CN111638803B - EEG data processing method, device, equipment and medium based on federated migration - Google Patents

Abstract

The invention discloses a method, device, device and medium for processing EEG data based on federation migration. The method includes: acquiring multiple sets of EEG data, and transmitting the multiple sets of EEG data to each of the local terminals respectively. The feature mapping model performs mapping processing to generate standard EEG data; the standard EEG data is transmitted to the brain activity recognition model in the local terminal for processing, and the generated model parameters are transmitted to the preset public server for aggregation; the updated parameters returned by the aggregation are received , update the brain activity recognition model, and process the standard EEG data obtained by remapping each feature mapping model based on the updated brain activity recognition model, until the total loss value generated by the brain activity recognition model is less than the preset value, and complete the process based on Federated transfer processing of multiple sets of EEG data. The present invention realizes the processing of EEG data on the premise of fully protecting the user's privacy, avoids obtaining EEG data in various ways, and improves the processing efficiency.

Description

EEG data processing method, device, equipment and medium based on federated migration

technical field

The present invention relates to the technical field of financial technology (Fintech), in particular to a federated migration-based EEG data processing method, device, equipment and medium.

Background technique

With the continuous development of financial technology (Fintech), especially Internet technology finance, more and more technologies (such as big data, cloud storage, machine learning, etc.) Higher requirements, such as requiring machine learning to process EEG data to reflect the user's psychological state, mental state, etc.

Currently, EEG (Electroencephalogram, electroencephalogram) is considered to be an ideal way to realize a non-invasive brain-computer interface (Brain-computer interface, BCI). Researchers continue to try to analyze complex brain activities from EEG signals. As a typical solution, machine learning can realize the recognition of various brain activities based on EEG signal data by processing and training a large amount of EEG data (ie EEG data) collected manually. However, it is difficult for a single EEG device user to generate enough required data for machine learning processing and training within a certain period of time, so that there is a problem of scarcity of EEG data for processing and training.

At the same time, the type of EEG data is different due to the number of electrodes, electrode positions, electrode types (dry electrodes, wet electrodes), sampling rate and other factors of the acquisition equipment, resulting in heterogeneity of EEG data due to different equipment. Different types of EEG data for heterogeneity often need to be processed separately to train different machine learning models. In this way, the scarcity of data used for processing training is exacerbated by heterogeneity.

In addition, EEG data has strong individual differences and can be used to reflect information such as individual fatigue, panic, alertness, and behavioral intentions; if the original EEG data is leaked and abused, it may lead to serious privacy violations , so that the privacy sensitivity of EEG data needs to be carefully considered in the process of processing EEG data through machine learning. To sum up, due to the scarcity, heterogeneity and privacy characteristics of EEG data, it is necessary to obtain a large amount of EEG data in various ways, resulting in low efficiency of processing EEG data by current machine learning methods, and It is difficult to process EEG data under the premise of fully protecting user privacy.

Contents of the invention

The main purpose of the present invention is to provide a method, device, device, and medium for processing EEG data based on federated migration, aiming to solve the problem of low processing efficiency of EEG data in the prior art and the difficulty of fully protecting user privacy. Technical issues of processing EEG data.

In order to achieve the above object, the present invention provides a method for processing EEG data based on federated migration, which includes the following steps:

Obtaining multiple sets of EEG data, and transmitting the multiple sets of EEG data to the local terminal respectively, and performing mapping processing on feature mapping models respectively corresponding to the multiple sets of EEG data, to generate standard EEG data;

The standard EEG data is transmitted to the brain activity recognition model in the local terminal for processing, and the generated model parameters are transmitted to the preset public server, so that the preset public server can compare the model parameters and the information sent by at least one other terminal. Other model parameters are aggregated to generate update parameters;

receiving the update parameters, updating the brain activity recognition model, and processing the standard EEG data obtained by remapping each feature mapping model based on the updated brain activity recognition model until the brain activity recognition model The generated total loss value is less than the preset value, and the processing of multiple sets of EEG data based on federated migration is completed.

Optionally, after completing the step of processing multiple sets of EEG data based on federated migration, the method further includes:

When the EEG data to be processed is received, the feature mapping model is called to perform mapping processing;

The EEG data to be processed by mapping is transmitted to the brain activity recognition model for classification processing, and the brain activity type corresponding to the EEG data to be processed is obtained.

Optionally, the multiple sets of EEG data are respectively transmitted to the local terminal, and the feature mapping models respectively corresponding to the multiple sets of EEG data are mapped, and the step of generating standard EEG data includes:

Preprocessing the multiple sets of EEG data according to the data attributes of the multiple sets of EEG data;

Transmitting the preprocessed multiple sets of EEG data to the local terminal respectively, and mapping the multiple sets of EEG data respectively according to the feature mapping models corresponding to the multiple sets of EEG data in the local terminal processing to generate the standard EEG data.

Optionally, after the step of processing the standard EEG data obtained by remapping each feature mapping model based on the updated brain activity recognition model, the method includes:

Obtain the processing result generated by processing the standard EEG data obtained by the remapping process with the updated brain activity recognition model, and determine the classification loss value corresponding to the updated brain activity recognition model according to the processing result ;

According to the standard EEG data processed by remapping, determine the dimensionality reduction loss value corresponding to each of the feature mapping models;

A total loss value of the updated brain activity recognition model is determined according to the classification loss value and the dimensionality reduction loss value.

Optionally, the step of determining the classification loss value corresponding to the updated brain activity recognition model according to the processing result includes:

reading the reference result of the standard EEG data, and performing similarity calculation between the reference result and the processing result to generate a first similarity value;

A classification loss value of the updated brain activity recognition model is determined according to the first similarity value.

