WO2025073504A1 - Method, device and system for adapting a data transmission following a change of attachment - Google Patents

Abstract

The invention relates to a method for adapting a data stream transmission in a session between a second access device (STB) and a terminal (Term), wherein the preceding data of the stream is transmitted between the terminal (Term) and a first access device (STA), wherein the method is implemented by the second access device (STB) and comprises obtaining, from the first access device (STA), at least one dynamic transmission characteristic (Caract) of the preceding data between the first access device (STA) and the terminal (Term), wherein the characteristic (Caract) is stored in a first file associated with a neural network (RN_STA_Em, RN_STA_Rec) specific to the session and to the first access device (STA), attaching the terminal (Term) to the second access device (STB) after obtaining the at least one characteristic (Caract), and adapting the transmission of the subsequent data of the stream between the terminal (Term) and the second access device (STB) to which the terminal is attached by modifying a second file associated with a neural network (RN_STB_Em, RN_STB_Rec) specific to the session and to the second access device (STB) on the basis of the at least one received characteristic.

Description

Method, device and system for adapting a data transmission following a change of attachment

The present disclosure belongs to the field of communication networks and more particularly to techniques for transferring a terminal from a first access device to a second access device of a communication network. It aims more particularly to optimize the transmission of data following the transfer of the terminal by taking advantage of the attachment of the terminal with the first access device.

Advances in Artificial Intelligence (AI) have opened many doors and enabled the application of these techniques in many fields. Among these, there is the field of digital signal processing for telecommunications.

A first approach to using artificial intelligence techniques in the telecommunications field consists of implementing physical layer functions with techniques derived from AI, such as neural networks. In this functional block approach (coding, communication channel estimation, decoding, etc.), we replace the classic algorithm used in current telecommunications networks with an element derived from AI. Thus, the same error correction and optimization algorithm can be implemented either by a standard algorithm or by means of a neural network. This allows us to benefit from the advantages of such an approach (performance, complexity, implementation flexibility) while remaining transparent from the point of view of the architecture of the signal processing chain, as well as the protocols used. In this approach, the usual functions of combating the deleterious effects of the propagation channel through equalization or error correction remain, only the way of implementing them changes, the new method being based on AI techniques.

A second, more disruptive approach abandons the objective of seamlessly integrating into a deterministic signal processing chain. It is no longer a matter of preserving the blocks as described in the first approach, but of considering a global communication channel and acting on communication without distinguishing between the different blocks. The degrees of freedom that AI then benefits from are greater, as well as, consequently, its optimization potential. In a bidirectional data exchange scheme between two learning systems, communicating through a given propagation channel, this learning characteristic allows them to gradually learn the most efficient method for communicating. By following common optimization metrics, the two systems learn to best exploit the shared propagation channel and thus define an optimized communication system. From this perspective, it is not just a matter of canceling the harmful effects of the channel to get as close as possible to a perfect channel (Gaussian channel), but rather of finding the optimal way to exploit a given channel. If we take this process to its conclusion, we arrive at what the state of the art describes as communication between two auto-coders who have developed their own "language".

This second approach, although promising, nevertheless faces two limitations:

It can only be implemented between two systems with similar learning capabilities, and benefiting from an initial “primary” two-way communication capability in order to enable the establishment of the feedback loop necessary for learning.

The principle of specific optimization for a given propagation channel has the direct consequence that this optimization is only relevant for this specific implementation. However, in certain contexts, such as mobile radio communications, the propagation channel is not stable, but evolves regularly, which requires a form of continuous learning.

In the context of mobile networks, these developments in the data propagation channel can take two forms:

• Soit les évolutions sont très rapides par rapport au temps nécessaire à l’apprentissage sur un canal statique. Dans ce cas, cette variabilité sera intégrée au sein d’un apprentissage, quitte à limiter les performances de la configuration qui en découlera, car mal adaptée au canal courant.
• Soit les évolutions sont lentes, c’est-à-dire que le phénomène concerné peut être considéré comme statique sur le temps d’apprentissage. La survenue de telles variation va donc nécessiter une mise à jour de l’apprentissage, pour adapter le modèle retenu à la nouvelle configuration du canal. Suivant l’ampleur de cette évolution, la remise en question de l’apprentissage existant sera plus ou moins profonde.

Continuous evolution: linked to natural phenomena of the channel, such as equipment movements, or changes in the propagation environment. These evolutions are progressive and take place, for example, between the terminal and a base station. They are detected and monitored, generally through the use of reference signals that allow channel knowledge to be updated. In the case of a communication system that has learned to communicate on the channel in question, two scenarios may arise:

• Either the changes are very rapid compared to the time required for learning on a static channel. In this case, this variability will be integrated into learning, even if it means limiting the performance of the resulting configuration, because it is poorly adapted to the current channel.
• Either the changes are slow, that is to say that the phenomenon concerned can be considered static over the learning time. The occurrence of such variations will therefore require an update of the learning, to adapt the chosen model to the new configuration of the channel. Depending on the extent of this change, the questioning of the existing learning will be more or less profound.

Discontinuous evolution: Here, the evolution of the radio channel is sudden and does not present continuity with the previous situation. It is no longer an evolution of the existing channel but rather a change to a new propagation channel. In the context of a telecommunications network, such evolutions are typically caused by a change of the attached access device during communication (also called handover in English).

