result in what may be considered a heterogeneous network comprising multiple layers. Compared to conventional communication networks, the heterogeneous network may achieve better overall performance through efficient multi-link joint operation, more flexible functionality sharing, and faster physical layer link switching between terrestrial networks and non-terrestrial networks.

The terrestrial communication system and the non-terrestrial communication system can be considered sub-systems of the communication system. In the example shown in FIG. 2, the communication system 100 includes electronic devices (ED) 110a, 110b, 110c, 110d (generically referred to as ED 110) , radio access networks (RANs) 120a, 120b, a non-terrestrial communication network 120c, a core network 130, a public switched telephone network (PSTN) 140, the Internet 150, and other networks 160. The RANs 120a, 120b include respective base stations (BSs) 170a, 170b, which may be generically referred to as terrestrial transmit and receive points (T-TRPs) 170a, 170b. The non-terrestrial communication network 120c includes an access node 172, which may be generically referred to as a non-terrestrial transmit and receive point (NT-TRP) 172.

Any ED 110 may be alternatively or additionally configured to interface, access, or communicate with any T-TRP 170a, 170b and NT-TRP 172, the Internet 150, the core network 130, the PSTN 140, the other networks 160, or any combination of the preceding. In some examples, ED 110a may communicate via an uplink and/or downlink transmission over a terrestrial air interface 190a with T-TRP 170a. In some examples, the EDs 110a, 110b, 110c, and 110d may also communicate directly with one another via one or more sidelink air interfaces 190b. In some examples, ED 110d may communicate via an uplink and/or downlink transmission over a non-terrestrial air interface 190c with NT-TRP 172.

The air interfaces 190a and 190b may use similar communication technology, such as any suitable radio access technology. For example, the communication system 100 may implement one or more channel access methods, such as code division multiple access (CDMA) , space division multiple access (SDMA) , time division multiple access (TDMA) , frequency division multiple access (FDMA) , orthogonal FDMA (OFDMA) , or single-carrier FDMA (SC-FDMA, also known as discrete Fourier transform spread OFDMA, DFT-s-OFDMA) in the air interfaces 190a and 190b. The air interfaces 190a and 190b may utilize other higher dimension signal spaces, which may involve a combination of orthogonal and/or non-orthogonal dimensions.

The non-terrestrial air interface 190c can enable communication between the ED 110d and one or multiple NT-TRPs 172 via a wireless link or simply a link. For some examples, the link is a dedicated connection for unicast transmission, a connection for broadcast transmission, or a connection between a group of EDs 110 and one or multiple NT-TRPs 172 for multicast transmission.

The RANs 120a and 120b are in communication with the core network 130 to provide the EDs 110a 110b, and 110c with various services such as voice, data, and other services. The RANs 120a and 120b and/or the core network 130 may be in direct or indirect communication with one or more other RANs (not shown) , which may or may not be directly served by core network 130, and may or may not employ the same radio access technology as RAN 120a, RAN 120b or both. The core network 130 may also serve as a gateway access between (i) the RANs 120a and 120b or EDs 110a 110b, and 110c or both, and (ii) other networks (such as the PSTN 140, the Internet 150, and the other networks 160) . In addition, some or all of the EDs 110a 110b, and 110c may include functionality for communicating with different wireless networks over different wireless links using different wireless technologies and/or protocols. Instead of wireless communication (or in addition thereto) , the EDs 110a 110b, and 110c may communicate via wired communication channels to a service provider or switch (not shown) , and to the Internet 150. PSTN 140 may include circuit switched telephone networks for providing plain old telephone service (POTS) . Internet 150 may include a network of computers and subnets (intranets) or both, and incorporate protocols, such as Internet Protocol (IP) , Transmission Control Protocol (TCP) , User Datagram Protocol (UDP) . EDs 110a 110b, and 110c may be multimode devices capable of operation according to multiple radio access technologies, and incorporate multiple transceivers necessary to support such.

FIG. 3 illustrates another example of an ED 110 and a base station 170a, 170b and/or 170c. The ED 110 is used to connect persons, objects, machines, etc. The ED 110 may be widely used in various scenarios including, for example, cellular communications, device-to-device (D2D) , vehicle to everything (V2X) , peer-to-peer (P2P) , machine-to-machine (M2M) , machine-type communications (MTC) , internet of things (IoT) , virtual reality (VR) , augmented reality (AR) , mixed reality (MR) , metaverse, digital twin, industrial control, self-driving, remote medical, smart grid, smart furniture, smart office, smart wearable, smart transportation, smart city, drones, robots, remote sensing, passive sensing, positioning, navigation and tracking, autonomous delivery and mobility, etc.

Each ED 110 represents any suitable end user device for wireless operation and may include such devices (or may be referred to) as a user equipment/device (UE) , a wireless transmit/receive unit (WTRU) , a mobile station, a fixed or mobile subscriber unit, a cellular telephone, a station (STA) , a machine type communication (MTC) device, a personal digital assistant (PDA) , a smartphone, a laptop, a computer, a tablet, a wireless sensor, a consumer electronics device, a smart book, a vehicle, a car, a truck, a bus, a train, or an IoT device, wearable devices (such as a watch, a pair of glasses, head mounted equipment, etc. ) , an industrial device, or an apparatus in (e.g. communication module, modem, or chip) or comprising the forgoing devices, among other possibilities. Future generation EDs 110 may be referred to using other terms. The base station 170a and 170b is a T-TRP and will hereafter be referred to as T-TRP 170. Also shown in FIG. 3, a NT-TRP will hereafter be referred to as NT-TRP 172. Each ED 110 connected to T-TRP 170 and/or NT-TRP 172 can be dynamically or semi-statically turned-on (i.e., established, activated, or enabled) , turned-off (i.e., released, deactivated, or disabled) and/or configured in response to one of more of: connection availability and connection necessity.

The ED 110 includes a transmitter 201 and a receiver 203 coupled to one or more antennas 204. Only one antenna 204 is illustrated to avoid congestion in the drawing. One, some, or all of the antennas 204 may alternatively be panels. The transmitter 201 and the receiver 203 may be integrated, e.g. as a transceiver. The transceiver is configured to modulate data or other content for transmission by at least one antenna 204 or network interface controller (NIC) . The transceiver is also configured to demodulate data or other content received by the at least one antenna 204. Each transceiver includes any suitable structure for generating signals for wireless or wired transmission and/or processing signals received wirelessly or by wire. Each antenna 204 includes any suitable structure for transmitting and/or receiving wireless or wired signals.

The ED 110 includes at least one memory 208. The memory 208 stores instructions and data used, generated, or collected by the ED 110. For example, the memory 208 could store software instructions or modules configured to implement some or all of the functionality and/or embodiments described herein and that are executed by one or more processing unit (s) (e.g., a processor 210) . Each memory 208 includes any suitable volatile and/or non-volatile storage and retrieval device (s) . Any suitable type of memory may be used, such as random access memory (RAM) , read only memory (ROM) , hard disk, optical disc, subscriber identity module (SIM) card, memory stick, secure digital (SD) memory card, on-processor cache, and the like.

The ED 110 may further include one or more input/output devices (not shown) or interfaces (such as a wired interface to the Internet 150 in FIG. 1) . The input/output devices or interfaces permit interaction with a user or other devices in the network. Each input/output device or interface includes any suitable structure for providing information to or receiving information from a user, and/or for network interface communications. Suitable structures include, for example, a speaker, microphone, keypad, keyboard, display, touch screen, etc.

The ED 110 includes the processor 210 for performing operations including those operations related to preparing a transmission for uplink transmission to the NT-TRP 172 and/or the T-TRP 170; those operations related to processing downlink transmissions received from the NT-TRP 172 and/or the T-TRP 170; and those operations related to processing sidelink transmission to and from another ED 110. Processing operations related to preparing a transmission for uplink transmission may include operations such as encoding, modulating, transmit beamforming, and generating symbols for transmission. Processing operations related to processing downlink transmissions may include operations such as receive beamforming, demodulating and decoding received symbols. Depending upon the embodiment, a downlink transmission may be received by the receiver 203, possibly using receive beamforming, and the processor 210 may extract signaling from the downlink transmission (e.g. by detecting and/or decoding the signaling) . An example of signaling may be a reference signal transmitted by the NT-TRP 172 and/or by the T-TRP 170. In some embodiments, the processor 210 implements the transmit beamforming and/or the receive beamforming based on the indication of beam direction, e.g. beam angle information (BAI) , received from the T-TRP 170. In some embodiments, the processor 210 may perform operations relating to network access (e.g. initial access) and/or downlink synchronization, such as operations relating to detecting a synchronization sequence, decoding and obtaining the system information, etc. In some embodiments, the processor 210 may perform channel estimation, e.g. using a reference signal received from the NT-TRP 172 and/or from the T-TRP 170.

Although not illustrated, the processor 210 may form part of the transmitter 201 and/or part of the receiver 203. Although not illustrated, the memory 208 may form part of the processor 210.

The processor 210, the processing components of the transmitter 201, and the processing components of the receiver 203 may each be implemented by the same or different one or more processors that are configured to execute instructions stored in a memory (e.g. in the memory 208) . Alternatively, some or all of the processor 210, the processing components of the transmitter 201, and the processing components of the receiver 203 may each be implemented using dedicated circuitry, such as a programmed field-programmable gate array (FPGA) , an application-specific integrated circuit (ASIC) , or a hardware accelerator such as a graphics processing unit (GPU) or an artificial intelligence (AI) accelerator.

The T-TRP 170 may be known by other names in some implementations, such as a base station, a base transceiver station (BTS) , a radio base station, a network node, a network device, a device on the network side, a transmit/receive node, a Node B, an evolved NodeB (eNodeB or eNB) , a Home eNodeB, a next Generation NodeB (gNB) , a transmission point (TP) , a site controller, an access point (AP) , a wireless router, a relay station, a terrestrial node, a terrestrial network device, a terrestrial base station, a base band unit (BBU) , a remote radio unit (RRU) , an active antenna unit (AAU) , a remote radio head (RRH) , a central unit (CU) , a distributed unit (DU) , a positioning node, among other possibilities. The T-TRP 170 may be a macro BS, a pico BS, a relay node, a donor node, or the like, or combinations thereof. The T-TRP 170 may refer to the forgoing devices or refer to apparatus (e.g. a communication module, a modem, or a chip) in the forgoing devices.

In some embodiments, the parts of the T-TRP 170 may be distributed. For example, some of the modules of the T-TRP 170 may be located remote from the equipment that houses the antennas 256 for the T-TRP 170, and may be coupled to the equipment that houses the antennas 256 over a communication link (not shown) sometimes known as front haul, such as common public radio interface (CPRI) . Therefore, in some embodiments, the term T-TRP 170 may also refer to modules on the network side that perform processing operations, such as determining the location of the ED 110, resource allocation (scheduling) , message generation, and encoding/decoding, and that are not necessarily part of the equipment that houses the antennas 256 of the T-TRP 170. The modules may also be coupled to other T-TRPs. In some embodiments, the T-TRP 170 may actually be a plurality of T-TRPs that are operating together to serve the ED 110, e.g. through the use of coordinated multipoint transmissions.

The T-TRP 170 includes at least one transmitter 252 and at least one receiver 254 coupled to one or more antennas 256. Only one antenna 256 is illustrated to avoid congestion in the drawing. One, some, or all of the antennas 256 may alternatively be panels. The transmitter 252 and the receiver 254 may be integrated as a transceiver. The T-TRP 170 further includes a processor 260 for performing operations including those related to: preparing a transmission for downlink transmission to the ED 110, processing an uplink transmission received from the ED 110, preparing a transmission for backhaul transmission to the NT-TRP 172, and processing a transmission received over backhaul from the NT-TRP 172. Processing operations related to preparing a transmission for downlink or backhaul transmission may include operations such as encoding, modulating, precoding (e.g. multiple input multiple output (MIMO) precoding) , transmit beamforming, and generating symbols for transmission. Processing operations related to processing received transmissions in the uplink or over backhaul may include operations such as receive beamforming, demodulating received symbols, and decoding received symbols. The processor 260 may also perform operations relating to network access (e.g. initial access) and/or downlink synchronization, such as generating the content of synchronization signal blocks (SSBs) , generating the system information, etc. In some embodiments, the processor 260 also generates an indication of beam direction, e.g. BAI, which may be scheduled for transmission by a scheduler 253. The processor 260 performs other network-side processing operations described herein, such as determining the location of the ED 110, determining where to deploy the NT-TRP 172, etc. In some embodiments, the processor 260 may generate signaling, e.g. to configure one or more parameters of the ED 110 and/or one or more parameters of the NT-TRP 172. Any signaling generated by the processor 260 is sent by the transmitter 252. Note that “signaling” , as used herein, may alternatively be called control signaling. Signaling may be transmitted in a physical layer control channel, e.g. a physical downlink control channel (PDCCH) , in which case the signaling may be known as dynamic signaling. Signaling transmitted in a downlink physical layer control channel may be known as Downlink Control Information (DCI) . Signaling transmitted in an uplink physical layer control channel may be known as Uplink Control Information (UCI) . Signaling transmitted in a sidelink physical layer control channel may be known as Sidelink Control Information (SCI) . Signaling may be included in a higher-layer (e.g., higher than physical layer) packet transmitted in a physical layer data channel, e.g. in a physical downlink shared channel (PDSCH) , in which case the signaling may be known as higher-layer signaling, static signaling, or semi-static signaling. Higher-layer signaling may also refer to Radio Resource Control (RRC) protocol signaling or Media Access Control –Control Element (MAC-CE) signaling.

The scheduler 253 may be coupled to the processor 260. The scheduler 253 may be included within or operated separately from the T-TRP 170. The scheduler 253 may schedule uplink, downlink, sidelink, and/or backhaul transmissions, including issuing scheduling grants and/or configuring scheduling-free (e.g., “configured grant” ) resources. The T-TRP 170 further includes a memory 258 for storing information and data. The memory 258 stores instructions and data used, generated, or collected by the T-TRP 170. For example, the memory 258 could store software instructions or modules configured to implement some or all of the functionality and/or embodiments described herein and that are executed by the processor 260.

Although not illustrated, the processor 260 may form part of the transmitter 252 and/or part of the receiver 254. Also, although not illustrated, the processor 260 may implement the scheduler 253. Although not illustrated, the memory 258 may form part of the processor 260.

The processor 260, the scheduler 253, the processing components of the transmitter 252, and the processing components of the receiver 254 may each be implemented by the same or different one or more processors that are configured to execute instructions stored in a memory, e.g. in the memory 258. Alternatively, some or all of the processor 260, the scheduler 253, the processing components of the transmitter 252, and the processing components of the receiver 254 may be implemented using dedicated circuitry, such as a programmed FPGA, a hardware accelerator (e.g., a GPU or AI accelerator) , or an ASIC.