Optionally, the step of determining the dimensionality reduction loss value corresponding to each of the feature mapping models according to the standard EEG data processed by remapping includes:

According to the corresponding relationship between the remapped standard EEG data and multiple sets of the EEG data, the remapped standard EEG data is divided into multiple sets of standard sub-data;

performing two-two combinations on multiple sets of standard sub-data, and performing reproduction processing on each set of standard sub-data in each combination to generate two sets of data points in each combination;

Perform mean value processing on each group of data points in each described combination, generate the average value of each group of data points in each described combination, and generate the average value of each described combination according to the average value of each group of data points in each described combination second similarity value;

A combined dimensionality reduction loss value of each combination is determined according to the second similarity value of each combination, and a dimensionality reduction loss value corresponding to each feature mapping model is generated according to each combination dimensionality reduction loss value.

Optionally, after the step of determining the total loss value of the updated brain activity recognition model according to the classification loss value and the dimensionality reduction loss value, the method further includes:

judging whether the total loss value is less than a preset value within a preset number of times, and if it is less than a preset value within a preset number of times, it is determined that the total loss value generated by the brain activity recognition model after updating is less than a preset value value;

If it is not less than the preset value within the preset times, the step of receiving the update parameters and updating the brain activity recognition model is performed.

Further, in order to achieve the above purpose, the present invention also provides an EEG data processing device based on federated migration, the EEG data processing device based on federated migration includes:

The acquiring module is used to acquire multiple sets of EEG data, and transmit the multiple sets of EEG data to the local terminal respectively, perform mapping processing on feature mapping models corresponding to the multiple sets of EEG data, and generate standard EEG data data;

A generating module, configured to transmit the standard EEG data to the brain activity recognition model in the local terminal for processing, and transmit the generated model parameters to a preset public server, so that the preset public server can compare the model parameters and at least Aggregate other model parameters sent by another terminal to generate updated parameters;

An update module, configured to receive the update parameters, update the brain activity recognition model, and process the standard EEG data obtained by remapping each feature mapping model based on the updated brain activity recognition model until the The total loss value generated by the brain activity recognition model is less than a preset value, and the processing of multiple sets of EEG data based on federated migration is completed.

Further, in order to achieve the above purpose, the present invention also provides an EEG data processing device based on federated migration, which includes a memory, a processor, and a memory stored on the memory and can be The federated migration-based EEG data processing program running on the processor, when the federated migration-based EEG data processing program is executed by the processor, implements the federated migration-based EEG data processing method as described above A step of.

Further, in order to achieve the above object, the present invention also provides a medium, on which a federated migration-based EEG data processing program is stored, and when the federated migration-based EEG data processing program is executed by a processor, the above The steps of the EEG data processing method based on federated migration.

The EEG data processing method based on federated migration of the present invention transmits multiple sets of EEG data acquired by multiple devices to the local terminal respectively, performs mapping processing with the feature mapping model connected to multiple devices, and generates a standard EEG data of the local terminal. After that, the standard EEG data is transmitted to the brain activity recognition model in the local terminal for processing, and the generated model parameters are transmitted to the preset public server for federated migration training. The model parameters of the preset public server and at least come from one The other model parameters of the other terminal are aggregated, and the updated parameters are generated and returned to the local terminal; the brain activity recognition model is updated by receiving the update parameters. Data remapping processing to obtain standard EEG data, based on the updated brain activity recognition model to process the remapped standard EEG data until the total loss value generated by the brain activity recognition model is less than the preset value, representing brain activity The recognition model converges and has a good processing effect before completing the processing of multiple sets of EEG data based on federated migration. Because multiple sets of EEG data are acquired by multiple devices, the scarcity problem of EEG data is avoided; at the same time, by mapping and processing each set of EEG data, each EEG data meets the processing requirements, avoiding the Each group of EEG data comes from different devices, resulting in the problem of heterogeneity; then, the standard EEG data of each terminal is processed through federated transfer learning, because the standard EEG data of each terminal in the process of federated transfer learning is retained in the terminal Local, so as to avoid the leakage of private information. Therefore, it is possible to process EEG data under the premise of fully protecting user privacy, and because EEG data does not have the problems of scarcity and heterogeneity, it is not necessary to obtain a large amount of EEG data in various ways, making the processing Efficiency is greatly improved.

Description of drawings

FIG. 1 is a schematic structural diagram of the device hardware operating environment involved in the embodiment scheme of the federated migration-based EEG data processing device of the present invention;

FIG. 2 is a schematic flow chart of the first embodiment of the federated migration-based EEG data processing method of the present invention;

FIG. 3 is a schematic diagram of functional modules of a preferred embodiment of the federated migration-based EEG data processing device of the present invention.

The realization of the purpose of the present invention, functional characteristics and advantages will be further described in conjunction with the embodiments and with reference to the accompanying drawings.

Detailed ways

It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

The present invention provides an EEG data processing device based on federated migration. Referring to FIG. 1 , FIG. 1 is a schematic structural diagram of the hardware operating environment of the device involved in the embodiment of the federated migration-based EEG data processing device of the present invention.

As shown in FIG. 1 , the federated migration-based EEG data processing device may include: a processor 1001 , such as a CPU, a communication bus 1002 , a user interface 1003 , a network interface 1004 , and a memory 1005 . Wherein, the communication bus 1002 is used to realize connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. Optionally, the network interface 1004 may include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 can be a high-speed RAM memory, or a stable memory (non-volatile memory), such as a disk memory. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001 .