This development results in the resetting of the learning protocol with a fallback to the primary communication mode during handover from one access device to another access device.

In the case of a change of the attached access device, restarting the learning process from zero is, according to known technology, inevitable since the terminal <-> access device pair has changed.

Returning to the primary communication mode will most often involve a significant degradation in the efficiency of the radio link and therefore in the quality of service provided by the network. In addition, the learning process requires time, not guaranteeing good transmission quality during this period. Finally, this learning process involves the consumption of radio resources and therefore a temporary loss of capacity.

We therefore see that discontinuous developments can be frequent and negatively impact network efficiency and the quality of the user experience, particularly when it is impossible to anticipate handover when the terminal disconnects from the original access device.

The present disclosure aims to remedy all or part of the limitations of the prior art solutions, in particular those set out above, by proposing a solution which makes it possible to optimize the transmission of data between a terminal and a new access device by using techniques relating to neural networks, by taking advantage of a transmission of data from the terminal with a previous access device, this transmission also using techniques relating to neural networks.

For this purpose, a method is proposed for adapting a transmission of a data stream of a session between a second access device and a terminal, the previous data of the stream being transmitted between said terminal and a first access device, said method being implemented by said second access device and comprising:
- obtaining from the first access device at least one dynamic characteristic of transmission of said previous data between the first access device and the terminal, said characteristic being represented in a first file associated with a neural network specific to the session and to the first access device,
- an attachment of said terminal to the second access device following the obtaining of at least one characteristic,
- an adaptation of the transmission of the following data of the flow between the terminal and the second access device to which the terminal is attached by modifying a second file associated with a neural network specific to the session and to the second access device according to the at least one characteristic received.

According to the prior art, when there is a discontinuity in a communication channel, for example during a change of attachment of a terminal, the terminal successively attaching to a first device then to a second access device, the terminal must reinitialize a learning process of its neural network to optimize the transmission of data of a session in progress between the terminal and the new access device to which it is now attached. The same applies to the neural networks of the access device hosting the terminal. More specifically, according to one example, the terminal must reinitialize a neural network for transmitting data to the second access device, and a neural network for receiving data from the second access device. Correspondingly, in this case, the second access device must initialize a transmitting neural network and a receiving neural network. In a handover context where the terminal does not benefit at any time from a simultaneous double attachment with the two access devices, the previous device and the new device, the adaptation method makes it possible to improve the learning process of the neural network of the second access device, or host device, and more precisely the files associated with the transmission and reception neural networks, these files comprising dynamic characteristics of transmission of the session data. Thus, the first access device to which the terminal was attached before the handover can advantageously transmit one or more characteristics of its own files (transmission and reception) to the second device to which the terminal attaches after the handover. Thus, the second device does not need to carry out a complete learning process of its neural network(s), and it is possible for it to reuse or exploit one or more of the transmitted characteristics, such as for example characteristics relating to the transmitted data, to the quality of service required for the session, to the propagation conditions probably equivalent for the two devices as well as parameters of the terminal. This method thus allows the second access device to include a neural network prepared for the continuity of the session during the handover and this makes it possible to accelerate the increase in quality of data transmission after the handover while saving communication network resources by making it possible to implement a faster learning process, including in a context where the terminal is never attached simultaneously to the two access devices.

According to one embodiment, the adaptation method is characterized in that the modification of the second file is initialized prior to the attachment of said terminal to the second device.

In a particular embodiment, the second access device can advantageously take into account the reception of a characteristic of a file associated with a neural network of the first device to adapt its own neural network, or its own neural networks, before the terminal attaches to the second access device. This makes it possible to further accelerate the learning process of the neural network of the second device for the current session(s) of the terminal. The first device can for example transmit the characteristic to the devices in its vicinity at the radio level, which are therefore likely to host the terminal.

According to a particular embodiment, the adaptation method is characterized in that the first file associated with a neural network specific to the session and to the first access device is specific to the reception of data from said terminal.

The second device can advantageously comprise two neural networks, a transmission network and a reception network and therefore two associated files, and the parameter received can then be specific to reception or transmission conditions. This makes it possible to adapt more quickly the routing conditions specific to the transmission or reception of session data.

According to one aspect of the invention, the adaptation method is characterized in that the at least one characteristic comprises a characteristic to be optimized as a function of a data transmission parameter between the second device and the terminal.

A characteristic of the data file may correspond to a transmission parameter, such as a transmission or reception power, a propagation speed or even a signal level. This characteristic may be reused or adapted by the second device knowing that the second device, probably close to the first access device, will probably have transmission parameters close to those of the first access device.

According to one aspect of the invention, the adaptation method is characterized in that the at least one characteristic comprises a characteristic to be transferred without modification from the first data file to the second file.

Knowing that certain characteristics of the file associated with the neural network of the first device are not linked to this first access device, but to the current session and to the terminal, or even to propagation conditions, these characteristics can be taken over without modification in the second file, associated with the second access device, and not give rise to a learning process, thus accelerating the optimization of the neural network(s) of the second access device.

According to one aspect of the invention, the adaptation method is characterized in that the at least one characteristic comprises information relating to a layer of the neural network or to a scalar parameter associated with a neuron of a layer.