Although the NT-TRP 172 is illustrated as a drone only as an example, the NT-TRP 172 may be implemented in any suitable non-terrestrial form, such as satellites and high altitude platforms, including international mobile telecommunication base stations and unmanned aerial vehicles, for example. Also, the NT-TRP 172 may be known by other names in some implementations, such as a non-terrestrial node, a non-terrestrial network device, or a non-terrestrial base station. The NT-TRP 172 includes a transmitter 272 and a receiver 274 coupled to one or more antennas 280. Only one antenna 280 is illustrated to avoid congestion in the drawing. One, some, or all of the antennas may alternatively be panels. The transmitter 272 and the receiver 274 may be integrated as a transceiver. The NT-TRP 172 further includes a processor 276 for performing operations including those related to: preparing a transmission for downlink transmission to the ED 110, processing an uplink transmission received from the ED 110, preparing a transmission for backhaul transmission to T-TRP 170, and processing a transmission received over backhaul from the T-TRP 170. Processing operations related to preparing a transmission for downlink or backhaul transmission may include operations such as encoding, modulating, precoding (e.g. MIMO precoding) , transmit beamforming, and generating symbols for transmission. Processing operations related to processing received transmissions in the uplink or over backhaul may include operations such as receive beamforming, demodulating received symbols, and decoding received symbols. In some embodiments, the processor 276 implements the transmit beamforming and/or receive beamforming based on beam direction information (e.g. BAI) received from the T-TRP 170. In some embodiments, the processor 276 may generate signaling, e.g. to configure one or more parameters of the ED 110. In some embodiments, the NT-TRP 172 implements physical layer processing, but does not implement higher layer functions such as functions at the medium access control (MAC) or radio link control (RLC) layer. As this is only an example, more generally, the NT-TRP 172 may implement higher layer functions in addition to physical layer processing.

The NT-TRP 172 further includes a memory 278 for storing information and data. Although not illustrated, the processor 276 may form part of the transmitter 272 and/or part of the receiver 274. Although not illustrated, the memory 278 may form part of the processor 276.

The processor 276, the processing components of the transmitter 272, and the processing components of the receiver 274 may each be implemented by the same or different one or more processors that are configured to execute instructions stored in a memory, e.g. in the memory 278. Alternatively, some or all of the processor 276, the processing components of the transmitter 272, and the processing components of the receiver 274 may be implemented using dedicated circuitry, such as a programmed FPGA, a hardware accelerator (e.g., a GPU or AI accelerator) , or an ASIC. In some embodiments, the NT-TRP 172 may actually be a plurality of NT-TRPs that are operating together to serve the ED 110, e.g. through coordinated multipoint transmissions.

The T-TRP 170, the NT-TRP 172, and/or the ED 110 may include other components, but these have been omitted for the sake of clarity.

One or more steps of the embodiment methods provided herein may be performed by corresponding units or modules, according to FIG. 4. FIG. 4 illustrates units or modules in a device, such as in the ED 110, in the T-TRP 170, or in the NT-TRP 172. For example, a signal may be transmitted by a transmitting unit or by a transmitting module. A signal may be received by a receiving unit or by a receiving module. A signal may be processed by a processing unit or a processing module. Other steps may be performed by an artificial intelligence (AI) or machine learning (ML) module. The respective units or modules may be implemented using hardware, one or more components or devices that execute software, or a combination thereof. For instance, one or more of the units or modules may be a circuit such as an integrated circuit. Examples of an integrated circuit includes a programmed FPGA, a GPU, or an ASIC. For instance, one or more of the units or modules may be logical such as a logical function performed by a circuit, by a portion of an integrated circuit, or by software instructions executed by a processor. It will be appreciated that where the modules are implemented using software for execution by a processor for example, the modules may be retrieved by a processor, in whole or part as needed, individually or together for processing, in single or multiple instances, and that the modules themselves may include instructions for further deployment and instantiation.

Additional details regarding the EDs 110, the T-TRP 170, and the NT-TRP 172 are known to those of skill in the art. As such, these details are omitted here.

The CSI compression framework in 5G NR is based on the principle of compressing the CSI before feedback from the UE to the terrestrial transmit and receive point (T-TRP) (e.g., gNB) . In various examples, the present disclosure makes reference to a gNB or a BS as an example T-TRP; it should be understood that this is only exemplary and is not intended to be limiting. This compression is achieved through dimensionality reduction techniques that exploit the inherent spatial and temporal correlations in the wireless channel. Specifically, the compression schemes operate on the precoding matrix indicators (PMIs) that represent the recommended precoding vectors or precoding matrix, or beamforming matrix, for different transmission layers and subbands (which may be a number of consecutive resource blocks) .

Two primary compression modes were specified for CSI feedback in 5G NR Release 15, namely Type I and Type II.Type I compression is a codebook-based approach, wherein the UE selects the precoding matrix from a standardized codebook and feeds back the corresponding PMI index. Type II compression, on the other hand, is a more flexible scheme that allows the UE to provide a wideband PMI along with a set of subband differential PMIs, enabling more granular CSI reporting.

Furthermore, the 3GPP standard supports an enhanced Type II (eType II) compression mode that introduces additional compression capabilities. eType II builds upon Type II by incorporating two enhancements: spatial domain compression and frequency domain compression. Spatial domain compression exploits the correlation across the angular domain, while frequency domain compression leverages the correlation across adjacent subbands.

The compression techniques specified in 5G NR Release 15 are non-data-driven, meaning that they rely on predefined codebooks and compression matrices derived from mathematical models and assumptions about the channel. However, with the advent of AI and ML techniques, there has been growing interest in exploring data-driven approaches for CSI compression, which can potentially provide better performance by adapting to the actual channel statistics and dynamics.

While the non-data-driven CSI compression techniques standardized in 5G NR Release 15 provide a baseline solution, there has been increasing interest in exploring data-driven approaches based on AI and ML for CSI compression. These AI/ML-based techniques have the potential to achieve better compression performance by adapting to the actual channel statistics and dynamics, rather than relying on predefined codebooks and assumptions.

One approach that has been studied is the use of a two-sided AI/ML model for CSI compression. In this framework, a pair of AI/ML model, usually deep neural networks, are jointly trained: an encoder model at the UE side and a decoder model at the gNB side. The encoder model at the UE compresses the high-dimensional CSI, usually a precoding matrix, into a low-dimensional representation, which is then fed back to the gNB. The decoder model at the gNB then reconstructs the original CSI from this compressed representation.

An advantage of the two-sided model is that it allows for end-to-end optimization of the compression and reconstruction processes. By jointly training the encoder and decoder models on representative channel data, the two-sided model can learn the most effective compression strategies and adapt to the specific channel characteristics of the deployment scenario.

During the training phase, the two-sided model is exposed to (trained on) a large dataset of channel realizations, typically obtained through simulations or field measurements. The encoder and decoder models are iteratively updated to minimize a loss function that quantifies the reconstruction error between the original CSI and the reconstructed CSI. Once trained, the encoder model is deployed at the UE, while the decoder model is deployed at the gNB.

One important aspect of the two-sided model is the need for collaboration between the UE-side and NW-side vendors during the training process. Since the encoder and decoder models are interdependent, their respective vendors coordinate to ensure compatibility and interoperability. Different approaches, such as sequential training, joint training, or model exchange, have been proposed to facilitate this collaboration.

The two-sided AI/ML model for CSI compression has shown promising results in terms of compression performance and adaptation to real-world channel conditions. However, challenges remain in areas such as inter-vendor collaboration, model complexity management, and deployment considerations. Ongoing research efforts aim to address these challenges and pave the way for the practical adoption of AI/ML-based CSI compression techniques in future wireless communication systems.

Some previously proposed approaches for AI/ML-based CSI compression techniques are now discussed.

To foster interoperability among two-sided AI/ML models for CSI compression developed by different vendors, a distributed training scheme based on a standardized dataset has been proposed. This approach introduces a central repository, a database accessible to all participating vendors, containing a vast collection of channel instance data. This data, typically derived from simulations or real-world measurements, encapsulates the statistical characteristics and dynamic behavior of actual wireless channels, offering a common ground for model training.

The central database serves as a unifying element, enabling UE and NW vendors to train their respective encoder and decoder models using the same underlying data distribution. This shared foundation theoretically paves the way for seamless interoperability, ensuring that the trained models can effectively communicate and cooperate regardless of their specific design or implementation.

FIG. 5 illustrates an example of this concept, depicting how both NW and UE vendors access the standardized dataset of precoder matrices (V) from the central database. Each vendor then independently trains their autoencoder (AE) models, consisting of an encoder model and a decoder model, using the shared data set. Upon completion of training, the NW vendor deploys the decoder portion of their model (usually called the CSI generator or CSI decompressor) , while the UE vendor deploys the encoder portion (usually called CSI compressor) , enabling end-to-end CSI compression and reconstruction.

As illustrated in FIG. 5, precoder matrices V are stored in the central database 502 and accessible to both NW and UE vendors, for NW-side vendor training and UE-side vendor training, respectively. Each matrix V may have dimensions N_BS x r, for example, where N_BS represents the number of antennas at the BS and r represents the rank resulting from singular value decomposition (SVD) used to generate V. At the NW-side vendor training, an AE model (consisting of an encoder model, denoted ENC_NW 504, and a decoder model, denoted DEC_NW 506) is trained by minimizing some loss function (e.g., min (|v’-v|²) ) . Similarly, at the UE-side vendor training, another AE model (consisting of an encoder model, denoted ENC_UE 508, and a decoder model, denoted DEC_UE 510) are trained by minimizing some loss function (e.g., min (|v’-v|²) ) . After training, the trained UE-side encoder ENC_UE 508 is used with the trained NW-side decoder DEC_NW 506 during inference.

The appeal of this approach lies in its simplicity and potential for minimizing direct collaboration efforts between vendors without sharing the trained models among the vendors (because the trained models are considered to be highly proprietary) . Rather than engaging in complex coordination and data exchange procedures, vendors can rely on the central database as a trusted source of training data, streamlining the development process and promoting interoperability.

However, relying solely on a standardized dataset has limitations. The inherent diversity in model architectures, parameter (or meta parameters) choices, training methodologies, and even the stochastic nature of AI/ML training algorithms can lead to discrepancies in the learned latent representations that is used as PMI mapping between the encoder and decoder. This means that even when trained on the same dataset, models from different vendors might develop different internal understandings of the information that is reflected on the distribution of the PMI mapping layer, potentially hindering seamless communication and cooperation during inference.

In sum, while a centralized training dataset offers a promising starting point for achieving interoperability, additional mechanisms are needed to ensure consistent and reliable performance across a diverse ecosystem of vendors and devices.

The sequential training approach offers a potential solution to the interoperability challenge inherent in two-sided AI/ML models for CSI compression. By establishing a predetermined order for model development, it ensures that the latent representations (PMI mapping layer) , the internal data structures learned by the models, are aligned between the encoder and decoder.

As outlined, the process typically begins with the NW-vendor taking the lead. They train their AE model using precoder matrices from the central database, creating a baseline latent representation. Subsequently, they share both the input data and the corresponding latent layer outputs, essentially providing a "key" for the UE vendor to unlock compatible encoding.

This approach streamlines the collaboration process and ensures compatibility between the final deployed models. However, it also introduces potential drawbacks that need to be considered. Primarily, it places the NW-vendor in a dominant position, as their initial model training dictates the performance ceiling for the UE-vendor's encoder. This can stifle innovation and limit the UE-vendor's ability to differentiate their products based on superior compression performance.

Furthermore, the inherent differences in design philosophies between UE-vendors and NW-vendors exacerbate the issue. UE-vendors, constrained by power and size limitations on user devices, often prioritize smaller, more efficient models even if they sacrifice some performance. The efficiency of these models is closely linked to the sparsity of the CSI data, which in turn is influenced by the degree of localized data. In contrast, NW-vendors, with less stringent hardware limitations, typically favor larger models with broader generalization capabilities, encompassing a wider range of scenarios and configurations.

FIGS. 6A and 6B illustrates an example of this disparity between NW-vendor priorities and UE-vendor priorities.

In this example, a BS 602 may communicate with a user equipment (UE) 604 over a number of different channels, which may include relays (e.g., via buildings 606) . The BS 602 may have a large number of antennas (e.g., N_BS = 512) while the UE 604 may have a smaller number of antennas (e.g., 16) . Additionally, the BS 602 may be concerned with channel conditions for all UEs in its service area 608, whereas the UE 604 may be concerned with only its own service area 610. This means that while the trained NW-side DEC_NW 506 may need to be more complex, ideally the trained UE-side ENC_UE 508 is preferably significantly smaller in scale, to avoid placing a large burden on the UE 604 (which has more limited computing power, battery power, memory resources, etc. compared to the BS 602) , as illustrated in FIG. 6B.

This disparity in priorities can lead to an imbalance in the sequential training process. When the NW-vendor dictates the latent representation, the UE-vendor's encoder model might be forced to learn irrelevant information, adding unnecessary complexity and potentially hindering efficiency. This ultimately disadvantages the UE-vendor and restricts their ability to optimize their models for specific use cases or device capabilities.

Generally, it may be expected that complexity and/or amount of communications in wireless communication networks will increase as technology advances. For example, the technical advances from 5G to 6G are discussed below, however the present disclosure is not intended to be limited to 6G embodiments or any other particular generation of technology. In 6G ultra-Massive MIMO systems, the scale of the wireless interface experiences a dramatic expansion compared to its 5G predecessor. This growth is evident not only in the sheer size of the infrastructure but also in the complexity of the information it handles.

FIG. 7 provides a visual comparison, highlighting the differences. While 5G Release 19 Massive MIMO systems typically operate with a bandwidth of 50 MHz, utilizing 64 antennas at the gNB and 8 at the UE, 6G envisions a much grander scale. With bandwidths reaching 400 MHz, gNB equipped with 512 antennas, and UEs with 16, the amount of CSI that needs to be processed and exchanged within the network increases exponentially.

This increase in scale has a cascading effect on various aspects of the system. The channel matrix, which represents the propagation characteristics between the transmitter and receiver, grows significantly in size. This, in turn, leads to more complex precoding matrices (V) , which are essential for beamforming and spatial multiplexing techniques. As a result, the dimensionality of the CSI explodes, creating a data deluge that challenges conventional methods of compression and feedback.

Furthermore, the increased number of antennas and wider bandwidth lead to a higher number of MIMO subchannels. This translates to a richer spatial and frequency domain, offering greater potential for spatial multiplexing and higher data rates. However, it also means that the AI/ML models used for CSI compression need to be more sophisticated and capable of handling this additional complexity. This results in larger models with increased training data requirements and computational demands.

The most immediate consequence of the scaled 6G air interface is the dramatic surge in storage requirements for the central database. Compared to 5G, the data footprint for 6G could potentially expand by a factor of 100 or even more. Let's break down the contributing factors to understand the magnitude of this data explosion.

Firstly, the increased bandwidth from 50 MHz in 5G to 400 MHz in 6G leads to a direct eightfold increase in the amount of frequency-domain information that needs to be stored. Secondly, the number of antennas at the base station jumps from 64 to 512, resulting in an eightfold increase in the spatial domain data. Considering both factors, we already face a 64-fold increase in data volume compared to 5G.

Additionally, factors like the number of users, the number of cells in the network, and the frequency of data collection contribute to the overall storage demand. Furthermore, decisions regarding data retention policies and the need for historical data for model training and analysis further influence the storage footprint.

Taking these factors into account, it becomes evident that maintaining such a massive-scale central database is no small feat.

FIG. 8 illustrates the potential challenges of managing this data deluge, highlighting the need for robust and scalable infrastructure, efficient data management strategies, and advanced data security measures.

As illustrated in FIG. 8, while in 5G R19, the precoding matrices V may be relatively small, the precoding matrices V in 6G are expected to be significantly larger. As a result, the central database in 6G (Release 20) is expected to be at least 100 times larger than that of 5G R19.

Beyond the data storage challenges, the increased scale of 6G necessitates a significant increase in the size and complexity of the AI/ML models themselves. As the volume of CSI requiring compression grows, so too does the need for more sophisticated models capable of handling this increased dimensionality and intricacy.

FIG. 9 depicts this challenge, illustrating how the expanded air interface and the greater number of MIMO subchannels contribute to the growth of AI/ML models. The models must be able to effectively capture the complex spatial and frequency correlations within the channel, requiring more parameters and intricate architectures. This directly translates to an increased demand for training data, as larger models need more diverse and representative examples to learn effectively and avoid overfitting.