Those skilled in the art can understand that the hardware structure of the federated migration-based EEG data processing device shown in FIG. 1 does not constitute a limitation to the federated migration-based EEG data processing device, and may include more or more Fewer components, or combinations of certain components, or different arrangements of components.

As shown in FIG. 1 , the memory 1005 as a medium may include an operating system, a network communication module, a user interface module, and an EEG data processing program based on federation migration. Among them, the operating system is a program that manages and controls EEG data processing equipment and software resources based on federated migration, and supports the operation of network communication modules, user interface modules, EEG data processing programs based on federated migration, and other programs or software; network The communication module is used to manage and control the network interface 1004; the user interface module is used to manage and control the user interface 1003.

In the hardware structure of the EEG data processing device based on federated migration shown in Figure 1, the network interface 1004 is mainly used to connect to the background server and perform data communication with the background server; the user interface 1003 is mainly used to connect to the client (client), Perform data communication with the client; the processor 1001 can call the federated migration-based EEG data processing program stored in the memory 1005, and perform the following operations:

Obtaining multiple sets of EEG data, and transmitting the multiple sets of EEG data to the local terminal respectively, and performing mapping processing on feature mapping models respectively corresponding to the multiple sets of EEG data, to generate standard EEG data;

The standard EEG data is transmitted to the brain activity recognition model in the local terminal for processing, and the generated model parameters are transmitted to the preset public server, so that the preset public server can compare the model parameters and the information sent by at least one other terminal. Other model parameters are aggregated to generate update parameters;

receiving the update parameters, updating the brain activity recognition model, and processing the standard EEG data obtained by remapping each feature mapping model based on the updated brain activity recognition model until the brain activity recognition model The generated total loss value is less than the preset value, and the processing of multiple sets of EEG data based on federated migration is completed.

Further, after the step of processing multiple sets of EEG data based on federated migration is completed, the processor 1001 may call the federated migration-based EEG data processing program stored in the memory 1005, and perform the following operations:

When the EEG data to be processed is received, the feature mapping model is called to perform mapping processing;

The EEG data to be processed by mapping is transmitted to the brain activity recognition model for classification processing, and the brain activity type corresponding to the EEG data to be processed is obtained.

Further, the step of transmitting the plurality of sets of EEG data to the local terminal respectively, performing mapping processing on feature mapping models corresponding to the plurality of sets of EEG data, and generating standard EEG data includes:

Preprocessing the multiple sets of EEG data according to the data attributes of the multiple sets of EEG data;

Transmitting the preprocessed multiple sets of EEG data to the local terminal respectively, and mapping the multiple sets of EEG data respectively according to the feature mapping models corresponding to the multiple sets of EEG data in the local terminal processing to generate the standard EEG data.

Further, after the step of processing the standard EEG data obtained by remapping each feature mapping model based on the updated brain activity recognition model, the processor 1001 may call the federated migration-based EEG data stored in the memory 1005. data handler, and do the following:

Obtain the processing result generated by processing the standard EEG data obtained by the remapping process with the updated brain activity recognition model, and determine the classification loss value corresponding to the updated brain activity recognition model according to the processing result ;

According to the standard EEG data processed by remapping, determine the dimensionality reduction loss value corresponding to each of the feature mapping models;

A total loss value of the updated brain activity recognition model is determined according to the classification loss value and the dimensionality reduction loss value.

Further, the step of determining the classification loss value corresponding to the updated brain activity recognition model according to the processing result includes:

reading the reference result of the standard EEG data, and performing similarity calculation between the reference result and the processing result to generate a first similarity value;

A classification loss value of the updated brain activity recognition model is determined according to the first similarity value.

Further, the step of determining the dimensionality reduction loss value corresponding to each of the feature mapping models according to the standard EEG data processed by remapping includes:

According to the corresponding relationship between the remapped standard EEG data and multiple sets of the EEG data, the remapped standard EEG data is divided into multiple sets of standard sub-data;

performing two-two combinations on multiple sets of standard sub-data, and performing reproduction processing on each set of standard sub-data in each combination to generate two sets of data points in each combination;

Perform mean value processing on each group of data points in each described combination, generate the average value of each group of data points in each described combination, and generate the average value of each described combination according to the average value of each group of data points in each described combination second similarity value;

A combined dimensionality reduction loss value of each combination is determined according to the second similarity value of each combination, and a dimensionality reduction loss value corresponding to each feature mapping model is generated according to each combination dimensionality reduction loss value.

Further, after the step of determining the total loss value of the updated brain activity recognition model according to the classification loss value and the dimensionality reduction loss value, the processor 1001 may call the federated migration-based The EEG data processing program, and perform the following operations:

judging whether the total loss value is less than a preset value within a preset number of times, and if it is less than a preset value within a preset number of times, it is determined that the total loss value generated by the brain activity recognition model after updating is less than a preset value value;

If it is not less than the preset value within the preset times, the step of receiving the update parameters and updating the brain activity recognition model is performed.

The specific implementation of the federated migration-based EEG data processing device of the present invention is basically the same as the following embodiments of the federated migration-based EEG data processing method, and will not be repeated here.

The invention also provides a federated migration-based EEG data processing method.

Referring to FIG. 2 , FIG. 2 is a schematic flowchart of a first embodiment of a method for processing EEG data based on federated migration in the present invention.