Advantageously, a dynamic transmission characteristic comprises information, for example represented in the file, of a layer of the neural network specific to the first device and to the current session, this layer being able to be supplemented by information indicating whether it must be modified or kept as is for training the neural network of the second device leading to the updating of the second file. As a replacement or in addition to this information on the layer, the characteristic may comprise information relating to a scalar parameter associated with a neuron of a layer. This scalar parameter may correspond to a weight or a bias associated with the neuron and may also comprise information indicating whether this parameter can be kept unchanged or modified for training the neural network of the second device and subsequently, for modifying the second file.

The various aspects of the adaptation process just described can be implemented independently of each other or in combination with each other.

The invention also relates to a method for optimizing a transmission of a data stream of a session between a second access device and a terminal, the previous data of the stream being transmitted between said terminal and a first access device, said method being implemented by said first access device and comprising:
- detection of the second access device in the vicinity of the first access device,
- a transmission to the second access device of at least one dynamic characteristic of transmission of said previous data between the first access device and the terminal, said characteristic being included in a first file associated with a neural network specific to the session and to the first access device.

The optimization method aims to enable a first access device to enable the adaptation of a neural network of a second access device, to which the terminal is likely to attach. By transmitting at least one characteristic of a file associated with a neural network to one or more access devices that it has detected, therefore probably located in its vicinity, it enables optimization of the transmission of data from a session with a terminal even though the terminal changes its attachment during the session, this session not being interrupted. This optimization thus makes it possible to guarantee a better quality of service to the terminal, whereas in the absence of this method, the learning process of the neural network associated with the second device would inevitably have deteriorated the quality of service for routing the data of the session during the learning period.

According to one aspect of the optimization method, the optimization method is characterized in that the transmission is carried out via equipment connected to the first access device and to the second access device.

The transmission of the parameter can advantageously be carried out via one or more intermediate devices allowing a communications network architecture which would not be completely meshed between the access devices involved.

According to one aspect of the optimization method, the adaptation method is characterized in that the transmission to the first access device of at least one dynamic transmission characteristic is carried out following the detachment of the terminal from the first access device.

According to one embodiment, the transmission of the characteristic is carried out following the detachment of the terminal, thus allowing fair use of resources and sending of the characteristic only on the condition that the terminal actually detaches from the first access device.

The various aspects of the optimization process just described can be implemented independently of each other or in combination with each other.

The invention also relates to a device for adapting a transmission of a data stream of a session between a second access device and a terminal, the previous data of the stream being transmitted between said terminal and a first access device, said adaptation device comprising a processor and a memory coupled to the processor and further comprising:
- an obtaining module, configured to obtain from the first access device at least one dynamic characteristic of transmission of said previous data between the first access device and the terminal, said characteristic being represented in a first file associated with a neural network specific to the session and to the first access device,
- an attachment module, configured to attach said terminal to the second access device following obtaining of the at least one characteristic,
- an adaptation module, configured to adapt the transmission of the following data of the flow between the terminal and the second access device to which the terminal is attached by modifying a second file associated with a neural network specific to the session and to the second access device according to the at least one characteristic received.

This adaptation device is configured to implement in all its embodiments the adaptation method which has just been described.

The invention also relates to a device for optimizing a transmission of a data stream of a session between a second access device and a terminal, the previous data of the stream being transmitted between said terminal and a first access device, comprising a processor and a memory coupled to the processor and further comprising:
- a detection module, configured to detect the second access device in the vicinity of the first access device,
- a transmitter, configured to transmit to the second access device at least one dynamic characteristic of transmission of said previous data between the first access device and the terminal, said characteristic being included in a first file associated with a neural network specific to the session and to the first access device.

This optimization device is configured to implement in all its embodiments the optimization method which has just been described.

The invention also relates to a terminal having established a session with a remote device via a first access device, the terminal being adapted to successively attach to a first access device and then to a second access device, said terminal comprising a processor and a memory coupled to the processor and further comprising:
- at least one attachment module, configured so that the terminal detaches from the first access device and then successively attaches to the two access devices,
- a modification module, configured to modify a characteristic, relating to the change of attachment of the terminal, of a data file associated with a neural network specific to the session and to the terminal, said file being managed by the terminal and associated with the transmission of a data flow of a session between the terminal and the second access device, prior to the attachment of the terminal to said second device,
- a module for transmitting data from the following session data to the second access device in accordance with the characteristic modified following attachment to the second access device.

The invention also relates to a system for adapting a transmission of a data stream of a session between a second access device and a terminal, the previous data of the stream being transmitted between said terminal and a first access device, said system comprising:
- an adaptation device as described above,
- an optimization device as described above,
- a terminal as described above.

The invention also relates to computer program products comprising a set of program code instructions which, when executed by at least one processor, configure said at least one processor to implement the respective adaptation and optimization methods according to any of the embodiments of the present disclosure.

The invention also relates to a computer-readable recording medium having recorded thereon a set of program code instructions which, when are executed by at least one processor, configure said at least one processor to implement an adaptation method or an optimization method according to any of the implementation modes of the present disclosure.