As illustrated in FIG. 9, the encoder and decoder models in 6G are expected to be significantly larger and more complex than the encoder and decoder models in 5G, due to the increase in size of the precoding matrices V in 6G.

The implications of larger AI/ML models are multifaceted. Training times become significantly longer, necessitating more powerful computing infrastructure and potentially leading to bottlenecks in the model development process. The computational resources required for model inference also increase, potentially straining the capabilities of UEs and gNBs. Furthermore, the energy consumption associated with training and running (inferencing) these models becomes a significant concern, impacting the overall energy efficiency and sustainability of the system.

The confluence of these two critical factors-the exponential growth in data storage needs and the demand for increasingly complex AI/ML models-poses a significant obstacle to the practical implementation of CSI compression based on a central database approach in 6G. The sheer scale of the data and the computational demands of larger models push the limits of current technological capabilities and raise concerns about the sustainability and efficiency of such a centralized approach.

The strain on network resources, the potential for bottlenecks in data transmission, and the energy consumption associated with managing and processing this massive amount of information all point to the need for a paradigm shift. The centralized model, while effective in previous generations with smaller-scale systems, struggles to keep pace with the demands of 6G.

As we move towards 6G, the landscape of antenna design is poised for a significant transformation. Moving beyond the traditional assumption of uniform planar arrays, 6G opens doors to explore more diverse and irregular antenna configurations. This shift presents both challenges and opportunities, particularly for CSI compression techniques.

The limitations of conventional CSI compression methods become apparent when dealing with non-conventional antenna arrays. Methods relying on predefined codebooks or structured assumptions about the channel struggle to effectively capture the unique characteristics and spatial correlations present in irregular antenna layouts.

However, this challenge also presents a fertile ground for innovation. One potential avenue for exploration is the CSI-RS antenna port muting scenario. By selectively deactivating specific antenna ports, either for energy-saving purposes or due to hardware limitations, we can create irregular antenna patterns. This deviation from uniformity provides an opportunity to develop more adaptable and efficient CSI compression techniques that can exploit the inherent sparsity and unique spatial correlations present in such configurations.

Eigenvector-based CSI compression emerges as a promising candidate in this context. By focusing on the dominant eigenvectors of the channel matrix, which represent the most significant spatial characteristics, we can achieve better compression ratios and improved reconstruction accuracy compared to conventional methods. Furthermore, incorporating AI/ML techniques allows us to develop models that can learn and adapt to the specific characteristics of different irregular antenna layouts, further enhancing performance and flexibility.

The potential benefits extend beyond simple port muting scenarios. Exploring more complex non-conventional antenna configurations, such as cylindrical or spherical arrays, could unlock even greater gains. These configurations offer unique beamforming capabilities and spatial diversity, enabling more efficient utilization of the wireless spectrum and potentially leading to significant improvements in CSI compression and overall system performance.

When antenna elements are arranged in a non-uniform fashion, this shift away from uniformity presents challenges for traditional codebook generation methods, particularly those relying on Discrete Fourier Transform (DFT) principles.

5G Release 18, as part of the 3GPP specifications for 5G networks, introduced the Type II and eType-II codebook. These codebooks utilize PMIs to describe channel characteristics in multi-antenna systems, facilitating efficient feedback of CSI. However, the underlying assumption of these codebooks, as with many DFT-based methods, is that the antenna elements are uniformly spaced within the array. This simplifies the process of codebook generation and allows for efficient mapping between precoding vectors and their corresponding PMIs via simple DFT.

However, when dealing with irregular antenna arrays, where the spacing between elements is non-uniform, the assumptions of DFT-based methods no longer hold true. The inherent relationship between the physical antenna layout and the signal's spatial characteristics becomes more complex, rendering traditional DFT-based codebook generation techniques ineffective.

While the eigenvector-channel still exist within a Fourier transformation in general, referring to a broader signal space beyond the limitations of conventional DFT assumptions, directly applying DFT-based methods to generate codebooks becomes problematic. The irregular arrangement of antennas disrupts the straightforward mapping between spatial characteristics and frequency domain representations, requiring new approaches to codebook generation.

This challenge becomes even more pronounced when considering scenarios where antenna elements can be dynamically activated or deactivated, such as in energy-saving techniques or when dealing with faulty hardware. The varying configuration of the antenna array further complicates the task of generating codebooks that accurately reflect the dynamic spatial characteristics of the system.

Therefore, exploring alternative methods for codebook generation becomes essential for harnessing the potential of irregular antenna arrays in 6G. These methods need to account for the non-uniform spacing of antenna elements and adapt to dynamic array configurations, paving the way for efficient CSI compression and feedback in future wireless communication systems.

To facilitate seamless data exchange between vendors and promote interoperability of AI/ML models for CSI compression, a central database can be introduced. This centralized repository, such as a 3GPP file server or a dedicated core network entity, serves as a platform where vendors can share and access standardized datasets for model training.

The process begins with NW and UE vendors registering their respective datasets within the central database. These datasets, curated and prepared according to predefined standards, encompass a wide range of channel realizations and scenarios, reflecting the diverse conditions encountered in real-world deployments. By making these datasets readily available, the central database fosters a collaborative environment where vendors can leverage a common pool of data to train their models.

UE vendors can then retrieve the relevant datasets from the central database to train their encoder models. Similarly, NW vendors can access the same datasets to train their decoder models. This shared access to training data ensures that the resulting models are compatible and can interoperate effectively, regardless of the specific vendor or implementation.

Reference is now made to FIG. 10, illustrating an example in which a central database is used for data exchange. In this example, the process includes a data collection step 1002, in which the collected data (e.g., in the form of a channel matrix, H, containing channel estimates) may be processed using rank-reduced SVD to obtain the precoding matrix V. Then at the next step 1004, the precoding matrix V is registered into the central database 1006. At some later time at step 1008, a NW vendor and/or UE vendor can retrieve V from the central database 1006.

To ensure smooth and consistent operation, the registration and retrieval procedures for the central database are standardized. This includes defining common data formats, metadata specifications, and access protocols. By adhering to these standards, vendors can seamlessly exchange data and ensure compatibility across different platforms and implementations.

Security and access control mechanisms are critical components of the central database infrastructure. Robust authentication and authorization protocols are implemented to ensure that only authorized vendors can access and contribute data. Additionally, data encryption and integrity checks are employed to protect the confidentiality and integrity of the stored information.

The registration process should encompass a well-defined set of steps, including data format validation, metadata submission, and quality checks. These measures guarantee that the submitted datasets adhere to the established standards and are suitable for model training. Additionally, access control mechanisms are implemented to ensure that only authorized vendors can register datasets and that appropriate permissions are granted for data access and modification.

Similarly, the retrieval process follows a standardized protocol to ensure consistency and ease of use. Vendors can query the database using specific criteria, such as scenario type, frequency range, or antenna configuration, to identify relevant datasets for their specific needs. The retrieved datasets are then delivered in a standardized format, facilitating seamless integration into the vendor's training pipelines.

By implementing these standardized procedures, the central database fosters a reliable and efficient platform for data exchange, promoting interoperability and accelerating the development of robust and high-performing AI/ML models for CSI compression.

The central database acts as a critical hub for data exchange and collaboration, requiring efficient and secure communication channels. Several options exist for vendors to interact with this central repository, each catering to different needs and scenarios.

Within the core network, Application Programming Interfaces (APIs) emerge as a powerful solution. Standardized APIs provide a structured and well-defined interface for vendors to register, retrieve, and query datasets. This approach facilitates efficient data transfer, enhances security, and simplifies integration with existing systems. Alternatively, established technologies like secure File Transfer Protocols (FTPs) can be used for uploading and downloading datasets. While simpler to implement, FTPs lack the flexibility and efficiency of APIs for complex data interactions. For vendors seeking more granular control and manipulation of data, direct interaction with the database using standardized query languages like SQL is an option. However, this requires expertise in database management and introduces potential security vulnerabilities if not managed carefully.

Moving to the wireless realm, the air interface offers a direct communication path between user equipment (UE) , base stations (gNBs) , and the central database. Dedicated control channels can be employed for low-latency data exchange, ensuring dedicated resources for this critical function. However, this approach necessitates allocating additional radio resources and raises concerns about potential interference.

Existing user plane data channels provide another avenue for transferring smaller datasets or model (herein model is not CSI encoder or decoder model but the model help to register, store, retrieve the data) updates. Leveraging existing infrastructure eliminates the need for additional resource allocation but may not be suitable for large datasets due to bandwidth limitations. In scenarios where device-to-device communication is available, sidelink channels offer an efficient means for local data exchange, reducing the burden on the core network. However, their range and reliability are limited compared to cellular connections.

Irrespective of the chosen communication method, establishing standardized signaling mechanisms and protocols is paramount. This includes defining consistent message formats, implementing robust authentication and authorization protocols, employing data encryption, and ensuring data integrity through checksums or other mechanisms. The chosen solution should prioritize scalability to accommodate the growing demands of 6G, while maintaining security, efficiency, and flexibility to adapt to diverse use cases and deployment scenarios.

To illustrate the process of data collection and registration within the central database, let's follow the journey of a single data sample, from its initial capture at the UE to its final destination in the central repository. This detailed breakdown will provide a clear understanding of the steps involved, which can be readily extrapolated to encompass the collection, registration, and retrieval of a multitude of data samples.

Reference is now made to FIG. 11, illustrating an example of how data collected at the UE may be transmitted to the central database in the network, in accordance with an example of the present disclosure. FIG. 11 shows data collection 1102 at the UE. For example, using a CSI-RS on one resource block (RB) or subband (f_k) , the UE performs channel estimation to obtain the channel matrix H (f_k) . Then, using rank-reduced SVD, a precoding matrix V is obtained.

Given that real-time processing isn't a strict requirement during data collection, the UE can be assigned a specialized role in gathering raw precoder data or matrices (V) . This involves dedicating a portion of the UE's resources specifically to this task, ensuring efficient and focused data acquisition. A precoder data, or a precoding matrix (V) is in terms of the subband (f_k) . To simplify the following discussion, we omit the f_k. The methods in the following discussion can be easily applied to various subbands or resource blocks or resource elements.

The process commences with the UE performing channel estimation on the downlink CSI-RS, reference signals (pilots) in the downlink from gNB to the UE. This step is essential, as it allows the UE to estimate the current state and characteristics of the wireless channel. Following this, the UE undertakes the task of decomposing the estimated channel to identify the optimal precoder, the beamforming vector that maximizes signal quality and minimizes interference. Once determined, this precoder data, usually in form of precoding matrix (number of BS antennas × number of MIMO layers) is transmitted back to the network via the uplink data channel.

To provide context and ensure the usability of the collected data, the UE also gathers and reports ancillary information. This includes details like the specific sub-bands or resource blocks (RBs) used during data collection and the geographical location where the data was obtained. These additional parameters offer valuable insights into the environmental conditions and system configuration under which the data was captured, enhancing the overall value and applicability of the dataset. Such ancillary information may be transmitted to the network with the precoder matrix V.

Optionally, to optimize resource utilization and reduce transmission overhead, the precoder data can be compressed before being sent to the network. Various compression techniques, such as those based on quantization or dimensionality reduction, can be employed, depending on the specific requirements and capabilities of the system.

This comprehensive data collection process ensures that the central database is populated with rich and informative datasets, providing a solid foundation for training robust and high-performing AI/ML models for CSI compression. an exemplary and generic design for the protocol and signaling between the UE, base station (gNB) , and network, focusing on the physical and MAC layers via the air interface.

Below is an exemplary and generic design for the protocol and signaling between the UE, base station (gNB) , and network, focusing on the physical and MAC layers via the air interface.

First is described an example data collection signaling:

Triggering data collection may include the following:

· The network can initiate data collection by sending a dedicated RRC message (e.g., AI_DataCollectionRequest) to the UE. This message can include parameters such as:

○ Target CSI-RS resources: Specify the CSI-RS resources the UE should use for channel estimation.

○ Data collection duration: Define the duration or number of time instances for data collection.

○ Reporting mode: Indicate whether the UE should report data periodically, upon completion, or based on specific events (e.g., handover, change in channel conditions) .

○ Data compression options: Specify if and how the UE should compress the collected precoder data.

Reporting collected data may include the following:

· The UE can report the collected data using existing uplink channels (e.g., PUCCH, PUSCH) or potentially via a dedicated control channel. The report may include:

○ Precoder matrices (V) : The raw or compressed precoder data collected from CSI-RS measurements.

○ Ancillary information: Additional parameters like RB allocation, geographical location, timestamp, etc.

○ Data quality indicators: Optional metrics indicating the quality or reliability of the collected data.

Next is described an example model update signaling. In this context, the term “model” may refer to the representation set (discussed further below) maintained at the central database and not the encoder/decoder model. The model (also referred to as the representation set) includes parameters that are useful for ensuring consistency in data registration and management across network entities. These parameters may include the representation set ID (e.g., an identifier that helps to ensure that the correct version of the representation set is used across different network entities) , weights and biases of the neural network (e.g., the learned parameters of a trained neural network used for implementing the encoder/decoder models) , hyperparameters used during training (e.g., including parameters such as learning rates, batch sizes and regularization strengths used during training of the neural network) , regularizer terms (e.g., terms used to help ensure the trained encoder/decoder model maintains efficient operation while avoiding overfitting) , and quantization parameters (e.g., parameters that help to ensure the quantization process at the UE and the dequantization process at the network entities are consistent, which may help to preserve the integrity of the compressed information) . These parameters may be communicated to the UE to maintain synchronization with the central datacenter.

A model update indication (also referred to as an indication of a representation set update) may include the following:

· The network can inform the UE of a model update (also referred to as a representation set update) by sending a dedicated RRC message (e.g., AI_ModelUpdate) . This message can include:

○ Model ID (also referred to as a representation set ID) : Unique identifier for the new model (also referred to as the new representation set) .

○ Model size (also referred to as a representation set size) : Size of the model parameters (also referred to as the representation set parameters) to be transferred.

○ Model transfer method (also referred to as a representation set transfer method) : Specify the method for transferring the model parameters (also referred to as the representation set parameters) (e.g., RRC signaling, user plane data channel) .

A model parameter transfer (also referred to as a representation set parameter transfer method) may include the following:

· The network can transfer the model parameters (also referred to as the representation set parameters) to the UE using:

○ RRC Signaling: This method is suitable for smaller models (also referred to as smaller representation sets) and offers reliable delivery through dedicated control channels.

○ User Plane Data Channel: This method is suitable for larger models (also referred to as larger representation sets) and utilizes existing data channels, but may have higher latency and potential impact on user traffic.

· The model parameters (also referred to as representation set parameters) can be sent in segments, with appropriate error detection and correction mechanisms to ensure reliable delivery.

Storing raw precoder data (V) directly in the central database can be inefficient due to its high dimensionality and potential redundancy. To optimize storage and retrieval, employing a sparse representation becomes crucial. This involves representing the data using a minimal set of essential components, discarding redundant or insignificant information.

When dealing with precoder matrices (V) obtained through specific decomposition techniques like SVD (singular value decomposition) , the resulting column vectors exhibit orthogonality. This implies that the data columns for each user are linearly independent and cannot be further simplified or compressed by expressing them as linear combinations of other vectors. In essence, they represent the most fundamental and irreducible components of the data from one reporting user.

However, despite the lack of inherent sparsity within individual precoder matrices, a different form of sparsity emerges when considering data collected under similar radio environmental conditions. Precoder matrices gathered in such environments tend to exhibit strong correlations and similarities. This means that when new data is introduced to the central database, it is likely to be highly correlated with existing data or can be closely approximated using the sparse representations of data already present in the database.