The embodiment of the present invention provides an embodiment of the EEG data processing method based on federated migration. It should be noted that although the logical order is shown in the flow chart, in some cases, the order may be different from that here. Perform the steps shown or described. Specifically, the federated migration-based EEG data processing method in this embodiment includes:

Step S10, acquiring multiple sets of EEG data, and transmitting the multiple sets of EEG data to the local terminal, and performing mapping processing on the feature mapping models corresponding to the multiple sets of EEG data, to generate standard EEG data.

The federated migration-based EEG data processing method in this embodiment is applied to federated-migrated local terminals, and is suitable for processing EEG data in cooperation with at least one other terminal through the local terminal through federated migration learning. Since the processing process of the EEG data inside the local terminal is similar to that inside the terminal of the other party, this embodiment first uses the local terminal as an example to describe the internal processing process. Among them, electroencephalogram (EEG) refers to scalp electroencephalogram, which is an electrical signal recorded by electrodes placed on the scalp and amplified by a signal amplifier. It is the synaptic activity of synchronous discharge of nerve cell clusters in the cerebral cortex to produce potential changes. General reflection on the scalp. EEG contains a wealth of brain activity information. When a person is in a state of high alertness and performing tasks of high difficulty, the main components of brain electrical activity tend to be β waves with lower amplitude and higher frequency; When alertness is low, the alpha wave activity is enhanced; when the person is sleepy, theta wave is significantly enhanced. It is generally believed that psychological stress, active thinking and attention will promote the movement of brain electrical activity to higher frequency bands and inhibit the activity of alpha waves. That is, EEG can be used to reflect information such as individual fatigue, panic, alertness, and behavioral intentions.

In this embodiment, the local terminal is connected with various acquisition devices capable of collecting EEG data, such as sensors, recorders, etc., so as to collect EEG data, ie, EEG data, through various acquisition devices. The collected EEG data are grouped according to each collection device, and the EEG data collected by the same collection device are regarded as a group of EEG data. Because the types of EEG data collected by the same collection device are the same, the data types of each group of EEG data are the same, but the types of EEG data between different groups are different. If a collection device includes two electrodes for collecting EEG data in the state of eyes closed and breathing, the EEG data formed by collecting EEG data from the collection device represent the data in the state of eyes closing and breathing, belong to the same data type; the other collection device contains two electrodes for collecting EEG data under the state of sound stimulation and light stimulation, and the EEG data formed by collecting EEG data by the collection device both represent light stimulation and sound stimulation The data in the state, belong to the same data type.

Further, multiple feature mapping models are set in the local terminal, and each feature mapping model is connected to an acquisition device, and each group of EEG data is transmitted according to the feature mapping model connected to the acquisition device, and transmitted to the respective connected The mapping process is carried out in the feature mapping model. The essence of the mapping process is the standardized processing of the format. After processing, the standard EEG data of each terminal is obtained, so that the EEG data collected by each acquisition device are consistent in format, and it is convenient to use federated transfer learning to pair the EEG data with the same format. EEG data are processed. Specifically, the multiple sets of EEG data are respectively transmitted to the local terminal, and the feature mapping models corresponding to the multiple sets of EEG data are mapped, and the steps of generating standard EEG data include:

Step S11, performing preprocessing on multiple sets of EEG data according to data attributes of multiple sets of EEG data;

Step S12, transmitting the preprocessed multiple sets of EEG data to the local terminal respectively, and according to the feature mapping models in the local terminal corresponding to the multiple sets of EEG data, for multiple sets of the EEG data Mapping processing is performed respectively to generate the standard EEG data.

Understandably, the EEG data collected by different acquisition devices have different attributes, which may be current data, voltage data, or electromagnetic wave data; and the collected raw EEG data may have factors that are not suitable for direct processing , but preprocessing is required. The way of this preprocessing is different according to different properties, and it includes at least two parts. The first part of preprocessing includes pre-amplification, high-pass filtering, low-pass filtering, post-stage amplification, level shifting, etc., and the second part of preprocessing Including A/D sampling, 50HZ notch and baseline drift and other signal processing methods. After collecting multiple sets of EEG data, preprocessing is carried out according to the data attributes of each EEG data in each set of EEG data, until all the EEG data in each set of EEG data have been pre-processed before being transmitted to their respective The mapping process is performed in the terminal.

Further, the multiple feature mapping models set in the local terminal are essentially fully connected neural networks, which can be two layers, three layers or more layers, and each layer is equipped with such as ReLU activation function. After the vectors of the previous layer are weighted and calculated to obtain the calculation results, the activation function is activated to transmit the calculation results to the next layer to continue the calculation until the calculation of each layer is completed. Through the feature mapping model connected to each acquisition device in the local terminal, the dimensionality reduction mapping process is performed on the preprocessed EEG data transmitted to it, and the interference data or EEG data that does not meet the requirements are screened out to reduce dimensionality. The processing of unnecessary data; at the same time, the necessary EEG data is mapped to meet the processing standards.

It should be noted that the output layer dimension of the neural network used as the feature mapping model in the local terminal is the same, so that after each group of EEG data is processed by each feature mapping model, it has standard EEG data of the same dimension, which is convenient for federated transfer learning. The consistency of processing dimensions of standard EEG data is conducive to the accuracy of processing results.

In addition, in order to facilitate the realization of mapping, a public feature subspace can be set in the local terminal, and each feature mapping model maps each group of EEG data after mapping to the public feature subspace; The data processed by the mapping all have common characteristics that meet the processing standards. After the local terminal finishes processing each group of EEG data transmitted thereto, the standard EEG data of the local terminal is obtained. Furthermore, the standard EEG data can be processed by means of federated transfer learning to obtain the model parameters of the brain activity recognition model used to recognize brain activity, so that the brain activity recognition model can identify brain activity more accurately.