•  : une représentation schématique d’un réseau de communication comprenant deux dispositifs d’accès dans lequel est mis en œuvre un procédé d’adaptation d’une transmission d’un flux de données d’une session entre un deuxième dispositif d’accès et un terminal, selon un mode de réalisation,
• : une représentation d’un réseau de neurones selon un mode de réalisation,
•  : un diagramme illustrant les principales étapes d’un procédé d’adaptation d’une transmission d’un flux de données d’une session entre un deuxième dispositif d’accès et un terminal, selon un mode de réalisation,
•  : une représentation d’un dispositif d’adaptation d’une transmission d’un flux de données d’une session entre un deuxième dispositif d’accès et un terminal selon un exemple,
•  : une représentation d’un dispositif d’optimisation d’une transmission d’un flux de données d’une session entre un deuxième dispositif d’accès et un terminal selon un exemple,
•  : une représentation d’un terminal adapté pour s’attacher successivement à un premier dispositif d’accès puis à un deuxième dispositif d’accès selon un exemple.

The invention will be better understood by reading the following description, given as a non-limiting example, and made with reference to the figures which represent:

• : a schematic representation of a communication network comprising two access devices in which a method is implemented for adapting a transmission of a data flow of a session between a second access device and a terminal, according to one embodiment,
• : a representation of a neural network according to one embodiment,
• : a diagram illustrating the main steps of a method for adapting a transmission of a data stream of a session between a second access device and a terminal, according to one embodiment,
• : a representation of a device for adapting a transmission of a data stream of a session between a second access device and a terminal according to an example,
• : a representation of a device for optimizing a transmission of a data flow of a session between a second access device and a terminal according to an example,
• : a representation of a terminal adapted to successively attach to a first access device then to a second access device according to an example.

In these figures, like references from one figure to another designate identical or similar elements. For reasons of clarity, the elements shown are not to scale, unless otherwise stated.

More generally, it should be noted that the methods of implementation and production considered above have been described as non-limiting examples, and that other variants are consequently conceivable.

In the remainder of the description, a communication network is presented comprising access devices (cellular network access stations, satellite type access stations, Wi-Fi access stations for example of the Box type or any other access station providing access to the communication network, for example of the Internet type). The stations and the network equipment can be instantiated in the form of specific equipment or in the form of virtualized functions.

We first refer to the which provides a schematic representation of a communication network comprising two access devices in which a method of adapting a transmission of a data flow of a session between a second access device and a terminal is implemented, according to one embodiment. This presents a communications network comprising two access devices, STA and STB, providing access to a cellular type communications network Res, independently of the network technology (3G, 4G, 5G, xG) to a terminal Term. The terminal Term (mobile terminal, laptop, local network access equipment for example of the box type, etc.) has, according to this embodiment, a transmission neural network RN_Term_Em and a reception neural network RN_Term_Rec making it possible to optimize communication, respectively in transmission and reception of data from a session with a remote device (terminal, server, machine, etc.) not shown in the .

These neural networks are thus made up of a succession of layers, as presented in the , representing in a more or less abstract way signal processing functions linked to the session data propagation channel, coding and decoding functions, hardware imperfection compensation functions, etc. Each layer is made up of a set of neurons to which learnable scalar parameters (weights and biases) are associated. The learning process makes it possible to modify these weights and biases to optimize the overall function performed by the neural network.

The Term terminal stores two files containing data relating to the neural networks RN_Term_Em and RN_Term_Rec respectively. These files describe the architecture and parameters (learnable weights and biases) of the neural networks, so that the contents of a file can be used to reproduce the neural network in question identically. This data file can be referred to as a standardized file format or a standard for representing a neural network.

Similarly, the STA and STB access stations have respectively a transmit neural network (RN_STA_Em and RN_STB_Em) and a receive neural network (RN_STA_Rec and RN_STB_Rec) whose role is to optimize the data session by the STA and STB stations. These neural networks are also transcribed into one or more data files whose data, such as the dynamic transmission characteristics, vary as the propagation channel evolves and the session is increasingly optimized, thus improving the transmission of the session data, in transmission and reception.

A data session is established between the terminal Term and a remote device (terminal, server, machine, etc.) not represented via the STA device of the communication network Res, the session data being routed via the communication Comm_A between the terminal Term and the access device STA. This can be an initialization of a session or a continuation of a session initiated via another access device. The respective neural networks of the terminal Term, RN_Term_Em and RN_Term_Rec, and of the STA device, RN_STA_Em and RN_STA_Rec, carry out a joint learning process with respect to this session allowing a good transmission and reception of data by the Terminal Term and the device STA. In the upstream channel, the transmitter neural network RN_Term_Em is trained jointly with the receiver neural network RN_STA_Rec. Similarly, in the downstream channel, the transmitter neural network RN_STA_Em is trained jointly with the receiver neural network RN_Term_Rec. Thus, the neurons representing a dynamic transmission characteristic specific to the session, the propagation channel or even a hardware interface of a device are assigned initial weights (and biases) which will be modified as the learning progresses until an optimal value is reached allowing routing of session data with a better quality of service, or even the best possible quality of service. The neural networks RN_Term_Em, RN_Term_Rec, RN_STA_Em and RN_STA_Rec are thus possibly different and are specific to the Term terminal and the STA device, to the transmission and reception of data, and are also specific to the established session. In fact, they are not directly transposable or adapted to a communication established for example between the Term terminal and the STB access device, for the same session or a separate session.