This observation opens up a new avenue for achieving efficient data storage. Instead of storing each new precoder matrix independently, we can leverage these correlations and express them using coefficients with respect to a data basis (also referred to as a representation set) formed by existing data or their sparse representations. This effectively reduces the storage requirements, as we only need to store the coefficients and the basis vectors (also referred to as representative vectors) forming the representation set, rather than the entire high-dimensional precoder matrices. In some examples, the present disclosure may refer to a representation set as a data basis. It should be understood that the term “basis” as used herein is not necessarily in the strictly mathematical sense of a basis.

Given the independent nature of each column within the precoder matrix (V) , they can be treated as distinct data units (v_i, i=1, 2…number of MIMO layers) for storage and management within the central database. Each column vector, representing a unique precoder or beamforming vector, is stored as a separate entry, allowing for granular access and manipulation of individual data points. This organizational structure facilitates efficient data processing and analysis, enabling the identification of specific precoders or the retrieval of subsets of data based on particular characteristics or conditions.

For each newly arriving data record, representing a single column beamforming vector or precoder, the central database undertakes a critical evaluation to determine its potential for sparse representation. The objective is to ascertain whether this new precoder can be effectively expressed as a sparse combination of existing data within the database. This evaluation process branches into two distinct scenarios:

Scenario 1: Sparse Representation is Possible: In the case, the new precoder can be approximately but accurately represented as a sparse combination of existing data points or basis vectors (also referred to as representative vectors) within the database. In some examples, the present disclosure may refer to a representative vector as a basis vector. It should be understood that the term “basis vector” as used herein is not necessarily in the strictly mathematical sense of a basis vector. This implies that the new precoder shares significant similarities and correlations with previously stored information. Consequently, the database incorporates this new data in its sparse form, storing only the relevant coefficients and the corresponding basis vectors (also referred to as corresponding representative vectors) , thereby minimizing storage requirements and optimizing efficiency.

Scenario 2: Sparse Representation is Infeasible: If the database fails to identify a suitable sparse combination to represent the new precoder, indicating that it possesses unique characteristics that distinguish it from existing data, a different approach is suggested. In such cases, the database may choose to store the new precoder in its raw form, or it may employ alternative dimensionality reduction techniques to create a new basis vector (also referred to as a new representative vector) that can be used to represent the new data along with other similar precoders that may be acquired in the future.

Regardless of the scenario, the process of storing precoder data in the central database begins with computing its sparse representation. This involves utilizing the "data basis" (also referred to as the representation set) derived from the database itself.

Scenario-1: When a precoder or beamforming vector, denoted as v_i, can be expressed in a sparse format, it signifies that it can be accurately reconstructed using a linear combination of a select few column vectors (also referred to as representative vectors) from the established common "data basis" (also referred to as the representation set) denoted as Ψ. This data basis (also referred to as the representation set) comprises a collection of representative precoder vectors that capture the essential characteristics of the data distribution within the central database.

The key to this sparse representation lies in the coefficient vector, s_i. This vector primarily consists of zeros, with only a few non-zero entries strategically placed. The indices of these non-zero elements correspond directly to the specific column vectors (also referred to as representative vectors) from the data basis (also referred to as the representation set) that contribute to the construction of v_i. The values of these non-zero entries represent the weights assigned to each basis vector (also referred to as representative vector) , indicating their relative contribution to forming the original precoder, as illustrated in FIG. 12.

As shown in the example of FIG. 12, a precoder matrix V may be expressed as a group of precoder vectors v_i (where i is the index of the MIMO layer, from 1 to r) . A given precoder vector v_i may be expressed or approximated as a linear combination of representative vectors, denoted Φ [1] to Φ [K] , stored in the central database. The representative vectors together may be considered to form the representation set, denoted Ψ. The coefficient vector s_i contains mostly zeros, with only a few non-zero entries (represented by dark bars in FIG. 12) corresponding to selected representative vectors. The selected representative vectors can be linearly combined, using the coefficients contained in the non-zero entries of the coefficient vector s_i, to approximately recover the precoder vector v_i. This means that each precoder vector v_i (and thus the precoder matrix V) may be stored in the central database using sparse representation.

This sparse representation offers a significant advantage in terms of storage efficiency. Instead of storing the entire high-dimensional precoder vector v_i, we only need to store the much smaller coefficient vector s_i and the associated data basis (also referred to as representation set) Ψ. This approach drastically reduces the storage footprint while preserving the ability to accurately reconstruct the original precoder when needed.

Scenario-2: When a new precoder vector, v_i, defies sparse representation using the existing data basis (also referred to as representation set) Ψ, it indicates that this vector possesses unique characteristics that set it apart from the current collection of basis vectors (also referred to as representative vectors) . Rather than discarding this valuable information, we can integrate it into the data basis (also referred to as representation set) itself, enriching its representational capacity and enhancing its ability to capture the diverse nature of precoder data.

One straightforward approach is to expand the data basis (also referred to as representation set) by appending v_i as a new column vector (also referred to as a new representative vector) . This effectively increases the dimensionality of the basis (also referred to as representation set) , providing a new dimension to represent precoders with similar characteristics to v_i. The corresponding sparse representation vector s_i (also referred to as a coefficient vector, not to be confused with the representative vector discussed above) for this new precoder would be a vector of zeros with a single '1' a t the end, signifying that v_i is directly represented by the newly added basis vector (also referred to as the newly added representative vector) , as shown in FIG. 13.

As shown in the example of FIG. 13, a precoder matrix V again may be expressed as a group of precoder vectors v_i. In this example, it is found that the representative vectors, denoted Φ [1] to Φ [K] , forming the representation set Ψ are non-representative of a given precoder vector v_i. That is, the precoder vector v_i cannot be approximated (to an acceptable level of accuracy) using a linear combination of the representative vectors. Accordingly, the precoder vector v_i may be added as a new representative vector, denoted Φ [K+1] , to the representation set Ψ. The representation set Ψ may thus be updated. Following the update, precoder vector v_i. can now be represented as a linear combination of the representative vectors, for example using a coefficient vector s_i, having only one non-zero entry (represented by a dark bar in FIG. 13) corresponding to the newly added representative vector Φ [K+1] as the selected representative vector.

This process of incorporating unique precoder vectors into the data basis (also referred to as the representation set) ensures that the central database continuously evolves and adapts to the ever-changing landscape of wireless channels and antenna configurations. It allows for the efficient representation of a wider range of precoders, enhancing the flexibility and effectiveness of AI/ML models for CSI compression.

To ensure the continued relevance and efficiency of the central database, we can implement a dynamic mechanism that tracks the usage and age of each basis vector (also referred to as representative vector) within the data basis (also referred to as representation set) . This involves associating a timestamp with each column vector (also referred to as representative vector) in the data basis (also referred to as representation set) , marking the time of its introduction or last update.

By monitoring these timestamps, the database can identify basis vectors (also referred to as representative vectors) that haven't been utilized by new arriving data for a significant period. If a specific basis vector (also referred to as representative vector) remains unused for an extended duration, it suggests that the data it represents may no longer be relevant to current channel conditions or deployment scenarios. In such cases, the database can automatically retire these outdated basis vectors (also referred to as outdated representative vectors) , ensuring that the data basis (also referred to as representation set) remains compact and reflects the evolving characteristics of the wireless environment.

This dynamic evolution of the data basis (also referred to as representation set) acknowledges the ever-changing nature of wireless communication networks. As network configurations, user equipment capabilities, and channel conditions evolve over time, so too should the data basis (also referred to as representation set) adapt to maintain its effectiveness and relevance. By removing outdated information and incorporating new representative data, the central database ensures its continued value as a foundation for training accurate and efficient AI/ML models for CSI compression.

The challenge of efficiently representing precoder data within the central database can be elegantly formulated as a mathematical optimization problem. The objective is to find an optimal coefficient vector s_i that minimizes a combination of two factors: reconstruction error and sparsity. The optimization problem may be formulated as follows:

Find a vector s_i (f_k) such that: α min (|| v_i (f_k) -Ψ (f_k) s_i (f_k) ||²) +β min (|s_i (f_k) |)

The first term, α×min (|| v_i –Ψs_i||²) , measures the reconstruction error, quantifying the difference between the original precoder vector v_i and its reconstruction using the data basis (also referred to as representation set) Ψ and the coefficient vector s_i. Minimizing this term ensures that the sparse representation accurately captures the essential information contained within the original precoder.

The second term, β×min (|s_i|) , promotes sparsity by minimizing the number of non-zero elements within the coefficient vector s_i. This encourages the use of a minimal set of basis vectors (also referred to as representative vectors) from Ψ to represent the precoder, thereby reducing storage requirements and improving efficiency.

The parameters α and β control the relative importance of these two objectives. Adjusting these parameters allows us to fine-tune the balance between reconstruction accuracy and sparsity, depending on the specific needs and constraints of the system.

A powerful tool for solving this optimization problem is the LASSO (Least Absolute Shrinkage and Selection Operator) algorithm. LASSO effectively identifies the most relevant basis vectors (also referred to as representative vectors) from Ψ while simultaneously minimizing the reconstruction error. Other suitable techniques for solving this optimization problem may be used.

To determine whether a sparse representation is feasible for a given precoder vector, a threshold is established. If the resulting L1-norm of the coefficient vector |s_i| falls below this threshold, it indicates that the precoder can be accurately represented using a sparse combination of existing basis vectors (also referred to as representative vectors) (Scenario 1, as illustrated in FIG. 12 for example) . Conversely, if |s_i| exceeds the threshold, it suggests that the precoder possesses unique characteristics that necessitate its inclusion in the data basis (also referred to as representation set) (Scenario 2, as illustrated in FIG. 13 for example) .

Recent data analysis from real-world 5G deployments has revealed a crucial characteristic of wireless channels: their inherent sparsity and spatial consistency. This means that high-dimensional MIMO signals, despite their complexity, often exhibit a significant degree of redundancy and predictability, particularly within localized regions. Channels in adjacent locations tend to share similar propagation characteristics, reflecting the common underlying physical environment and scattering properties.

This observation aligns with theoretical expectations and holds significant implications for AI/ML-based CSI compression. By leveraging localized data, models can effectively exploit the spatial consistency of channels, leading to improved compression ratios and enhanced reconstruction accuracy. This is because models trained on local data can better capture the nuances and correlations specific to that region, leading to more efficient (i.e. sparser) representations and better generalization within that specific area.

As we move towards 6G, with even higher dimensional channels and more complex antenna configurations, the sparsity and spatial consistency become even more pronounced. This presents a valuable opportunity to further optimize CSI compression techniques by incorporating localized learning and data-driven approaches.

Within the s vector space, which represents the sparse representation of precoder data, this spatial consistency manifests as a distinct distribution. Precoder vectors from nearby locations tend to cluster together, reflecting their shared characteristics and underlying correlations. This spatial distribution can be effectively captured and exploited by AI/ML models trained on localized datasets, leading to more efficient and accurate CSI compression. A generic diagram is illustrated in FIG. 14.

FIG. 14 illustrates an example BS 1402 in communication with a plurality of UEs. Notably, a first group of UEs 1404A may be grouped in one location and may be served by a first vendor (e.g., denoted UE-vendor-A) , while a second group of UEs 1404B may be grouped in another location and may be served by a second vendor (e.g., denoted UE-vendor-B) . Each UE-vendor may be a respective network entity, which may have modules similar to those illustrated in FIG. 4. FIG. 14 further illustrates example coefficient vectors in a simplified s-domain (or sparse domain) 1406. The coefficient vectors, that is the s_i vectors, corresponding to the first group of UEs 1404A are shown without hatching, and the coefficient vectors corresponding to the second group of UEs 1404B are shown with hatching. The spatial correlation of each group of UEs 1404A, 1404B may result in their corresponding precoder vectors being correlated, which may be reflected in their corresponding coefficient vectors being clustered together in the s-domain 1406.

The expandable nature of the "data basis" (also referred to as representation set) Ψ offers a crucial advantage for preserving the spatial relationships and topological properties of precoder vectors within the sparse domain, represented by the s_i vectors. This capability is essential or at least useful for maintaining spatial consistency and separability, which are critical aspects for both NW and UE vendors when retrieving data from the central database.

Spatial Consistency: By incorporating new precoder vectors into the data basis (also referred to as representation set) , we ensure that the spatial correlations between nearby locations are captured and preserved. This allows the models to effectively learn and exploit these relationships, leading to improved prediction accuracy and better generalization within specific regions.

Spatial Separability: The data basis (also referred to as representation set) also facilitates the separation of precoders from distinct locations or environments. This enables vendors to selectively retrieve data that is relevant to their specific needs, avoiding the unnecessary processing of irrelevant or out-of-distribution data. For example, a UE vendor might be interested in training a model specifically for urban environments, while an NW vendor might require data for a rural deployment. The data basis allows for efficient retrieval of these specific subsets of data, optimizing training efficiency and model performance.

Benefits for NW and UE Vendors: This ability to preserve spatial relationships within the sparse domain offers several benefits for both NW and UE vendors:

Improved Model Performance: By leveraging spatially consistent data, vendors can train models that are better adapted to specific environments and exhibit higher accuracy in predicting precoders for nearby locations.

Reduced Complexity: The ability to selectively retrieve relevant data based on spatial location reduces the amount of data that needs to be processed and stored, leading to lower computational complexity and improved efficiency.

Enhanced Generalization: Models trained on spatially diverse data from the central database can better generalize to new and unseen locations, improving their overall robustness and adaptability.

The expandable data basis (also referred to as representation set) , therefore, plays a crucial role in preserving spatial information within the sparse domain, enabling efficient and effective utilization of precoder data for training AI/ML models for CSI compression.

While data-driven approaches offer the flexibility of an expandable basis (also referred to as representation set) , methods relying on fixed basis like the Discrete Fourier Transform (DFT) and the Cosine Discrete Transform (CDT) face limitations when dealing with irregular antenna arrays and diverse channel conditions. These fixed basis are predetermined and lack the ability to adapt or expand to incorporate new information or unique characteristics of the data.

This inflexibility becomes a significant drawback when dealing with irregular antenna configurations, where the spatial relationships between antenna elements deviate from the uniform antenna array assumptions underlying DFT and CDT. As a result, these transforms struggle to accurately capture the complex spatial correlations and nuances present in such scenarios, leading to suboptimal performance in tasks like beamforming and CSI compression.

This limitation is evident in the challenges faced by the DFT-based Type II codebooks introduced in Release 18 of 5G NR. While these codebooks offer efficient CSI compression for uniform antenna arrays, they are ill-equipped to handle irregular antenna layouts, leading to performance degradation and reduced accuracy.

The lack of extensibility in fixed basis restricts their ability to adapt to the evolving landscape of wireless communication systems. As 6G embraces more diverse antenna configurations and explores the potential of non-conventional arrays, the need for data-driven approaches with expandable bases becomes increasingly crucial for achieving optimal performance and flexibility in CSI compression and other critical tasks.

The flexibility of the central database as disclosed herein allows for diverse data retrieval strategies, catering to the specific needs and priorities of both network and UE vendors. While NW vendors may opt for a more comprehensive approach, extracting a larger data basis (also referred to as representation set) to encompass a broader range of scenarios and configurations, UE vendors often prioritize the advantages of localized data, focusing on smaller, more specific subsets of the data basis (also referred to as representation set) .

The example in FIG. 15 exemplifies this concept. In this example, the BS 1402 is in communication with a first group of UEs 1404A (serviced by a first vendor UE-vendor-A) and a second group of UEs 1404B (serviced by a second vendor UE-vendor-B) , as discussed previously. UE-vendor-A and UE-vendor-B each extract distinct sub-data bases (also referred to as subsets of the representation set 1502) (e.g., UE-vendor-A extracts the representation subset 1504A and UE-vendor-B extracts the representation subset 1504B) , reflecting their specific areas of interest within the overall area. This localized approach allows them to capitalize on the spatial consistency and correlations present within their respective regions, leading to more efficient and accurate models for their target deployments.