Step S20, the standard EEG data is transmitted to the brain activity recognition model in the local terminal for processing, and the generated model parameters are transmitted to the preset public server, so that the preset public server can compare the model parameters with at least one other party The other model parameters sent by the terminal are aggregated to generate update parameters;

With the development of science and technology, "machine learning" has become one of the core research fields of artificial intelligence, and how to continue machine learning on the premise of protecting data privacy and meeting legal and compliance requirements is a trend that the field of machine learning is now concerned about , In this context, people have proposed the concept of "federated learning". Federated learning uses technical algorithms to encrypt the built model. Both sides of the federation can also conduct model training to obtain model parameters without giving their own data. Federated learning protects user data privacy through parameter exchange under the encryption mechanism, data and the model itself It will not be transmitted, nor can it guess the other party's data, so there is no possibility of leakage at the data level, and it will not violate stricter data protection laws such as GDPR (General Data Protection Regulation, "General Data Protection Regulation"), etc. While maintaining a high degree of data integrity, data privacy is guaranteed.

The federated transfer learning in this embodiment is a type of federated learning, which refers to the use of transfer learning to overcome data or Insufficient labeling. In this implementation, multiple sets of acquired EEG data are transferred to different feature mapping models for mapping processing. After obtaining the standard EEG data of the local terminal, the standard EEG data of the local terminal forms a data set, which can be learned through federated transfer learning. Process it to generate update parameters for updating the brain activity recognition model in the local terminal; by using a large amount of collected EEG data as the basis for updating the brain activity recognition model, the brain activity recognition model can recognize brain activity more accurate.

Furthermore, multiple feature mapping models in the local terminal are connected to the EEG recognition model in the local terminal. In the process of processing the standard EEG data of the local terminal with the help of federated transfer learning, the mapped standard EEG The electrical data is transmitted to the EEG recognition model in the local terminal, and processed by the EEG recognition model to obtain model parameters of the EEG recognition model in the local terminal. The model parameters are essentially model gradients used to update and optimize the parameters of the EEG recognition model, so as to make the EEG recognition more accurate.

It should be noted that after the preprocessing of each group of EEG data, manual intervention is used to label each group of EEG data, and then the EEG data carrying the labeled information are mapped to the common feature subspace respectively. By transmitting various data in the public feature subspace and their corresponding labeling information to the EEG recognition model for processing, the training of the EEG recognition model is realized, and update parameters for optimizing and updating the EEG recognition model are generated.

Furthermore, both the local terminal and the other party’s terminal are connected to the pre-set preset public server. While the local terminal is processing its internal standard EEG data to generate model parameters, the other party’s terminal is also processing its internal standard EEG data. The data is processed to generate other model parameters of other terminals. The local terminal encrypts the generated model parameters. The encryption method can be an additive homomorphic encryption such as Paillier's scheme, and transmits the encrypted model parameters to the preset public server. The weight value preset by each terminal performs weighted aggregation on the model parameters and model parameters from at least one other terminal to generate update parameters for updating the parameters of the brain activity recognition model.

Step S30, receiving the updated parameters, updating the brain activity recognition model, and processing the standard EEG data obtained by remapping each feature mapping model based on the updated brain activity recognition model until the brain activity The total loss value generated by the activity recognition model is less than a preset value, and the processing of multiple sets of EEG data based on federated migration is completed.

Further, the preset public server sends the generated update parameters to each terminal respectively, and after receiving the update parameters, the local terminal uses it to update the parameters of the brain activity recognition model in the local terminal, so as to realize brain activity recognition Model updates. Thereafter, calculate the total loss value for the updated brain activity recognition model, and set a preset value for determining the size of the total loss value; compare the calculated total loss value with the preset value, and determine whether the total loss value is less than the preset value . If it is less than the preset value, it means that the brain activity recognition model updated based on a large amount of EEG data has a small loss value for EEG data processing; it is determined that the training of the brain activity recognition model is completed, and the processing of various EEG data is completed. If it is not less than the preset value, it means that the loss value of the EEG data processing is relatively large, and training needs to be continued. Among them, during the update process of the brain activity recognition model, multiple feature mapping models are also remapping their respective EEG data to obtain standard EEG data. When continuing training, process the remapped standard EEG data based on the updated brain activity recognition model, generate new model parameters and transmit them to the preset public server, obtain updated parameters to update the brain activity recognition model, and so on. Until the total loss value generated by the updated brain activity recognition model is less than the preset value, the processing of multiple sets of EEG data based on federated migration is completed.

Furthermore, after the training is completed to obtain the brain activity recognition model for the classification and recognition of brain activity, if the EEG data that needs to be classified and recognized is received, the EEG data is used as the EEG data to be processed. The feature mapping model in the local terminal is called for mapping processing, and the EEG data to be processed after mapping is transmitted to the brain activity recognition model for classification, and the type of brain activity is obtained to reflect the brain activity reflected by the EEG data to be processed.