When a transfer, also called handover, of the current HDVR session occurs from the STA device to the STB device, a Comm_B communication is established between the Term terminal and the STB device for routing the session data via the STB device. This handover is implemented without the Term terminal being at any given time simultaneously attached to both the STA and STB devices, not allowing the handover to be prepared, according to known techniques (in English "Make before Break"). In this case, it is rather a Handover mechanism (in English "Break before Make") where the Comma_A communication between the Term terminal and the STA device is broken before the Comm_B communication is established.

The HDVR handover can occur for different reasons such as a loss of connectivity with the STA device, a detection of better connectivity of the terminal with the STB device, a movement of the Term terminal bringing it closer to the STB device, a configuration, for example dynamic, of the Term terminal preferring a connection with the STB device. Knowing that the neural networks of the Term terminal are optimized for the Comm_A communication and that the neural networks of the STB device RN_STB_Em and RN_STB_Rec are not optimized for the Comm_B communication established for the routing of the session data, a learning process of these neural networks must be initialized, in the absence of any adaptation process.

Indeed, in the absence of any method for adapting a transmission of a data stream of a session between the access STB device and the Term terminal, when the Term terminal attaches or connects to the STB device, the terminal and the STB device must carry out a joint learning process of their neural networks RN_Term_Em, RN_Term_Rec, RN_STB_Em and RN_STB_Rec so that the transmission of the session data between the Term terminal and the STB device via the Comm_B communication is optimized. The purpose of the learning is always here to allow the transmission of the session data while guaranteeing a good quality of service with respect to the propagation channel established between the Term terminal and the STB device.

The method for adapting a transmission of a data stream of a session between the access STB device and the terminal thus allows the STA device to transmit to the STB device, for example prior to the attachment of the Term terminal to the STB device, one or more dynamic transmission characteristics, represented in the Fich file associated with the neural network RN_STA_Em and/or RN_STA_Rec, to the STB station. The STA device can thus transmit to the devices which are neighboring it, and therefore to the STB device, one or more dynamic transmission characteristics. These characteristics can be transmitted in the form of a file or a stream or in a protocol for example. According to one example, the STA device can also transmit the entire file or files associated with its neural network(s). Thus, the STB device will have the transmitted characteristics, or even the data file, prior to the attachment of the Term terminal and will be able to avoid complete training of its neural networks for the Comm_B communication to be established.

Indeed, certain dynamic characteristics of file transmission, relating to the session, to the STA device, to routing conditions, can be reused within the RN_STB_Em and RN_STB_Rec neural networks. The optimization time will be reduced and the transmission characteristics, specific to the session via the STB station, will be adapted more quickly to this Comm_B communication than if a complete learning process of these characteristics had had to be carried out. Thus, for example, the modifications relating to the characteristics of the propagation channel between the terminal and the STB device and the characteristics relating to the hardware of the STB device will involve learning of the neural networks, but this will be less significant than if a complete learning of these neural networks had had to be carried out because all the neurons or characteristics do not have to repeat a complete learning process. The STA device can decide to transmit the entire Fich file representing the neural network RN_STA_Em and/or RN_STA_Rec or to transmit only part of the file, and to transmit for example only certain dynamic transmission characteristics which will have little or no need to be modified for optimization, this sending being able to be carried out via a file exchange from the file comprising the transmission characteristics associated with a neural network, or by any other means (flow, protocol).

Thus, the STA device can transmit one or more Caract characteristics to the STB device, these characteristics being able to be transmitted in a file, such as that associated with the neural networks of the STA device or in two files associated with each neural network (specific to transmission and reception) of the STA device, or even in a file extracted from the file associated with the neural network, or even in a stream or a protocol or any other means of transmitting the characteristic. The STA device can thus transmit a characteristic to the STB device or a set of dynamic transmission characteristics represented, that is to say included or translated into information usable by the STB device, in a file or a stream. According to one example, learnable parameters also called scalar parameters, represented for example by weights associated with neurons, can correspond to certain dynamic transmission characteristics transmitted, and these learnable parameters can be kept as is or modified during the adaptation or learning process implemented by the STB device. In the case where the access devices are satellite access devices, the characteristics, in particular transmission characteristics, may include satellite navigation data (speed vector, position, etc.) which data may be used to determine radio parameters (Dopler, elevation angle, MIMO configuration, etc.). These satellite characteristics and radio parameters may be used to update the second file associated with the neural network of the terminal for communication with the second STB access device.

Once it has received dynamic transmission characteristics Caract, for example in a file from the STA device, the STB device creates or updates its own file representing a neural network, and in particular modifies the neurons and the weights of the neurons of its neural network to adapt it to the communication Comm_B. The second file, associated with its neural network, may be identical to the first file, associated with the neural network of the first STA device, or may include one or more characteristics transmitted by the STA device and may be specifically associated with a neural network for transmitting and receiving data (RN_STB_Em or RN_STB_Rec) or the file may be common to both transmitting and receiving neural networks (RN_STB_Em and RN_STB_Rec).

We then refer to the which describes a representation of a neural network according to one embodiment.