In contrast, the BS 1402 is concerned with channel conditions over a broader region 1408. Thus, the NW vendor (which services the BS 1402) retrieves the entire data basis (also referred to as the representation set 1502) , encompassing the data relevant to both UE-vendor-A and UE-vendor-B, as well as additional data from other regions or scenarios. This broader perspective allows the network vendor to develop models with greater generalizability and adaptability, capable of handling a wider range of channel conditions and user equipment configurations.

Localized data not only allows for tailored model training but also presents opportunities for further optimization of the sparse representation, specifically for UE vendors. By focusing on a specific region or subset of the data basis, UE vendors can achieve a more compact and efficient representation of precoders within their domain of interest.

An example is illustrated in FIG. 16. For instance in FIG. 16, in the case of UE-vendor A, a localized sparse vector subspace that contains a coefficient vector s_{UE_A} is derived by extracting the relevant portions from the global s vector space 1406 (also referred to as the global s-domain 1406) . This extraction process involves selecting the entries in s that correspond to the representative vectors within the sub-data basis (also referred to as representation subset 1504A) used by UE-vendor A. The resulting subspace retains the essential information needed to represent precoders within the specific region of interest while discarding irrelevant components, leading to a more compact and efficient representation. Consider a selected coefficient vector s_{UE_A} that corresponds to the representation subset 1504A (i.e., s_{UE_A} only has non-zero entries corresponding to the representative vectors contained in the representation subset 1504A) . s_{UE_A} can be stored at the central datacenter in an expanded vector form 1602 that includes entries corresponding to all the representative vectors contained in the full representation set 1502. For more efficient use of resources, s_{UE_A} can be represented in a more compact vector form 1604A corresponding to only the representative vectors contained in the representation subset 1504A. This means that a coefficient vector within the region of interest can be represented in a more compact vector form by the UE-vendor A, compared to the representation of the same coefficient vector at the central datacenter.

To ensure proper data retrieval and reconstruction, the UE vendor records the specific extraction method used to generate s_{UE_A} in the more compact form 1604A. This information (denoted as Mapping_A) , along with the sub-data basis (also referred to as the representation subset 1504A) itself, enables the UE vendor to effectively utilize the sparse representation for training their models and accurately reconstructing precoders within their target region. As will be discussed further below, this stored mapping information may be used to convert a coefficient vector from the compact vector form 1604A (which corresponds to the representation subset 1504A) back to the expanded vector form 1602 (which corresponds to the full representation set 1502) . Similarly, UE-vendor B may use another mapping (denoted as Mapping_B) to map a coefficient vector s_{UE_B}, selected to be representable using the representation subset 1504B used by UE-vendor B, from the expanded vector form 1602 to a more compact vector form 1604B. It may be noted that, because the representation subset 1504A used by UE-vendor A and the representation subset 1504B used by UE-vendor B can be different in size, the compact vector forms 1604A and 1604B may also be different in size.

FIG. 17 illustrates an example of training a UE-side encoder (that is, training a neural network to perform a precoder encoding task, which will be deployed at a UE) . This training may be performed by any network entity, such as at a UE-vendor. In FIG. 17, the training process for a UE-side encoder (in this example, training performed by UE-vendor A) in the context of CSI compression using a central database and sparse representation can be viewed as an optimization problem within the Minimum Mean Square Error (MMSE) framework. The objective is to minimize the difference between the original sparse representation (s_{UE_A}) and the output of the encoder network (s_{UE_A}=ENC (v; β) ) when provided with the synthesized precoder data (v) and the corresponding sparse representation (s_{UE_A}) .

Here's a breakdown of the process illustrated in the example of FIG. 17:

1. Acquiring Data Basis and Sparse Representations: The UE vendor A first retrieves the relevant "data basis" (also referred to as representation subset) (Ψ_UE-A 1504A as aforementioned a subset of Ψ; note that the (f_k) notation may be omitted for convenience) and a set of data samples in their sparse representation form (or compact vector form 1604A) (s_{UE_A} as aforementioned a subset of s; in some examples, s_{UE_A} may be simplified to s_A) from the central database. These form the foundation for training the encoder model 1702A (which may be denoted ENC_{UE_A}) . The information about mapping from s to s_{UE_A} and Ψ to Ψ_UE-A is recorded as mapping A. The mapping information may be in the form of a transformation matrix, for example, and may be recorded at the central datacenter.

2. Synthesizing Precoder Data: Using the retrieved representation subset (Ψ_UE-A) and the sparse representations of the coefficient vectors (s_{UE_A}) , the UE vendor synthesizes the corresponding precoder vectors (v) . This involves multiplying the data basis matrix (also referred to as the representation subset matrix) with the sparse representation vectors to reconstruct the original precoder data: v = Ψ_UE-A s_UE-A

3. Training the Encoder Network: The synthesized data pairs, consisting of the precoders (v) and their corresponding sparse representations of the coefficient vectors (s_{UE_A}) , are used to train the encoder network (s’_{UE_A} =ENC (v; β) , where s’_{UE_A} denotes the output of the encoder to differentiate from the ground-truth s_{UE_A}) . The training process aims to optimize the encoder's parameters (β) to minimize the mean squared error between the original sparse representation (s_{UE_A}) and the encoder's output when given the synthesized precoder data and its sparse representation. MMSE Optimization Framework: The training objective can be mathematically expressed as: min (|s_{UE_A}-ENC (Ψ_UE-A s_{UE_A}; β) |²) or more simply min (|s_{UE_A}-s’_{UE_A}|²) (note that this is equivalent to min (|s’_{UE_A}-s_{UE_A}|²) ) . This equation reflects the goal of minimizing the discrepancy between the original and encoded sparse representations. The specific method employed to achieve this optimization is left to the UE vendor's discretion. Deep neural networks are a common choice, allowing the model to learn the optimal parameters (β) through a data-driven approach. However, other optimization techniques or machine learning algorithms may also be explored.

This framework offers UE vendors the flexibility to design and optimize their encoder models based on their specific requirements and priorities. They can choose the model architecture, training algorithms, and optimization techniques that best suit their needs, while still ensuring compatibility with the standardized data basis and sparse representation format used within the central database via recorded mapping information.

FIG. 18 illustrates an example of training a NW-side encoder (that is, training a neural network to perform a precoder decoding task, which will be deployed at a network node, such as a BS or TRP) . This training may be performed by any network entity, such as at a NW-vendor (which may have modules similar to those shown in FIG. 4) . Unlike the UE-side encoder, the network-side decoder operates within a slightly different context, as illustrated in FIG. 18. While the data basis (also referred to as the representation set 1502) (Ψ_NW) employed at the network side is primarily composed of representative vectors that are not necessarily strictly orthogonal to each other, its expandable nature introduces a margin of compression.

Consequently, the role of the network-side decoder 1802 (v=DEC_NW (s; θ) ) extends beyond mere reconstruction. It also acts as a compressor, further refining the sparse representation received from the UE and recovering the original precoder data with minimal loss of information.

During training, the original precoder data is synthesized from coefficient vectors s sampled from the dataset maintained by the central datacenter. For example, as shown in FIG. 18, coefficient vectors may be obtained (in the original expanded vector form 1602) , then multiplied by the representation set Ψ_NW to obtain the precoder vector v. The optimization objective for the decoder network is to minimize the mean squared error between its output (which may be denoted v’) and the original precoder data. This can be expressed mathematically as:

min (|DEC_NW (s; θ) -v|²)

where s represents the coefficient vector (which, during inference, would be the sparse representation received from the UE; and during training is the sampled training data) , θ denotes the decoder's parameters, and v is the original precoder vector.

Similar to the UE-side encoder, the specific approach for training the decoder network is left to the NW vendor's discretion. Neural networks offer a powerful and flexible solution, allowing the model to learn the optimal parameters (θ) through a data-driven process. However, other optimization techniques or machine learning algorithms may also be explored based on the NW vendor's specific requirements and preferences.

FIG. 19 illustrates an example of how the trained encoder model and trained decoder model may be used in inference. After training (e.g., as described above with respect to FIGS. 17 and 18) , the trained encoder model may be deployed to a UE 110 (e.g., the encoder model trained by UE-vendor A, denoted ENC_{UE_A} 1702A may be deployed to a UE_A 110A serviced by UE-vendor A; and the encoder model trained by UE-vendor B, denoted ENC_{UE_B} 1702B may be deployed to a UE_B 110B serviced by UE-vendor B) and the trained decoder model (e.g., denoted DEC_NW 1802) may be deployed to a network node 170 such as a BS or T-TRP. Mapping information (e.g., Mapping_A and Mapping_B) may be provided from the central datacenter to the network entity (e.g., via control signalling or other suitable signalling) .

The UE_A 110A and UE_B 110B may be similar to the ED 110 as previously described, including the units and modules shown in FIG. 4. For simplicity, such units and modules are omitted from FIG. 19. The trained encoder model ENC_{UE_A} 1702A may be provided to the UE_A 110A by the UE-vendor A after training as previously described with respect to FIG. 17. In some examples, the encoder model ENC_{UE_A} 1702A may be a submodule of any of the previously described units or modules of the UE_A 110A. For example, the encoder model ENC_{UE_A} 1702A may be implemented in the logic of a processing module of UE_A 110A. In general, the encoder model ENC_{UE_A} 1702A may be implemented as software, hardware or a combination of software and hardware (e.g., implemented as an integrated circuit and/or by software instructions executed by a processor, among other possibilities) .

The UE_A 110A may receive, via a receiving module, a reference signal that the UE_A 110A uses to calculate the channel matrix H, which may then be decomposed into the precoder matrix V. Each vector of the precoder matrix V, is referred to as a precoder vector v_A, The UE_A 110A may generate a precoder vector v_A and use the trained encoder model ENC_{UE_A} 1702A to encode the precoder vector v_A into a coefficient vector (which may be denoted as s_{UE_A} or more simply s_A) that in this case is in a compressed vector form 1604A. The coefficient vector may be used as a precoding matrix indicator (denoted PMI_A) . This PMI_A (e.g., the coefficient vector s_{UE_A}in compressed vector form 1604A) may be transmitted to the network node 170 using suitable signalling.

The network node 170 may be similar to the T-TRP 170 previously described with reference to FIG. 4, and may include the units and modules as discussed with respect to that figure. At the network node 170, the PMI_A (that is, the coefficient vector s_{UE_A}in compressed vector form 1604A) may be received via a receiving module. The network node 170 has stored, in a memory, mapping information (e.g., Mapping_A) that is used to transform or map the coefficient vector from the compressed vector form 1604A to the expanded vector form 1602. This enables the coefficient vector, in the expanded vector form 1602, to be decoded using the trained decoder model DEC_NW 1802, to recover the precoder vector v’_A. The trained decoder model DEC_NW 1802 may be provided to the network node 170 by the NW-vendor after training as previously described with respect to FIG. 18. In some examples, the decoder model DEC_NW 1802 may be trained by the network node 170 itself. In some examples, the decoder model DEC_NW 1802 may be a submodule of any of the previously described units or modules of the network node 170. For example, the decoder model DEC_NW 1802 may be implemented in the logic of a processing module of the network node 170. In general, the decoder model DEC_NW 1802 may be implemented as software, hardware or a combination of software and hardware (e.g., implemented as an integrated circuit and/or by software instructions executed by a processor, among other possibilities) .

Similar to the operations of the UE_A 110A, the UE_B 110B may generate a precoder vector v_B and use the trained encoder model ENC_{UE_B} 1702B to encode the precoder vector v_B into a precoding matrix indicator (denoted PMI_B) that in this case is a coefficient vector s_{UE_B} in a compressed vector form 1604B. This may be transmitted to the network node 170 using suitable signalling. At the network node 170, the mapping information (e.g., Mapping_B) is used to transform or map the coefficient vector from the compressed vector form 1604B to the expanded vector form 1602, which is in turn decoded using the trained decoder model DEC_NW 1802, to recover the precoder vector v’_B.

The proposed approach of utilizing sparse representation and a representation set (also referred to as a data basis, although not intended to be limited to the strictly mathematical meaning of “basis” ) for CSI compression offers a robust foundation for achieving interoperability between models developed by different vendors. By basing the mapping of the PMI (e.g., coefficient vectors in compressed vector form) from different UE vendors’ encoder back onto the same expanded vector form (as mapping-A from s_{UE_A} to s and mapping-B from s_{UE_B} to s are well recorded) , which encodes the essential information about the precoder using a standardized format, we establish a common language that transcends specific model architectures or training methods.

This approach empowers UE vendors to leverage the advantages of localized data. By focusing on a specific subset of the data basis relevant to their target deployment region, they can develop smaller and more efficient encoder models. This translates to more concise PMI mappings, reducing overhead and improving communication efficiency.

Furthermore, the proposed framework imposes no restrictions on the specific models or training methods employed by either UE or NW vendors. This allows for flexibility and innovation, enabling vendors to explore and implement the AI/ML techniques that best suit their needs and expertise. This is because of the standardized s domain.

To facilitate seamless integration and operation, a certain degree of over-the-air signaling is introduced. These signaling mechanisms enable the exchange of information between the UE and the network, such as the data basis, sparse representations, and model updates.

Building upon the previous discussions, here's a design for the protocol and signaling mechanisms at the physical a nd MAC layers to facilitate the integration and operation of AI/ML-based CSI compression with a central database and sparse representation:

System information and UE capability signaling may include:

· Data Basis Indication: The network broadcasts information about the available data basis (e.g., Ψ_NW) through system information messages. This can include:

○ Data basis ID: A unique identifier for the data basis.

○ Basis vector descriptions: Parameters like dimensionality, quantization methods, and potentially a subset of representative basis vectors.

· UE Capability Reporting: The UE reports its supported AI/ML functionalities and capabilities, including:

○ Supported data basis IDs: Indicate the data bases the UE can work with.

○ Model capabilities: Report the complexity and performance characteristics of supported encoder models.

○ Local data availability: Optionally indicate the availability of locally collected data for model fine-tuning.

Data collection and registration may include:

· Triggering Data Collection: The network sends a dedicated RRC message (e.g., AI_DataCollectionRequest) to the UE, specifying the CSI-RS resources, data collection duration, and reporting mode.

· UE Data Collection: The UE performs channel estimation on the specified CSI-RS resources and extracts the precoder matrix (V) . It then computes the sparse representation (s) using the data basis and reports both s and ancillary information to the network.

Model update and synchronization may include the following. In this context, the “model” refers to the representation set maintained at the central database and not the encoder/decoder model:

· Model Update Indication (also referred to as Representation Set Update Indication) : The network informs the UE of a new model update via an RRC message (e.g., AI_ModelUpdate) , including the model ID (also referred to as the representation set ID) and transfer method.

· Model Parameter Transfer (also referred to as Representation Set Parameter Transfer) : Depending on the chosen method, the network transfers the model parameters (also referred to as the representation set parameters) to the UE through RRC signaling or the user plane data channel, ensuring reliable delivery with segmentation and error correction mechanisms.

· Synchronization: The network can optionally provide a synchronization signal (e.g., via RRC or MAC control element) to ensure that the UE and gNB are using the same data basis and model versions (also referred to as representation set version) .

Physical Layer Enhancements may include:

· CSI-RS Configuration: The network can configure the CSI-RS resources dynamically based on feedback from the UEs and the gNB, ensuring sufficient measurements for accurate channel estimation and data collection. This feedback mechanism can include information such as channel conditions, interference levels, and signal quality metrics reported by the UEs, for example. By dynamically adjusting the CSI-RS configuration, the network can optimize resource allocation and improve overall system performance.