The EEG data processing method based on federated migration of the present invention transmits multiple sets of EEG data acquired by multiple devices to the local terminal respectively, performs mapping processing with the feature mapping model connected to multiple devices, and generates a standard EEG data of the local terminal. After that, the standard EEG data is transmitted to the brain activity recognition model in the local terminal for processing, and the generated model parameters are transmitted to the preset public server for federated migration training. The model parameters of the preset public server and at least come from one The other model parameters of the other terminal are aggregated, and the updated parameters are generated and returned to the local terminal; the brain activity recognition model is updated by receiving the update parameters. Data remapping processing to obtain standard EEG data, based on the updated brain activity recognition model to process the remapped standard EEG data until the total loss value generated by the brain activity recognition model is less than the preset value, representing brain activity The recognition model converges and has a good processing effect before completing the processing of multiple sets of EEG data based on federated migration. Because multiple sets of EEG data are acquired by multiple devices, the scarcity problem of EEG data is avoided; at the same time, by mapping and processing each set of EEG data, each EEG data meets the processing requirements, avoiding the Each group of EEG data comes from different devices, resulting in the problem of heterogeneity; then, the standard EEG data of each terminal is processed through federated transfer learning, because the standard EEG data of each terminal in the process of federated transfer learning is retained in the terminal Local, in order to avoid the leakage of private information. Therefore, it is possible to process EEG data under the premise of fully protecting user privacy, and because EEG data does not have the problems of scarcity and heterogeneity, it is not necessary to obtain a large amount of EEG data in various ways, making the processing Efficiency is greatly improved.

Further, based on the first embodiment of the federated migration-based EEG data processing method of the present invention, a second embodiment of the federated migration-based EEG data processing method of the present invention is proposed.

The difference between the second embodiment of the federated migration-based EEG data processing method and the federated migration-based EEG data processing method in the first embodiment is that the updated brain activity recognition model based on each feature map After the step of processing the standard EEG data obtained by the model remapping process, it also includes:

Step a1, obtaining the processing result generated by processing the standard EEG data obtained by the remapping process with the updated brain activity recognition model, and according to the processing result, determine the corresponding brain activity recognition model after the update. The classification loss value;

Step a2, according to the standard EEG data processed by remapping, determine the dimensionality reduction loss value corresponding to each of the feature mapping models;

In this embodiment, after the brain activity recognition model is updated by updating parameters, the total loss value corresponding to the updated brain activity recognition model is also calculated. The total loss value includes at least two aspects; one is the classification loss value between the predicted classification result obtained by the brain activity recognition model for classifying each EEG data and the actual classification result of each EEG data; Dimensionality reduction mapping is performed on each EEG data for the feature mapping model, resulting in the dimensionality reduction loss value.

Further, while the local terminal is updating the brain activity recognition model therein by preset update parameters returned by the public server, the feature mapping model of the local terminal is also updating its own training. Each feature mapping model maps each group of preprocessed EEG data transmitted to it in different mapping ways, so as to optimize the standard EEG data obtained through each mapping. Therefore, the updated brain activity recognition model classifies and recognizes the standard EEG data obtained by the remapping process, generates processing results, and then determines the updated brain activity recognition model through the loss between the processing results and the real reference results The corresponding classification loss value. For the standard EEG data obtained by remapping multiple sets of EEG data, if the similarity of the mapped standard EEG data between each set of EEG data is higher, it means that the mapping between each feature mapping model The more similar, the smaller the dimensionality reduction loss. Therefore, the dimensionality reduction loss value corresponding to the mapping process of the EEG data by each feature mapping model can be determined through the similarity between the standard EEG data obtained by remapping multiple sets of EEG data.

Wherein, the step of determining the classification loss value corresponding to the updated brain activity recognition model according to the processing result includes:

Step a11, reading the reference result of the standard EEG data, and performing similarity calculation between the reference result and the processing result to generate a first similarity value;

Step a12, according to the first similarity value, determine the classification loss value of the brain activity recognition model after updating.

Further, the updated brain activity recognition model in the local terminal classifies and recognizes the remapped standard EEG data to generate a classification result, which is the processing result. After obtaining the processing result, the real classification result of the standard EEG data is determined according to the annotation information carried by the standard EEG data, and the real classification result is read as a reference result of the standard EEG data. Further, the similarity calculation is performed on the reference result and the processing result to generate a first similarity value, so as to represent that the local terminal classifies the standard EEG data through the first similarity value, and the obtained classification result is consistent with the real classification result The difference between, that is, the classification accuracy of the local terminal midbrain activity recognition model.

Furthermore, the similarity between the processing results and the reference results represents the accuracy of the classification of the brain activity recognition model. The higher the similarity between the two, the more accurate the classification, and the smaller the classification loss caused by the classification; otherwise, if The lower the similarity, the more inaccurate the classification, and the greater the classification loss value caused by the classification. Therefore, in order to represent the size of the classification loss value by the degree of similarity, the corresponding relationship between the similarity value and the classification loss value can be preset. After generating the first similarity value of the local terminal, compare the first similarity value with the corresponding relationship to determine the corresponding classification loss value in the corresponding relationship, and the corresponding classification loss value is the updated value of the local terminal. The classification loss value of the brain activity recognition model.

Further, the step of determining the dimensionality reduction loss value corresponding to each of the feature mapping models according to the standard EEG data processed by remapping includes:

Step a21, dividing the remapped standard EEG data into multiple sets of standard sub-data according to the correspondence between the remapped standard EEG data and multiple sets of the EEG data;

Step a22, combining multiple sets of standard sub-data in pairs, and reproducing each set of standard sub-data in each combination to generate two sets of data points in each combination;

Step a23, performing mean value processing on each group of data points in each of the combinations, generating the average value of each group of data points in each of the combinations, and generating the average value of each group of data points in each of the combinations. The second similarity value of the combination;

Step a24: Determine the combined dimensionality reduction loss value of each combination according to the second similarity value of each combination, and generate a dimensionality reduction value corresponding to each feature mapping model according to each combination dimensionality reduction loss value loss value.