The neural network represented in the is for example the neural network (RN_STA_Em) of emission of the STA access device presented in the . Binary elements (eb) 15 emitted by an output interface of the STA device to the Term terminal are converted into a signal (sig) adapted to the propagation conditions of a channel established between the STA device and the Term terminal. This signal is optimized by progressively modifying dynamic transmission characteristics associated with the neural network RN_STA_Em and represented in a data file associated with this neural network RN_STA_Em. The neural network RN_STA_Em is thus composed of several layers C1, C2....,,C7, each of the layers intervening for the transmission of data via the optimized sig signal, that is to say adapted to the propagation channel between the STA device and the terminal so that the data routed between the STA device and the Term terminal guarantee good quality of service for the session and subsequently for the service associated with the data. Each layer is composed of neurons as represented for layer C1 composed of neurons n1, n2, n3. At the output of a neuron, the data generated can be represented by the value y which is here equal to a weighted sum of the neuron's inputs to which we add a bias b_i. The weighting is done by a set of weights w_i, each input being associated with a weight. Each neuron n1, n2, n3 is therefore assigned different sets of weights w_i, which will evolve during the learning process just like the biases b_i, this information w_i and b_i can be called scalar parameters. For example, the data y is calculated according to the following formula y = phi (sum w_i x_i + b_i) where x_i are input values of neuron n1. In the following layers, C2 to C7, these inputs are the outputs of the neurons of the previous layers, such as layer C1 for layer C2. Other equations are possible to represent the calculation carried out by a neuron. All layers of the RN_STA_Em neural network are thus optimized, via neurons, weights and biases, for the transmission of data to the terminal, guaranteeing good quality of service for this transmission depending on the propagation channel between the STA device and the Term terminal.

Furthermore, with regard to a transmission of data from the STA device to a terminal, according to one example, the neural network RN_STA_Em is trained jointly with the reception neural network RN_Term_Rec of the terminal so that the transmission quality actually benefits from a good quality of service. The terminal can for its part accelerate its learning process by taking advantage, for example, of a reception neural network used for receiving data from another device distinct from the STA device. According to one embodiment, the different layers C1, C2,....., C7 and/or the scalar parameters independently represent dynamic data processing characteristics and these characteristics are themselves represented in a file associated with the neural network. According to another embodiment, the correspondence between the dynamic processing characteristics and the layers C1,...,C7 is not as direct and several layers of neurons or even several neurons intervene for a single characteristic. The adaptation method as described allows weights and/or biases or even one or more layers of a neural network, these characteristics representing particular embodiments, to be transferred. Dynamic transmission characteristics associated with a neural network of a device, for example STA, are thus transmitted and exploited by another device, such as the STB device, thus making it possible to reduce the learning process of a certain number of weights and biases or even layers of neurons in the neural network of this other STB device for a current data session.

We then refer to the which presents a diagram illustrating the main steps of a method for adapting a transmission of a data flow of a session between a second access device and a terminal, according to one embodiment.

In a step 200, an optimization process Optim of a neural network is implemented for a communication between a terminal and a first access device. This process aims to weight the n dynamic transmission characteristics in such a way that the optimized neural network comprises layers of neurons as well as scalar parameters such as weights and/or biases allowing good transmission of the data of the communication routed. The process aims to ensure that each neural network of the terminal, for the transmission and reception of the session data, and of the first access device are optimized, each file associated with a neural network possibly comprising neurons and different scalar parameter values.

In a step 201, a second access device obtains, during an Obt procedure from the first access device, at least one dynamic characteristic of transmission of the previous data of the session between the first access device and the terminal, said characteristic being, according to this embodiment, included in a first file associated with a neural network specific to the session and to the first access device. The second device is close to the first device in a communication network and may possibly accommodate the terminal during a future handover. The first device thus transmits to the second device one or more dynamic characteristics, i.e. neurons or layers of neurons or even scalar parameters or even calculation formulas associated with the neural networks, corresponding to characteristics in the neural networks used for carrying out the various functions of the signal processing chain in the uplink (transmission of data by the terminal) and downlink (reception of data from the terminal). According to one example, to accelerate learning by the second device, the first device may indicate all the layers of the neural network or more generally the weights associated with the neurons, which may be generalized as dynamic transmission characteristics, which must require optimization, and those which may be reused without modification by the second device. The transmission of the file by the first device may be carried out directly, using a direct interface between the two devices or via separate equipment attached to the first and second devices. It should be noted that according to one embodiment, this transfer is carried out following detection of the detachment of the terminal from the first device.

According to one example, during a step 202, the second device initiates an Init optimization of the uplink voice reception even before the terminal attaches to the second device. The second device therefore optimizes its neural network associated with reception with parameters, which it can recover by listening to data transmissions from the terminal to the first device or from the first device to the terminal. This procedure, implemented according to one embodiment, makes it possible to accelerate the learning process of the second device, at least as regards the reception neural network. The fact that the terminal only uses a transmission neural network, independently of the device to which it attaches, also allows this early initialization of the learning of the reception neural network of the second device, since, in this case, the characteristics linked to the terminal and to the associated neural network do not differ during a handover.

In a step 203, the terminal attaches itself, during an Attach procedure, to the second device, after having detached itself from the first device. This attachment procedure does not differ from an attachment according to the techniques of the prior art. Thus, the terminal authenticates itself with an entity of a communication network and obtains configuration parameters to be able to connect to the communication network.