· Uplink Channel Selection: The UE and gNB can dynamically select the appropriate uplink channel for data reporting based on payload size and channel conditions. Feedback from the UEs regarding current channel quality, signal-to-noise ratio (SNR) , and other relevant metrics can be used to make informed decisions about the most suitable channel for transmitting data. This dynamic selection process helps to ensure that data is transmitted efficiently and with minimal errors, adapting to changing network conditions in real-time.

· Hybrid Automatic Repeat Request (HARQ) : Implement HARQ mechanisms to guarantee reliable delivery of CSI data and model updates. HARQ combines error detection, error correction, and retransmission protocols to enhance data reliability. In this context, feedback from the receiver (e.g., either the UE or gNB) is useful for indicating whether a packet was received correctly or if a retransmission is necessary. This feedback loop may allow for continuous monitoring and correction of transmission errors, thereby maintaining the integrity and accuracy of the CSI data and model updates. The feedback mechanism may involve a continuous loop where the UEs periodically report back to the gNB with updated CSI, including parameters such as link quality, interference, and signal strength. The gNB analyzes this feedback to adjust the CSI-RS configurations and optimize resource allocation dynamically. This helps to ensure that the network adapts to real-time conditions and maintains optimal performance. Uplink channel selection may benefit from a real-time feedback system where UEs send regular updates on channel conditions. This data may enable the gNB to make timely decisions about the best uplink channel for data reporting. The process may involve selecting channels that can handle the current data payload efficiently, reducing latency and improving data throughput. HARQ mechanisms typically rely heavily on feedback for error correction. When a data packet is received, the receiver checks for errors and sends an acknowledgment (ACK) if the packet is error-free or a negative acknowledgment (NACK) if errors are detected. The sender uses this feedback to either proceed with the next packet or retransmit the erroneous packet. This process helps to ensure high data reliability and efficient use of network resources. By leveraging feedback from UEs, the network can continuously optimize CSI-RS configurations, uplink channel selection, and HARQ processes. This proactive approach helps to reduce the impact of adverse conditions such as interference and poor signal quality, which may help to enhance the overall efficiency and robustness of the network.

MAC Layer Enhancements may include:

· Resource Allocation: The MAC layer scheduler should prioritize data collection and model update transmissions while ensuring fairness and efficiency for other traffic.

· Scheduling Algorithms: Adaptive scheduling algorithms can be employed to dynamically allocate resources based on real-time channel conditions and traffic demands, ensuring timely and efficient data exchange.

· Data Aggregation: The gNB can aggregate data reports from multiple UEs to reduce overhead and improve efficiency before sending them to the central database.

Security Enhancements may include:

· Secure Boot and Secure Storage: Implement secure boot mechanisms and secure storage solutions on UEs and gNBs to protect the integrity of AI/ML models and prevent unauthorized modifications.

· Authentication and Authorization: Utilize strong authentication and authorization protocols to ensure that only trusted entities can access and contribute data to the central database and receive model updates.

· Secure Communication Channels: Employ encryption and integrity protection mechanisms for all communication between the UE, gNB, and the central database.

This protocol and signaling design provide a comprehensive framework for integrating AI/MLbased CSI compress ion with a central database and sparse representation into the existing NR framework. This framework ensures interoperabilit y, efficiency, and security while offering flexibility for vendors to innovate and optimize their implementations.

Reference is now made to FIG. 20.

FIG. 20 illustrates an example apparatus, which may be an implementation of a datacenter 2000 (also referred to as a database) as previously described. The datacenter 2000 may be a central datacenter 2000 of a wireless communication system (e.g., the communication system 100 of FIG. 1) . The datacenter 2000 may be a network entity of the wireless communication system in addition to the devices shown in FIG. 1 (e.g., in addition to the ED 110, T-TRP 170 or NT-TRP 172) ; for example, the datacenter 2000 may be a server or computing system at the core network 130 and may be in capable of communication with any of the ED 110, T-TRP 170 and/or NT-TRP 172. The datacenter 2000 may include modules similar to the modules shown in FIG. 4, however for simplicity not all units or modules of the datacenter 2000 are shown in FIG. 20. The datacenter 2000 may include a receiving unit or module 2002 (also referred to as a receiving module 2002 for simplicity) , a transmitting unit or module 2004 (also referred to as a transmitting module 2004 for simplicity) and a processing unit or module 2006 (also referred to as a processing module 2006 for simplicity) , the functions of which have been previously discussed with respect to FIG. 4. The datacenter 2000 may also include a memory unit storing a training dataset 2008 and the same or additional memory unit storing a representation set 2010.

The receiving module 2002, transmitting module 2004 and/or processing module 2006 may be implemented as software, hardware or a combination of software and hardware. For example, the receiving module 2002, transmitting module 2004 and/or processing module 2006 may be a circuit such as an integrated circuit. Examples of an integrated circuit includes a programmed FPGA, a GPU, or an ASIC. For instance, the receiving module 2002, transmitting module 2004 and/or processing module 2006 may be at least partly implemented as logic, such as a logical function performed by a circuit, by a portion of an integrated circuit, or by software instructions executed by a processor, among other possibilities. The training dataset 2008 and the representation set 2010 may be stored in one or more physical memory units of the datacenter 2000, such as a RAM, ROM, hard disk, optical disc, and any other suitable volatile and/or non-volatile storage and retrieval device (s) .

The datacenter 2000 may receive communications via the receiving module 2002, such as a precoder vector from an ED 110 that is a UE, requests for training data sampled from the training dataset 2008, etc. The datacenter 2000 may transmit communications via the transmitting module 2004, such as update information about the representation set 2010, requested data samples, etc. The processing module 2006 may perform operations to carry out the various functions of the datacenter 2000 as disclosed herein, such as adding data to the training dataset 2008 and updating the representation set 2010, among others.

FIG. 21 is a flowchart illustrating an example method 2100 for maintaining a training dataset and a representation set. The method 2100 may be implemented at a datacenter of a wireless communication system (e.g., the datacenter 2000 of FIG. 20) or by any suitable network entity. The method 2100 may be performed by a functional unit, such as a processor or any other suitable unit. For simplicity, the method 2100 will be described in the context of operations at a datacenter, however this is not intended to be limiting.

At the start of the method 2100, a UE may broadcast or report its availability to collect data. For simplicity, the following description with refer to a single UE however it should be understood that this is not intended to be limiting.

Optionally, at an operation 2102, a set of one or more representative vectors that defines a representation set (e.g., the representation set 2010 of FIG. 20) is maintained in a memory 2020 of the datacenter. A dataset (e.g., the training dataset 2008 of FIG. 20) may also be maintained in a memory 2020 of the datacenter.

At an operation 2104, data including a precoder vector associated with a UE is received. In some examples, data collection by the UE may be initiated by the datacenter or other network entity. For example, a dedicated RRC message may be sent to the UE to trigger data collection, and the data may be received from the UE in response to the control message. The control message may indicate relevant parameters such as an indication of the CSR-RS resources the UE should use for channel estimation, an indication of the duration or number of time instances over which the UE should collect data, an indication of how the UE should report data (e.g., periodically, upon completion, based on specific events such as handover, change in channel conditions, etc. ) and/or an indication of any compression the UE should use to compress the collected data, for example.

Data from the UE may be received over existing uplink channels (e.g., PUCCH or PUSCH) or via a dedicated control channel, for example. The received data from the UE may also include ancillary information related to the data collection by the UE for generating the associated precoder vector. For example, the ancillary information may include information about a frequency sub-band used in the data collection, information about a resource block used in the data collection, information about an antenna configuration used in the data collection, and/or information about a geographical location during the data collection, for example. The data received from the UE may also include metrics indicating the quality or reliability of the collected data.

At an operation 2106, the precoder vector is converted to a coefficient vector. The coefficient vector represents a linear combination of one or more selected representative vectors from the representation set defined by one or more representative vectors. The coefficient vector may thus be a sparse representation of the precoder vector based on the representation set. The coefficient vector is useable for training a neural network to perform a precoder vector encoding task or a precoder vector decoding task, as discussed above.

Converting the precoder vector to the coefficient vector may include checking whether the precoder vector can be sparsely represented using the current representation set. This may include computing a candidate vector to represent the precoder vector as a linear combination of one or more candidate representative vectors from the current representation set. The candidate vector is then compared to a sparsity threshold. For example, the magnitude of the candidate vector (e.g., |s_i|, where s_i is the candidate vector) may be compared to a predefined threshold value. If the magnitude is less than or equal to the threshold value, then the sparsity threshold may be satisfied; if the magnitude is greater than the threshold value, the candidate vector may fail to satisfy the sparsity threshold. If the candidate vector satisfies the sparsity threshold this means the precoder vector can be sparsely represented as a coefficient vector using the current representation set. The candidate vector that satisfies the sparsity threshold is thus a coefficient vector that represents the precoder vector.

If the candidate vector fails to satisfy the sparsity threshold, this means the precoder vector cannot be sparsely represented as a coefficient vector using the current representation set. The representation set may be updated by including the precoder vector as a new representative vector in the representation set. Then the precoder vector can be converted to a coefficient vector after the representation set is updated. For example, if the precoder vector is added as a new k+1 representative vector in the updated representation set, then the precoder vector may be represented as a sparse coefficient vector where the only non-zero entry is the (k+1) -th entry.

It may be noted that after the representation set is updated (e.g., by adding the precoder vector as a new representative vector or by other operations disclosed herein) , the datacenter may send a control message (e.g., RRC message) to inform the UE and other network entities of the update. The model update message may include a unique identifier for the updated representation set, as well as information for the representative vectors such as dimensionality, quantization methods, and other parameters.

At an operation 2108, the datacenter may store the coefficient vector (generated at operation 2106) in a dataset (e.g., the training dataset 2008) . The coefficient vector may be included in a dataset that is useable for training a neural network to perform a precoder vector encoding task or a precoder vector decoding task. For example, the coefficient vector may be part of the data used for performing the training discussed above with respect to FIG. 17 or with respect to FIG. 18. If the data received from the UE at operation 2104 included ancillary data, the ancillary data may be stored in association with the corresponding coefficient vector in the dataset. This ancillary data may be useful for sampling data from the dataset for a spatial characteristic of interest. For example, the coefficient vector may be stored with ancillary data about a particular geographical location associated with the data collection. Then the coefficient vector may be included in sampled data where the particular geographical location is of interest.

In some examples, the datacenter may generate one or more synthetic coefficient vectors based on one or more coefficient vectors converted from real-world precoder vectors. For example, the real-world coefficient vectors may be clustered and a synthetic coefficient vector may be generated based on a cluster. A synthetic coefficient vector generated in this way may help to augment real-world data in the dataset, and may the synthetic coefficient vector may be useable for training a neural network similar to real-world coefficient vectors. When sampled data is provided to a NW-vendor or UE-vendor for training a neural network, the data may include synthetic coefficient vectors as well as real-world coefficient vectors.

At an operation 2110, the datacenter may provide coefficient vectors sampled from the dataset for training a neural network to perform a precoder vector encoding task (e.g., the training illustrated in FIG. 17) or a precoder vector decoding task (e.g., the training illustrated in FIG. 18) . As previously discussed, the datacenter may sample coefficient vectors from across the entire dataset or from a selected sub-dataset. For example, if a NW-vendor requests data for training a decoder model that is intended for a network entity (e.g., a T-TRP 170) that services a large geographical area, the coefficient vectors provided by the datacenter to train the decoder model may be sampled from across the entire dataset. In another example, if a UE-vendor requests data for training an encoder model that is intended for UEs in a particular spatial region of interest, the coefficient vectors provided by the datacenter to train the encoder model may be sampled from a selected sub-dataset that contains coefficient vectors associated with the spatial region of interest (e.g., based on the ancillary information associated with each coefficient vector) . The sub-data may correspond to a representation subset that includes only representative vectors that represent the sub-dataset and that omits at least one representative vector that is not used to represent the sub-dataset.

In some examples, the datacenter may transmit coefficient vectors (e.g., sampled from a sub-dataset) in a compact vector form that is smaller than the original expanded vector form used to store the coefficient vectors in the dataset. For example, if the coefficient vectors are sampled from a sub-dataset corresponding to a representation subset that omits a particular representative vector, then the compact vector form may be obtained by omitting entries of the coefficient vectors that correspond to the omitted representative vector (since the entries corresponding to the omitted representative vector should all be zero) .

Mapping information that is used to map (or transform) from the compact vector form to the expanded vector form may be stored. The mapping information is also used to map (or transform) between the representation subset and the representation set. For example, the mapping information may be a transformation matrix. Different UE-vendors may use different sub-datasets to train their encoder models and thus have different mapping information. This mapping information may also be used by a network entity to map the compact vector form to the expanded vector form prior to decoding a precoder vector, as discussed above. The mapping information may be transmitted (e.g., in a control message) to the network entity, for example.

The datacenter may perform operations to ensure the representation set is up-to-date and contains information to represent the dynamic, real-time characteristics of wireless channels. For example, the datacenter may update the representation set by including a unique precoder vector, that cannot be sparsely represented using the current representation set, as a new representative vector in an updated representation set. In another example, the datacenter may track the usage and/or age of each representative vector in the representation set. Each representative vector may be associated with a timestamp (or other time indicator) indicating the time that representative vector was added to the representation set or the time that representative vector was last used to represent a precoder vector. Based on this tracked usage and/or age, an out-of-date representative vector (e.g., having a timestamp that is older than a preset time threshold, such as 1 day) may be removed from the representation set.

Following updating of the representation set, the datacenter may send a control message to inform the UE and other network entities of the updated representation set. For example, the control message may include an updated identifier of the updated representation set and/or information about the representative vectors such as the number of representative vectors, dimensionality of representative vectors, index of an added or removed representative vector, etc.

FIG. 22 is a flowchart illustrating an example method 2200 for training a neural network to perform a precoder vector encoding task. The method 2200 may be implemented at a network entity that is a UE-vendor, for example, or other suitable network entity that may deploy the trained neural network at a UE. The method 2200 may be performed by a functional unit, such as a processor or any other suitable unit.

At an operation 2202, a selected subset of representative vector (s) selected from the representation set (e.g., defined by the set of all representative vectors maintained at the datacenter) is obtained by the UE-vendor. The selected representation subset is smaller than the representation set and may include only representative vector (s) associated with a characteristic of interest (e.g., spatial characteristic of interest, such as a particular geographical region) . The representation subset may contain representative vectors that together are sufficient to sparsely represent precoder vectors associated with the characteristic of interest (e.g., able to sparsely represent precoder vectors generated by UEs within a particular geographical region) .

For example, the UE-vendor may send a request to the datacenter for representative vector (s) associated with characteristic (s) of interest. The datacenter may select representative vector (s) from the representation set that are associated with the characteristic (s) of interest (e.g., based on the ancillary data stored in association with each representative vector) and transmit the selected representative vector (s) to the UE-vendor in a control message or using other signalling. The representation subset defined by the selected representative vector (s) may be assigned a unique identifier.

At an operation 2204, the UE-vendor obtains coefficient vectors. Each coefficient vector represents a linear combination of one or more selected representative vectors from the representation subset (obtained at operation 2202) .

For example, the UE-vendor may send a request to the datacenter for coefficient vectors associated with characteristic (s) of interest. The datacenter may sample coefficient vectors from the sub-dataset associated with the characteristic (s) of interest (e.g., based on the ancillary data stored in association with each coefficient vector) and transmit the sampled coefficient vectors to the UE-vendor in a control message or using other signalling. The sampled coefficient vectors can each be converted to a respective precoder vector using only representative vectors contained in the representation subset. The sampled coefficient vectors may be transmitted to the UE-vendor in a compact vector form corresponding to the representation subset (and smaller than the expanded vector form corresponding to the representation set) as discussed above.