In order to reflect the degree of similarity of the remapped standard EEG data among multiple sets of EEG data, this embodiment is based on the difference between the standard EEG data obtained through remapping processing and the original multiple sets of EEG data before processing. The corresponding relationship is to divide the remapped standard EEG data into multiple sets of standard sub-data, and a set of standard sub-data is correspondingly generated by a set of EEG data. Thereafter, each group of standard sub-data is permuted and combined in pairs to obtain a plurality of combinations each including two groups of standard sub-data. For the three sets of standard sub-data A, B, and C involved, the three are arranged and combined in pairs, and the resulting combination includes AB, AC, and BC. Thereafter, each group of standard sub-data in each combination is reproduced separately through a method such as the kernel Hilbert space, and each standard EEG data contained in it is reproduced from the common feature subspace to the kernel through the feature map. The Hilbert subspace realizes the reproduction of standard EEG data as data points, so as to obtain two sets of data points in the combination. For example, for the above-mentioned combination BC, where the standard EEG data contained in the two sets of standard sub-data B and C are [b1, b2, b3, b4, b5] and [c1, c2, c3, c4, c5] respectively, after reproduction Come out and get two sets of data points [m1, m2, m3, m4, m5] and [n1, n2, n3, n4, n5].

Further, mean value processing is performed on each group of data points in each combination to generate the average value of each group of data points in each combination; and then the similarity calculation is performed on the average values of the two groups of data points in each combination to generate each combination The second similarity value in . Among them, the essence of the mean value processing for each group of data points is to calculate the centroid points of each data point contained in each group, and accordingly, the similarity calculation of the average value is to calculate the distance between two centroid points. For the above combination BC, calculate the centroid point m of the data points [m1, m2, m3, m4, m5] of the standard sub-data B, and the centroid of the data points [n1, n2, n3, n4, n5] of the standard sub-data C Point n represents the second similarity value between standard sub-data B and C through the distance between m and n.

Furthermore, after each combination generates the second similarity value, each second similarity value can represent the degree of similarity of the standard sub-data obtained by mapping the EEG data through the respective feature mapping models in each combination. The higher the second similarity value of the combination, the higher the similarity of the feature mapping model corresponding to the combination to the EEG data mapping processing, and the smaller the combined dimensionality reduction loss value of the combination; otherwise, it represents the EEG data mapping processing. The lower the similarity of , the greater the combined dimensionality reduction loss value of the combination. Therefore, in order to characterize the magnitude of the combined dimensionality reduction loss value through the degree of similarity, the corresponding relationship between the similarity value and the combined dimensionality reduction loss value can be preset. After generating the second similarity value of each combination, compare the second similarity value with the corresponding relationship to determine the corresponding combination dimensionality reduction loss value in the correspondence relationship, and the corresponding combination dimensionality reduction loss value is the combination The combined dimensionality reduction loss value caused by the corresponding feature mapping model performing dimensionality reduction mapping on the respective EEG data. In this way, the combined dimensionality reduction loss value of each combination can be obtained. Furthermore, the combined dimensionality reduction loss values are added and processed to obtain the dimensionality reduction loss value representing the overall loss caused by each feature mapping model in the mapping process of EEG data, that is, the dimensionality reduction loss value corresponding to the mapping process with the local terminal. loss value.

Step S33, according to the classification loss value and the dimensionality reduction loss value, determine the total loss value of the brain activity recognition model after updating.

Further, after generating the classification loss value of the updated brain activity recognition model in the local terminal, and representing the overall dimensionality reduction loss value, the classification loss value and the dimensionality reduction loss value are summed to generate an updated The total loss value of each brain activity recognition model after the current brain activity recognition model is updated, the brain activity recognition model classifies the EEG data, and the degree of similarity between the obtained classification result and the real result is high or low. The smaller the total loss value, the higher the similarity of the representation, and the more accurate the classification of the brain activity recognition model; otherwise, the less accurate the classification of the representation is, and it is necessary to continue to update the training until the total loss value is smaller and the classification of the representation is accurate.

In this implementation, the classification loss value corresponding to the brain activity recognition model and the dimensionality reduction loss value corresponding to the feature mapping model reflect the total loss value of the brain activity recognition model. Fully consider the standard EEG data obtained by mapping the EEG data with the feature mapping model, the impact of the resulting brain activity recognition model on the classification accuracy of the standard EEG data, and the impact of the brain activity recognition model itself on the standard EEG data. impact on classification accuracy. It makes the generation of the total loss value more accurate, completes the processing of standard EEG data according to the total loss value, and ends the training of the brain activity recognition model more accurately, which is conducive to improving the classification of EEG data by the brain activity recognition model accuracy.

Further, based on the second embodiment of the federated migration-based EEG data processing method of the present invention, a third embodiment of the federated migration-based EEG data processing method of the present invention is proposed.

The difference between the third embodiment of the federated migration-based EEG data processing method and the second embodiment of the federated migration-based EEG data processing method is that according to the classification loss value and the dimensionality reduction loss value , after the step of determining the total loss value of the updated brain activity recognition model, further comprising:

Step S34, judging whether the total loss value is less than the preset value within the preset number of times, if it is less than the preset value within the preset number of times, then determine the total loss value generated by the brain activity recognition model after the update less than the preset value;

Step S35, if not less than the preset value within the preset times, execute the step of receiving the update parameters and updating the brain activity recognition model.