In a step 204, the second device adapts, during an Adapt procedure, the transmission of the following data of the session flow between the terminal and the second access device by modifying a second file associated with a neural network specific to the session and to the second access device as a function of dynamic transmission characteristics that it received in step 201, for example transmitted in a file. In the case where the second device has already initiated the optimization process in step 202, this adaptation step comprises the continuation of the optimization of the reception neural network and the initialization of the optimization procedure of the transmission neural network and the updating or creation of the file(s) associated with the corresponding neural networks. In the case where the dynamic transmission characteristic received does not have to be modified, it is simply copied from the first file (relating to the first device) to the second file (relating to the second device). This may occur if the dynamic transmission characteristic concerns the current session (identifiers, parameters relating to the terminal or one of its interfaces, context of the handover in a rural or urban area). Certain characteristics relating, for example, to the second device or to the propagation channel and disturbances will not simply be copied but updated depending on the attachment of the terminal to the second device. This update includes the creation of new neurons or new layers of neurons, for example specific to the second device, and the modification of one or more weights of the neuron in the neural network and therefore its importance in the optimization process of the neural network, specific to the current session and to the second device ensuring the routing of the session data in place of the first device. In the case where a file associated with the neural network of the first device has been transferred, this modification also includes the deletion of neurons that would no longer be useful in the second file, such as, for example, the characteristics intrinsically linked to the first device. Thus the weights of neurons representing a signal quality, a distance from the terminal, a number of lost packets, latency, routing, representing certain dynamic transmission characteristics, can be modified depending on the attachment to the second device.

We then refer to the which provides a representation of a device for adapting a transmission of a data stream of a session between a second access device and a terminal according to an example, the previous data of the stream being transmitted between said terminal and a first access device.

For example, the adaptation device 100 comprises a processing unit 130, equipped for example with a microprocessor μP, and controlled by a computer program 110, stored in a memory 120 and implementing the adaptation method according to the invention. At initialization, the code instructions of the computer program 110 are for example loaded into a RAM memory, before being executed by the processor of the processing unit 130. The processor of the processing unit 130 controls the operations of the device 100. Such an adaptation device 100, which can be instantiated in an access device such as a Wi-Fi access station (hot Spot, box, etc.) or cellular (eNodeB, 5G gateway, etc.) or even satellite, comprises

- an obtaining module 101, configured to obtain from the first access device at least one dynamic characteristic of transmission of said previous data between the first access device and the terminal, said characteristic being represented in a first file associated with a neural network specific to the session and to the first access device,
- an attachment module 102, configured to attach said terminal to the second access device following the obtaining of the at least one characteristic,
- an adaptation module 103, configured to adapt the transmission of the following data of the flow between the terminal and the second access device to which the terminal is attached by modifying a second file associated with a neural network specific to the session and to the second access device as a function of the at least one characteristic received.

We then refer to the which provides a representation of a device for optimizing a transmission of a data flow of a session between a second access device and a terminal, the previous data of the flow being transmitted between said terminal and a first access device.

For example, the optimization device 400 comprises a processing unit 430, equipped for example with a microprocessor μP, and controlled by a computer program 410, stored in a memory 420 and implementing the optimization method according to the invention. At initialization, the code instructions of the computer program 410 are for example loaded into a RAM memory, before being executed by the processor of the processing unit 430. The processor of the processing unit 430 controls the operations of the device 400. Such an optimization device 400 can be instantiated in an access device such as a Wi-Fi access station (hot Spot, box, etc.) or cellular (eNodeB, 5G gateway, etc.) or even satellite, comprises

- a detection module 401, configured to detect the second access device in the vicinity of the first access device,
- a transmitter 402, configured to transmit to the second access device at least one dynamic characteristic of transmission of said previous data between the first access device and the terminal, said characteristic being included in a first file associated with a neural network specific to the session and to the first access device.

We then refer to the which provides a representation of a terminal adapted to attach successively to a first access device then to a second access device.

For example, the terminal 300 comprises a processing unit 330, equipped for example with a microprocessor μP, and controlled by a computer program 310, stored in a memory 320. At initialization, the code instructions of the computer program 310 are for example loaded into a RAM memory, before being executed by the processor of the processing unit 330. The processor of the processing unit 330 controls the operations of the terminal 300. Such a terminal 300 may be for example a smartphone, a PC or access equipment of a local network such as a home gateway, adapted to attach successively to a first access device then to a second access device, the terminal having established a session with a remote device via a first access device, and comprises:
- at least one attachment module 301, configured so that the terminal detaches from the first access device and then successively attaches to the two access devices,
- a modification module 302, configured to modify a characteristic, relating to the change of attachment of the terminal, of a data file associated with a neural network specific to the session and to the terminal, said file being managed by the terminal and associated with the transmission of a data flow of a session between the terminal and the second access device, prior to the attachment of the terminal to said second device,
- a module (303) for transmitting the following data of the session to the second access device in accordance with the characteristic modified following attachment to the second access device.

The entities described and included in the devices described in relation to figures 4, 5 and 6 may be hardware or software. Figures 4, 5 and 6 illustrate only one particular way, among several possible ones, of carrying out the method detailed above, in relation to the preceding figures. Indeed, the technique of the invention is carried out indifferently on a reprogrammable computing machine (a PC computer, a DSP processor or a microcontroller) executing a program comprising a sequence of instructions, or on a dedicated computing machine (for example a set of logic gates such as an FPGA or an ASIC, or any other hardware module).