At an operation 2206, the UE-vendor uses the obtained coefficient vectors to train a neural network to perform a precoder vector encoding task. To train the neural network, a coefficient vector is converted to the corresponding precoder vector using the representation subset (e.g., by multiplying the coefficient vector with the representation subset to obtain the corresponding precoder vector) . Then the precoder vector and corresponding coefficient vector is a data pair that is used to train the neural network. For example, the precoder vector is provided as input to the neural network and an output generated by the neural network is obtained. The neural network may then be trained to minimize an error between the output generated by the neural network and the coefficient vector corresponding to the precoder vector, using suitable machine learning techniques (e.g., to minimize the MMSE) . In some examples, the neural network may be trained using a transfer learning approach.

Transfer learning involves taking an existing pre-trained model, which has been trained on a different but related task, and further training it using data specific to a target task. For example, a pre-trained model that may be fine-tuned, using the obtained coefficient vectors, to perform the precoder vector encoding task could be a general-purpose encoder/decoder neural network that has been trained on a large dataset for similar purposes such as image recognition, natural language processing, or even a preliminary version of precoder vector encoding.

The further training using the coefficient vectors adapts this pre-trained model to the specific nuances of the precoder vector encoding task. This approach leverages the knowledge the model has already acquired during its pre-training phase, which can help to accelerate the training process and/or improve the model's performance. Unlike training from scratch, where the neural network starts with randomly initiated parameter values (or weights) and gradually learns the task, transfer learning starts with a pre-trained model and fine-tunes it using data specific to the desired task, in this case a precoder vector encoding task.

The further training with transfer learning differs from the MMSE approach in that it refines an already functional model rather than building a model from the ground up. While the MMSE approach focuses on minimizing the mean squared error between the output and the target coefficient vector, transfer learning involves using the pre-trained model's parameters as a starting point and adjusting them to minimize the MMSE in the context of the new, more specific task. This method can result in faster convergence and potentially better performance due to the model's initial training on a broader dataset.

At an operation 2208, the trained neural network may be deployed, for example to a UE serviced by the UE-vendor. The UE-vendor may use suitable signaling to deploy the trained neural network to a UE. For example, a control message (e.g., dedicated RRC message) may contain information about the trained neural network, such as an identifier of the encoder model implemented by the trained neural network, indication of the size of the encoder model (e.g., number of model parameters) and/or method by which the trained neural network will be deployed to the UE (e.g., via control messages, over a data channel, etc. ) . The parameters of the trained neural network may be deployed to the UE via control messages (e.g., using RRC signaling) where the neural network is relatively small (e.g., smaller number of parameters) ; a data channel (e.g., user plane data channel) may be suitable where the neural network is larger (e.g., larger number of parameters) . In some examples, parameters of the trained neural network may be sent in segments, with appropriate error detection and/or correction mechanisms to help ensure more reliable deployment. Additionally, synchronization signals (e.g., from the datacenter, from the UE-vendor or some other network entity) may be used to ensure that the encoder model deployed at the UE is aligned and up-to-date with a decoder model deployed at a network entity. This synchronization enables effective communication and data processing, and helps to ensure seamless integration and functionality within he wireless communication system.

At an operation 2210, the UE-vendor may update the representation subset. For example, the UE-vendor update the representation subset to include a new representative vector (e.g., by obtaining a new representative vector from the datacenter that corresponds to a new characteristic of interest, such as a new spatial characteristic of the UEs serviced by the UE-vendor) and/or to exclude an out-of-date representative vector (e.g., a representative vector that has not been used for at least a predetermined period of time) . The method 2200 may then return to the operation 2204. The UE-vendor may obtain new data corresponding to the updated representation subset (e.g., a new set of coefficient vectors representing linear combinations of the representative vectors in the updated representation subset) and the operation 2206 may be performed again to retrain the neural network to perform the precoder vector encoding task using the new set of coefficient vectors. The UE-vendor may then deploy the retrained neural network to the UE as described above.

FIG. 23 is a flowchart illustrating an example method 2300 for training a neural network to perform a precoder vector decoding task. The method 2300 may be implemented at a network entity that is a NW-vendor, for example, or other suitable network entity that may deploy the trained neural network at a network node (e.g., deploy at a TRP) . The method 2300 may be performed by a functional unit, such as a processor or any other suitable unit.

At an operation 2302, a set of one or more representative vectors defining a representation set is obtained by the NW-vendor. For example, the NW-vendor may obtain the full representation set that is maintained at the datacenter. The NW-vendor may send a request to the datacenter for the representation set and the datacenter may transmit the representation set to the NW-vendor, for example using control messages.

At an operation 2304, the NW-vendor obtains coefficient vectors. Each coefficient vector represents a linear combination of one or more selected representative vectors from the representation set, and each coefficient vector corresponds to a precoder vector. For example, the NW-vendor may send a request to the datacenter for coefficient vectors sampled from across the dataset maintained by the datacenter. The datacenter may sample coefficient vectors from the dataset and transmit the sampled coefficient vectors to the NW-vendor in a control message or using other signalling. The sampled coefficient vectors can each be converted to a respective precoder vector using representative vectors contained in the representation set. The sampled coefficient vectors may be obtained by the NW-vendor in the expanded vector form.

At an operation 2306, the NW-vendor uses the obtained coefficient vectors to train a neural network to perform a precoder vector decoding task. To train the neural network, a coefficient vector is converted to the corresponding precoder vector using the representation set (e.g., by multiplying the coefficient vector with the representation subset to obtain the corresponding precoder vector) . Then the precoder vector and corresponding coefficient vector is a data pair that is used to train the neural network. For example, the coefficient vector is provided as input to the neural network and an output generated by the neural network is obtained. The neural network may then be trained to minimize an error between the output generated by the neural network and the precoder vector corresponding to the coefficient vector, using suitable machine learning techniques (e.g., to minimize the MMSE) . In some examples, the neural network may be trained using a transfer learning approach, in which a pre-trained model is further trained using the coefficient vectors.

This transfer learning may be similar to the process previously described with respect to the operation 2206. For example, the obtained coefficient vectors may be used to fine-tune a pre-trained model that has been pre-trained on a related task such as image recognition, natural language processing, or a preliminary version of precoder vector decoding.

At an operation 2308, the trained neural network may be deployed, for example to a network entity (e.g., a TRP) serviced by the NW-vendor. The NW-vendor may use suitable signaling to deploy the trained neural network to a network entity. For example, a control message (e.g., dedicated RRC message) may contain information about the trained neural network, such as an identifier of the decoder model implemented by the trained neural network, indication of the size of the decoder model (e.g., number of model parameters) and/or method by which the trained neural network will be deployed to the network entity (e.g., via control messages, over a data channel, etc. ) . The parameters of the trained neural network may be deployed to the network entity via control messages (e.g., using RRC signaling) where the neural network is relatively small (e.g., smaller number of parameters) ; a data channel (e.g., user plane data channel) may be suitable where the neural network is larger (e.g., larger number of parameters) . In some examples, parameters of the trained neural network may be sent in segments, with appropriate error detection and/or correction mechanisms to help ensure more reliable deployment. Additionally, synchronization signals (e.g., from the datacenter, from the NW-vendor or some other network entity) can be used to ensure that the decoder model deployed at the network entity is aligned and up-to-date with the encoder model deployed at the UE. This alignment enables effective communication and data processing, helping to ensure seamless integration and functionality within the wireless communication system.

In some examples, the trained neural network may be trained at the same network entity that performs the training. In such cases, deployment of the trained neural network may be simply storing the trained parameters in memory. After deployment, a coefficient vector may be received (e.g., from a UE) and the coefficient vector may be decoded using the trained neural network to recover the precoder vector (e.g., using the method 2400 discussed further below) .

At an operation 2310, an updated representation set may be received. For example, after the representation set is updated at the datacenter (e.g., by addition of a new representative vector or removal of an out-of-date representative vector) , the datacenter may send an update message to the NW-vendor with information about the updated representation set (e.g., including an updated identifier, identification of a removed representative vector and/or an added representative vector) . The method 2300 may then return to the operation 2304 where new coefficient vectors corresponding to the updated representation set are obtained. Then the operation 2306 may be performed again to retrain the neural network to perform the precoder vector decoding task using the new set of coefficient vectors. The NW-vendor may then deploy the retrained neural network as described above.

FIG. 24 is a flowchart illustrating an example method 2400 for using a trained neural network to decode a precoder vector. The method 2400 may be implemented at a network entity, such as a TRP. The method 2400 may be performed by a functional unit, such as a processor or any other suitable unit. The trained neural network may have been trained using the method 2300 described above, for example, and deployed to the network entity.

At an operation 2402, a coefficient vector is received (e.g., from a UE) . The coefficient vector corresponds to a precoder vector and may be received as a form of PMI. As described above, the coefficient vector may be a sparse representation of the precoder vector, where the coefficient vector represents a linear combination of one or more selected representative vectors from a representation set defined by one or more representative vectors.

In some examples, the coefficient vector may be received in a compact vector form that corresponds to a representation subset that is smaller than the representation set. In order for the precoder vector to be decoded from the coefficient vector, the coefficient vector may need to be converted to an expanded vector form that corresponds to the representation set. At an operation 2404, the coefficient vector may be transformed from the compact vector form to the expanded vector form using a mapping. The mapping may be a transformation matrix, for example, that transforms between the compact vector form and the expanded vector form. The mapping may be obtained beforehand (e.g., obtained from the datacenter via control signaling) . An identifier associated with the mapping (or associated with the representation subset associated with the mapping) may be provided by or associated with the UE, to enable the network entity to apply the proper mapping to convert the coefficient vector to the expanded vector form. The decoding may then be performed on the coefficient vector in the expanded vector form.

At an operation 2406, the trained neural network is used to decode the precoder vector from the coefficient vector.

The decoded precoder vector may then be used for various operations. For example, at an operation 2408, the decoded precoder vector may be provided for use in optimizing network operations and resource allocation, among other possibilities.

In some examples, the decoded precoder vector may be stored with the corresponding original coefficient vector. This may form a data pair that may be used for further training and/or refinement of the neural network. For example, data pairs formed in this manner can be analyzed (e.g., using machine learning or statistic techniques) to identify patterns or discrepancies between the decoded and original data, which may provide insights that can guide adjustments to the neural network's architecture or parameters. Further training or refinement can be achieved by using these data pairs to fine-tune the neural network, helping to improve its accuracy and robustness. The stored data pairs may enable the neural network to be trained on real-world scenarios, which may help to enhance the model’s performance over time by updating it with new data. This iterative process may help to ensure that the neural network remains effective and reliable, adapting to changing conditions and improving its predictive capabilities.

In some examples, information regarding the accuracy of the decoded precoder vectors may be provided as feedback to further refine the training of the neural network. The feedback can include metrics such as the error rate, signal-to-noise ratio (SNR) , bit error rate (BER) , and/or any discrepancies between the expected and actual performance of the decoded precoder vectors. This information may be collected from the network entity and sent back to the datacenter or the entity responsible for training the neural network (e.g., the NW-vendor) .

The feedback can be used to identify specific areas where the neural network's performance can be improved. For example, if the feedback indicates a high error rate for certain types of channels or conditions, the training process can be adjusted to focus more on those scenarios (e.g., the NW-vendor may request more training data from the datacenter specific to those scenarios) . The neural network can be retrained using additional data that reflects these conditions or by modifying the training algorithm to better handle the identified issues.

Moreover, the feedback can be used to perform incremental updates to the neural network. Instead of retraining the entire model from scratch, the neural network can undergo fine-tuning sessions where only certain layers or parameters are adjusted based on the feedback (e.g., by fixing hidden layers and focusing the training on only the final output layer) . This process ensures that the model remains up-to-date with the latest operational data and can adapt to new patterns or changes in the network environment.

By repeatedly incorporating feedback from the network entity after the neural network has been deployed, the trained neural network can be fine-tuned to reflect up-to-date real-world conditions, which may help the neural network to achieve higher accuracy and reliability, ultimately enhancing the overall efficiency and performance of the wireless communication system.

FIG. 25 is a flowchart illustrating an example method 2500 for using a trained neural network to encode a precoder vector. The method 2500 may be implemented by a processor executing instructions at an ED, such as a UE. The trained neural network may have been trained using the method 2200 described above, for example, and deployed to the UE.

The trained neural network may have been deployed to the UE by the UE obtaining parameters of the trained neural network via a control signal and/or over a data channel, for example as discussed above.

At an operation 2502, a precoder vector is obtained by the UE. The UE may obtain a precoder vector based on one or more parameters indicated by a control message from a network entity. For example, a control message (e.g., RRC message) to the UE may specify the CSI-RS resources, data collection duration and reporting mode, among other parameters. The UE may then perform channel estimation on the specified CSI-RS resources and extract the precoder matrix (the vectors of which are the precoder vectors) .

At an operation 2504, a trained neural network is used to encode the precoder vector into a coefficient vector. The coefficient vector is a sparse representation of the precoder vector. The coefficient vector represents a linear combination of one or more selected representative vectors from a representation subset defined by one or more representative vectors.

At an operation 2506, the coefficient vector is transmitted. For example, the coefficient vector may be transmitted as a PMI. In some examples, ancillary information (e.g., related to data collection for the precoder vector, as discussed above) may be transmitted with the coefficient vector.

Examples of the present disclosure also provide a protocol and signaling design at the physical and MAC layers. For example, at the datacenter, information about the representation set, such as a unique identifier for the representation set and parameters of representative vectors (e.g., dimensionality, quantization methods, etc. ) may be broadcasted through system information messages. The datacenter may also provide a network entity with a requested representation set or representation subset (which may be identified using a unique identifier different from that of the representation set) , and/or with requested data sampled from the dataset of coefficient vectors. The datacenter may optionally provide a synchronization signal (e.g., via RRC or MAC control element) to help ensure that the same representation set version and encoder/decoder model versions are being used by UEs and network entities.

The UE may report its supported AI/ML functionalities, such as the representative vectors that the UE is able to work with (e.g., the representative vectors forming the representation subset that the encoder model, deployed at the UE, is trained on) . The UE may implement multiple trained encoder models. The UE may report the complexity and performance characteristics of its supported encoder models. Optionally, the UE may indicate the availability of locally collected data (e.g., locally collected precoder vectors) to be used for fine tuning a model.

A network entity, such as the datacenter or other scheduler, may trigger data collection at the UE via a dedicated RRC message or other control message. The control message may indicate the resources for performing channel estimation (e.g., the CSI-RS resources) , data collection duration and reporting mode, for example.

A network entity, such as a UE-vendor, may transfer to the UE the parameters of a trained neural network that implements the encoder model. The parameters may be transferred via control signaling (e.g., RRC signaling) or via a data channel, for example depending on the number of parameters to be transferred. Various error detection and correction mechanisms may be used to ensure reliable delivery of the model parameters. Optionally, a synchronization signal may be provided to ensure the UE is using an up-to-date model.

Various feedback mechanisms may be implemented to help ensure satisfactory performance of the trained models. For example, UEs and TRPs at which the trained encoder/decoder models are deployed may provide precoder vector and coefficient vector data pairs that may be used for further training of the models. Performance of the trained models may also be reported. Each UE and TRP may dynamically select an appropriate channel for data reporting, for example based on payload size and/or channel conditions. HARQ mechanisms may also be implemented for greater reliability.

Data reported from multiple UEs may be grouped together at a TRP, to more efficiently communicate the data to the datacenter and/or reduce overhead.

Based on feedback from the UEs and TRPs, resources for channel estimation and data collection (e.g., CSI-RS resources) may be dynamically configured, to help ensure sufficient up-to-date measurements.

A MAC layer scheduler may perform operations to prioritize data collection and model update transmissions, while balancing such communications with other traffic. For example, various adaptive scheduling algorithms may be used for dynamic resource allocation, based on real-time channel conditions and/or traffic demands.

Various security mechanisms may be implemented. For example, secure boot mechanisms and secure storage solutions may be implemented on UEs and TRPs (and other network entities) to help prevent unauthorized modifications to the trained encoder and decoder models. Authentication and/or authorization protocols may be used to ensure that only trusted entities can access and contribute data to the datacenter. Authentication and/or authorization protocols may be used to help ensure that models and model updates are communicated to only trusted entities. All communications between the UE, TRP, network entities and the datacenter may employ suitable encryption and integrity protection mechanisms.

The present disclosure encompasses various embodiments, including not only method embodiments, but also other embodiments such as apparatus embodiments and embodiments related to non-transitory computer readable storage media. Embodiments may incorporate, individually or in combinations, the features disclosed herein.

Although this disclosure refers to illustrative embodiments, this is not intended to be construed in a limiting sense. Various modifications and combinations of the illustrative embodiments, as well as other embodiments of the disclosure, will be apparent to persons skilled in the art upon reference to the description.

Features disclosed herein in the context of any particular embodiments may also or instead be implemented in other embodiments. Method embodiments, for example, may also or instead be implemented in apparatus, system, and/or computer program product embodiments. In addition, although embodiments are described primarily in the context of methods and apparatus, other implementations are also contemplated, as instructions stored on one or more non-transitory computer-readable media, for example. Such media could store programming or instructions to perform any of various methods consistent with the present disclosure.

The terms "system" and "network" may be used interchangeably in embodiments of this application. "At least one" means one or more, and "a plurality of" means two or more. The term "and/or" describes an association relationship of associated objects, and indicates that three relationships may exist. For example, A and/or B may indicate the following three cases: Only A exists, both A and B exist, and only B exists, where A and B may be singular or plural. The character "/" usually indicates an "or" relationship between associated objects. "At least one of the following items (pieces) " or a similar expression thereof indicates any combination of these items, including a single item (piece) or any combination of a plurality of items (pieces) . For example, "at least one of A, B, or C" includes A, B, C, A and B, A and C, B and C, or A, B, and C, and "at least one of A, B, and C" may also be understood as including A, B, C, A and B, A and C, B and C, or A, B, and C. In addition, unless otherwise specified, ordinal numbers such as "first" and "second" in embodiments of this application are used to distinguish between a plurality of objects, and are not used to limit a sequence, a time sequence, priorities, or importance of the plurality of objects.

It should be understood that examples of the present disclosure may be embodied as a method, an apparatus, a non-transitory computer readable medium, a processing module, a chipset, a system chip or a computer program, among others. An apparatus may include a transmitting module configured to carry out transmitting steps described above and a receiving module configured to carry out receiving steps described above. An apparatus may include a processing module, processor or processing unit configured to control or cause the apparatus to carry out examples disclosed herein.

Although the present disclosure describes methods and processes with steps in a certain order, one or more steps of the methods and processes may be omitted or altered as appropriate. One or more steps may take place in an order other than that in which they are described, as appropriate.

Although the present disclosure is described, at least in part, in terms of methods, a person of ordinary skill in the art will understand that the present disclosure is also directed to the various components for performing at least some of the aspects and features of the described methods, be it by way of hardware components, software or any combination of the two. Accordingly, the technical solution of the present disclosure may be embodied in the form of a software product. A suitable software product may be stored in a pre-recorded storage device or other similar non-volatile or non-transitory computer readable medium, including DVDs, CD-ROMs, USB flash disk, a removable hard disk, or other storage media, for example. The software product includes instructions tangibly stored thereon that enable a processing device (e.g., a personal computer, a server, or a network device) to execute examples of the methods disclosed herein. The machine-executable instructions may be in the form of code sequences, configuration information, or other data, which, when executed, cause a machine (e.g., a processor or other processing device) to perform steps in a method according to examples of the present disclosure.

The present disclosure may be embodied in other specific forms without departing from the subject matter of the claims. The described example embodiments are to be considered in all respects as being only illustrative and not restrictive. Selected features from one or more of the above-described embodiments may be combined to create alternative embodiments not explicitly described, features suitable for such combinations being understood within the scope of this disclosure.

All values and sub-ranges within disclosed ranges are also disclosed. Also, although the systems, devices and processes disclosed and shown herein may comprise a specific number of elements/components, the systems, devices and assemblies could be modified to include additional or fewer of such elements/components. For example, although any of the elements/components disclosed may be referenced as being singular, the embodiments disclosed herein could be modified to include a plurality of such elements/components. The subject matter described herein intends to cover and embrace all suitable changes in technology.

Claims

A method comprising:

receiving data including a precoder vector associated with a user equipment (UE) ;

converting the precoder vector to a coefficient vector representing a linear combination of one or more selected representative vectors from a representation set defined by one or more representative vectors, the coefficient vector being useable for training a neural network to perform a precoder vector encoding task or a precoder vector decoding task.
The method of claim 1, wherein converting the precoder vector to the coefficient vector comprises:

computing a candidate vector to represent the precoder vector as a linear combination of one or more candidate representative vectors from the representation set; and

determining that the candidate vector satisfies a sparsity threshold;

wherein, after the candidate vector is determined to satisfy the sparsity threshold, the candidate vector is the coefficient vector and the one or more candidate representative vectors is the one or more selected representative vectors.
The method of claim 1, wherein converting the precoder vector to the coefficient vector comprises:

determining that the representation set is non-representative of the precoder vector;

updating the representation set by including the precoder vector as a new representative vector in the representation set; and

converting the precoder vector to the coefficient vector after the updating.
The method of claim 3, wherein determining that the representation set is non-representative of the precoder vector comprises:

computing a candidate vector to represent the precoder vector as a linear combination of one or more candidate representative vectors from the representation set prior to the updating; and

determining that the candidate vector fails to satisfy a sparsity threshold.
The method of any one of claims 1 to 4, further comprising:

tracking usage and/or age of each representative vector in the representation set; and

updating the representation set by removing an out-of-date representative vector based on the tracked usage and/or age.
The method of any one of claims 3 to 5, further comprising:

transmitting a control message indicating an update to the representation set.
The method of any one of claims 1 to 6, further comprising:

transmitting a control message indicating one or more parameters to be used by the UE to generate the precoder vector;

wherein the precoder vector is received in response to the control message.
The method of any one of claims 1 to 7, wherein the received data further includes ancillary information related to data collection by the UE for generating the precoder vector, and wherein the coefficient vector corresponding to the precoder vector is associated with the ancillary information.
The method of claim 8, wherein the ancillary information includes one or more of:

information about a frequency sub-band used in the data collection;

information about a resource block used in the data collection;

information about an antenna configuration used in the data collection; and

information about a geographical location during the data collection.
The method of any one of claims 1 to 9, further comprising:

providing the coefficient vector to the neural network for training to perform the precoder vector encoding task or the precoder vector decoding task.
The method of any one of claims 1 to 10, further comprising:

generating a synthetic coefficient vector based on one or more coefficient vectors converted from precoder vectors;

wherein the synthetic coefficient vector is also useable for training the neural network to perform the precoder vector encoding task or the precoder vector decoding task.
The method of claim 11, further comprising:

providing the synthetic coefficient vector to the neural network for training to perform the precoder vector encoding task or the precoder vector decoding task.
The method of any one of claims 1 to 12, wherein:

the method is performed at a datacenter of a wireless communication system that includes the UE;

the set of one or more representative vectors is maintained in a memory of the datacenter; and

the coefficient vector is stored in a dataset, maintained in the memory of the datacenter, that is useable for training the neural network to perform the precoder vector encoding task or the precoder vector decoding task.
The method of claim 13, further comprising:

transmitting coefficient vectors in the dataset to train the neural network to perform the precoder vector decoding task, wherein the transmitted coefficient vectors are sampled from across the entire dataset.
The method of claim 13, further comprising:

transmitting coefficient vectors in the dataset to train the neural network to perform the precoder vector encoding task, wherein the transmitted coefficient vectors are sampled from a selected sub-dataset of the dataset, the selected sub-dataset containing coefficient vectors associated with a spatial characteristic of interest, and wherein the transmitted coefficient vectors correspond to a selected representation subset that omits at least one representative vector from the representation set and that is associated with the spatial characteristic of interest.
The method of claim 15, further comprising:

transmitting mapping information for mapping between the representation subset and the representation set.
The method of claim 15 or 16, wherein the transmitted coefficient vectors are transmitted in a compact vector form that is smaller than an expanded vector form used to store the coefficient vectors in the dataset, the compact vector form omitting entries corresponding to at least one representative vector omitted from the representation subset.
The method of claim 17 when dependent on claim 16, wherein the mapping information also maps between the compact vector form and the expended vector form.
A method comprising:

obtaining a selected subset of one or more representative vectors from a set of one or more representative vectors defining a representation set, the selected subset of one or more representative vectors defining a representation subset that is smaller than the representation set;

obtaining coefficient vectors, wherein each coefficient vector represents a linear combination of one or more selected representative vectors from the representation subset, and wherein the linear combination corresponds to a precoder vector; and

training a neural network to perform a precoder vector encoding task, wherein during training each coefficient vector is converted to the corresponding precoder vector and the corresponding precoder vector is used to train the neural network.
The method of claim 19, wherein training the neural network comprises:

providing the corresponding precoding vector as input to the neural network and obtaining an output; and

training the neural network to minimize an error between the output generated by the neural network and the coefficient vector corresponding to the precoding vector.
The method of claim 19 or claim 20, wherein the coefficient vectors are sampled from a selected sub-dataset containing coefficient vectors associated with a spatial characteristic of interest, and each coefficient vector in the sub-dataset being convertible to a respective precoder vector using the representation subset.
The method of claim 21, wherein each coefficient vector is obtained in a compact vector form corresponding to the representation subset, the compact vector form being smaller than an expanded vector form corresponding to the representation set.
The method of any one of claims 19 to 22, further comprising:

deploying the trained neural network to perform the precoder vector encoding task.
The method of claim 23, wherein the trained neural network is deployed at a user equipment (UE) .
The method of any one of claims 19 to 24, further comprising:

updating the representation subset to include a new representative vector from the representation set or to exclude an out-of-date representative vector from the representation subset;

obtaining another set of coefficient vectors that corresponds to the updated representation subset; and

retraining the neural network to perform the precoder vector encoding task using the another set of coefficient vectors.
A method comprising:

obtaining a set of one or more representative vectors defining a representation set;

obtaining coefficient vectors, wherein each coefficient vector represents a linear combination of one or more selected representative vectors from the representation set, and wherein the linear combination corresponds to a precoder vector; and

training a neural network to perform a precoder vector decoding task, wherein during training each coefficient vector is used to train the neural network.
The method of claim 26, wherein training the neural network comprises:

providing the coefficient vector as input to the neural network and obtaining an output; and

training the neural network to minimize an error between the output generated by the neural network and the precoder vector corresponding to the coefficient vector.
The method of any one of claims 26 to 27, further comprising:

receiving an update to the representation set, wherein the neural network is retrained responsive to the update.
The method of any one of claims 26 to 28, further comprising:

receiving feedback relevant to performance of the trained neural network; and

using the feedback to further refine training of the neural network.
The method of any one of claims 26 to 29, wherein:

the neural network is trained using transfer learning, wherein a pre-trained neural network is further trained using the coefficient vectors.
The method of any one of claims 26 to 30, further comprising:

deploying the trained neural network to perform the precoder vector decoding task.
The method of claim 31, further comprising:

receiving a coefficient vector; and

decoding a precoder vector from the coefficient vector using the trained neural network.
The method of claim 32, further comprising:

obtaining a mapping between a representation subset and the representation set, the representation subset being defined by a selected subset of one or more representative vectors from the representative set;

wherein:

the coefficient vector is received in a compact vector form that corresponds to the representation subset; and

the coefficient vector is transformed from the compact vector form to an expanded vector form corresponding to the representation set using the mapping, the decoding being performed on the coefficient vector in the expanded vector form.
A method comprising:

receiving a coefficient vector that represents a linear combination of one or more selected representative vectors from a representation set defined by one or more representative vectors, and wherein the linear combination corresponds to a precoder vector; and

decoding the precoder vector from the coefficient vector using a trained neural network, wherein the trained neural network has been trained to perform a precoder vector decoding task.
The method of claim 34, further comprising:

obtaining a mapping between a representation subset and the representation set, the representation subset being defined by a selected subset of one or more representative vectors from the representative set;

wherein:

the coefficient vector is received in a compact vector form that corresponds to the representation subset; and

the coefficient vector is transformed from the compact vector form to an expanded vector form corresponding to the representation set using the mapping, the decoding being performed on the coefficient vector in the expanded vector form.
The method of any one of claims 34 to 35, further comprising:

storing the decoded precoder vector and the corresponding original coefficient vectors as a data pair for further analysis and/or refinement of the trained neural network.
The method of any one of claims 34 to 36, further comprising:

transmitting the decoded precoder vector for use in optimizing network operations and/or resource allocation.
A method comprising:

obtaining a precoder vector;

using a trained neural network to encode the precoder vector into a coefficient vector represents a linear combination of one or more selected representative vectors from a representation subset defined by one or more representative vectors, wherein the linear combination corresponds to the precoder vector; and

transmitting the coefficient vector.
The method of claim 38, wherein the coefficient vector is transmitted as a precoding matrix indicator (PMI) .
The method of claim 38 or 39, further comprising:

obtaining parameters of the trained neural network from a control signal.
The method of any one of claims 38 to 40, further comprising:

updating the trained neural network based on feedback received from a network entity.
The method of any one of claims 38 to 41, wherein:

the precoder vector is obtained based on one or more parameters indicated by a control message from a network entity.
The method of any one of claims 38 to 42, further comprising:

transmitting ancillary information related to the precoder vector along with the coefficient vector.
A network entity comprising:

a memory; and

a processor configured to execute instructions stored in the memory to cause the network entity to carry out the method of any one of claims 1 to 37.
The network entity of claim 44, wherein the network entity is a datacenter.
The network entity of claim 44, wherein the network entity services one or more user equipment (UEs) .
The network entity of claim 44, wherein the network entity services one or more base stations (BSs) .
An apparatus comprising:

a memory; and

a processor configured to execute instructions stored in the memory to cause the apparatus to carry out the method of any one of claims 38 to 43.
A non-transitory computer readable medium having machine-executable instructions stored thereon, wherein the instructions, when executed by a network entity, cause the network entity to perform the method of any one of claims 1 to 37.
The non-transitory computer readable medium of claim 49, wherein the network entity is a datacenter.
The non-transitory computer readable medium of claim 49, wherein the network entity services one or more user equipment (UEs) .
The non-transitory computer readable medium of claim 49, wherein the network entity services one or more base stations (BSs) .
A non-transitory computer readable medium having machine-executable instructions stored thereon, wherein the instructions, when executed by an apparatus, cause the apparatus to perform the method of any one of claims 38 to 43.
A processing module configured to control a network entity to cause the network entity to carry out the method of any one of claims 1 to 37.
The processing module of claim 54, wherein the network entity is a datacenter.
The processing module of claim 54, wherein the network entity services one or more user equipment (UEs) .
The processing module of claim 54, wherein the network entity services one or more base stations (BSs) .
A processing module configured to control an apparatus to cause the apparatus to carry out the method of any one of claims 38-43.
A computer program characterized in that, when the computer program is run on a computer, the computer is caused to execute the method of any one of claims 1 to 43.