In the case where the invention is implemented on a reprogrammable computing machine, the corresponding program (i.e. the sequence of instructions) may be stored in a removable storage medium (such as for example a USB key, a floppy disk, a CD-ROM or a DVD-ROM) or not, this storage medium being partially or totally readable by a computer or a processor.

Claims (15)

Method for adapting a transmission of a data stream of a session between a second access device (STB) and a terminal (Term), the previous data of the stream being transmitted between said terminal (Term) and a first access device (STA), said method being implemented by said second access device (STB) and comprising:
- obtaining (Obt) from the first access device (STA) at least one dynamic transmission characteristic (Caract) of said previous data between the first access device (STA) and the terminal (Term), said characteristic (Caract) being represented in a first file associated with a neural network (RN_STA_Em, RN_STA_Rec) specific to the session and to the first access device (STA),
- an attachment (Attach) of said terminal (Term) to the second access device (STB) following the obtaining of at least one characteristic (Caract),
- an adaptation (Adapt) of the transmission of the following data of the flow between the terminal (Term) and the second access device (STB) to which the terminal is attached by modifying a second file associated with a neural network (RN_STB_Em, RN_STB_Rec) specific to the session and to the second access device (STB) according to the at least one characteristic received. Adaptation method, according to claim 1, in which the modification of the second file is initialized prior to the attachment of said terminal to the second device.
Adaptation method, according to one of claims 1 or 2, in which the first file associated with a neural network (RN_STA_Rec) specific to the session and to the first access device is specific to the reception of data from said terminal.
Adaptation method according to one of claims 1 to 3, in which the at least one characteristic comprises a characteristic to be optimized as a function of a data transmission parameter between the second device and the terminal.
Adaptation method according to one of claims 1 to 4, in which the at least one characteristic comprises a characteristic to be transferred without modification from the first data file to the second file.
Adaptation method according to one of claims 1 to 5, in which the at least one characteristic comprises information relating to a layer of the neural network or to a scalar parameter associated with a neuron of a layer.
Method for optimizing a transmission of a data stream of a session between a second access device (STB) and a terminal (Term), the previous data of the stream being transmitted between said terminal (Term) and a first access device (STA), said method being implemented by said first access device (STA) and comprising:
- detection of the second access device (STB) in the vicinity of the first access device (STA),
- a transmission to the second access device (STB) of at least one dynamic transmission characteristic (Caract) of said previous data between the first access device (STA) and the terminal (Term), said characteristic (Caract) being included in a first file associated with a neural network (RN_STA_Em, RN_STA_Rec) specific to the session and to the first access device.
Optimization method, according to claim 7, wherein the transmission is carried out via equipment connected to the first access device and to the second access device.
Optimization method, according to claim 7 or claim 8, in which the transmission to the first access device of at least one dynamic transmission characteristic is carried out following the detachment of the terminal from the first access device.
Device (100) for adapting a transmission of a data stream of a session between a second access device and a terminal, the previous data of the stream being transmitted between said terminal and a first access device, said adaptation device comprising a processor (130) and a memory (110) coupled to the processor (130) and further comprising:
- an obtaining module (101), configured to obtain from the first access device at least one dynamic transmission characteristic (Caract) of said previous data between the first access device and the terminal, said characteristic being represented in a first file associated with a neural network specific to the session and to the first access device,
- an attachment module (102), configured to attach said terminal to the second access device following the obtaining of the at least one characteristic,
- an adaptation module (103), configured to adapt the transmission of the following data of the flow between the terminal and the second access device to which the terminal is attached by modifying a second file associated with a neural network specific to the session and to the second access device as a function of the at least one characteristic received.
Device (400) for optimizing a transmission of a data stream of a session between a second access device and a terminal, the previous data of the stream being transmitted between said terminal and a first access device, said optimization device comprising a processor (430) and a memory (410) coupled to the processor (430) and further comprising:
- a detection module (401), configured to detect the second access device in the vicinity of the first access device,
- a transmitter (402), configured to transmit to the second access device at least one dynamic transmission characteristic (Caract) of said previous data between the first access device and the terminal, said characteristic being included in a first file associated with a neural network specific to the session and to the first access device.
Terminal (Term) adapted to attach successively to a first access device and then to a second access device, the terminal having established a session with a remote device via a first access device, said terminal comprising a processor (330) and a memory (310) coupled to the processor (330) and further comprising:
- at least one attachment module (301), configured so that the terminal detaches from the first access device (STA) then successively attaches to the two access devices (STB),
- a modification module (302), configured to modify a characteristic, relating to the change of attachment of the terminal, of a data file associated with a neural network specific to the session and to the terminal, said file being managed by the terminal and associated with the transmission of a data flow of a session between the terminal and the second access device, prior to the attachment of the terminal to said second device,
- a module (303) for transmitting the following data of the session to the second access device in accordance with the characteristic modified following attachment to the second access device. System for adapting a transmission of a data stream of a session between a second access device and a terminal, the previous data of the stream being transmitted between said terminal and a first access device, said system comprising:
- an adaptation device according to claim 9,
- an optimization device according to claim 10,
- a terminal according to claim 12.
Program comprising instructions which, when the program is implemented by a processor, lead to implementing the adaptation method according to one of claims 1 to 6. Program comprising instructions which, when the program is implemented by a processor, lead to implementing the optimization method according to one of claims 7 to 9.