In this embodiment, in order to ensure the stability of the accurate classification of the brain activity recognition model and eliminate the influence of accidental factors, a mechanism is provided to determine whether the total loss value is less than the preset value within a preset number of times. Among them, the preset number of times is preferably a continuous number of times, and can also be set to a number of times with fewer intervals. If the preset number of times is set to 5 times, then it is preferably set so that the total loss value generated within 5 consecutive times is less than the preset time. The value may also be set such that among the total loss values generated for six consecutive times, there are five total loss values that are less than the preset value.

Further, after the brain activity recognition model in the local terminal is updated and generates the total loss value, it calls the preset value and compares it with the total loss value to determine whether the total loss value is less than the preset value; if it is less than the preset value, then Perform cumulative counting to obtain the count value. Furthermore, by judging whether the count value is greater than or equal to the preset number of times, it is judged whether the total loss value is less than the preset value within the preset number of times; if the count value is greater than or equal to the preset number of times, it means that the updated brain activity recognition model has The characteristic of accurate classification, it is determined that the total loss value is less than the preset value within the preset number of times. Then it is determined that the total loss value generated by the updated brain activity recognition model is less than the preset value, the processing of multiple sets of EEG data is completed, and the training of the brain activity recognition model based on federated migration is ended. Subsequently, the updated brain activity recognition model is used to classify each EEG data, so as to achieve the purpose of accurate classification.

Furthermore, if the count value is less than the preset number of times, it is determined that the total loss value is not less than the preset value within the preset number of times, which is not enough to show that the updated brain activity recognition model has the characteristics of accurate classification. At this time, it is determined that the total loss value is not less than the preset value within the preset number of times, and the brain activity recognition model needs to continue to be based on federated transfer learning, receive the update parameters returned by the preset public server, update the brain activity recognition model, and The standard EEG data is processed to train the brain activity recognition model until the generated total loss value is less than the preset value within the preset times.

It should be noted that when continuing to process standard EEG data, since each feature mapping model will map each group of EEG data in different ways, the essence of continuing processing of standard EEG data is to The standard EEG data mapped by the new mapping method are processed. In this way, in the case of a small amount of data, a large amount of standard EEG data can be obtained through different mapping methods, and then the brain activity recognition model can be trained through a large amount of standard EEG data, so that the brain activity recognition obtained after training The model is more accurate in processing EEG data.

In this embodiment, the mechanism for processing EEG data is completed by setting the total loss value generated by the continuously updated brain activity recognition model to be less than the preset value, ensuring the stability of accurate classification of the brain activity recognition model and eliminating accidental factors The impact on classification is beneficial to the accurate classification of EEG data by the brain activity recognition model.

The invention also provides an EEG data processing device based on federation migration.

Referring to FIG. 3 , FIG. 3 is a schematic diagram of functional modules of the first embodiment of the federated migration-based EEG data processing device of the present invention. The EEG data processing device based on federation migration includes:

The acquiring module 10 is configured to acquire multiple sets of EEG data, and transmit the multiple sets of EEG data to the local terminal respectively, perform mapping processing on the feature mapping models respectively corresponding to the multiple sets of EEG data, and generate a standard EEG data. electrical data;

The generating module 20 is used to transmit the standard EEG data to the brain activity recognition model in the local terminal for processing, and transmit the generated model parameters to the preset public server, so that the preset public server can compare the model parameters and Aggregating other model parameters sent by at least one other terminal to generate updated parameters;

The update module 30 is configured to receive the update parameters, update the brain activity recognition model, and process the standard EEG data obtained by remapping each feature mapping model based on the updated brain activity recognition model until The total loss value generated by the brain activity recognition model is less than a preset value, and the processing of multiple sets of EEG data based on federated migration is completed.

Further, the EEG data processing device based on federation migration also includes:

A calling module, configured to call the feature mapping model for mapping processing when receiving the EEG data to be processed;

The processing module is configured to transmit the EEG data to be processed through the mapping process to the brain activity recognition model for classification processing, and obtain the brain activity type corresponding to the EEG data to be processed.

Further, the acquisition module 10 includes:

A preprocessing unit, configured to preprocess multiple sets of EEG data according to data attributes of multiple sets of EEG data;

The mapping unit is used to transmit the preprocessed multiple sets of EEG data to the local terminal, and according to the feature mapping models corresponding to the multiple sets of EEG data in the local terminal, to multiple sets of the EEG data. The electrical data are respectively subjected to mapping processing to generate the standard EEG data.

Further, the update module 30 also includes:

An acquisition unit, configured to acquire a processing result generated by processing the standard EEG data obtained through remapping processing with the updated brain activity recognition model, and determine the updated brain activity recognition model according to the processing result The corresponding classification loss value;

The first determination unit is configured to determine the dimensionality reduction loss value corresponding to each of the feature mapping models according to the standard EEG data processed by remapping;

The second determining unit is configured to determine a total loss value of the brain activity recognition model after updating according to the classification loss value and the dimensionality reduction loss value.

Further, the acquisition unit is also used for:

reading the reference result of the standard EEG data, and performing similarity calculation between the reference result and the processing result to generate a first similarity value;

A classification loss value of the updated brain activity recognition model is determined according to the first similarity value.

Further, the first determination unit is also used for:

According to the corresponding relationship between the remapped standard EEG data and multiple sets of the EEG data, the remapped standard EEG data is divided into multiple sets of standard sub-data;

performing two-two combinations on multiple sets of standard sub-data, and performing reproduction processing on each set of standard sub-data in each combination to generate two sets of data points in each combination;

Perform mean value processing on each group of data points in each described combination, generate the average value of each group of data points in each described combination, and generate the average value of each described combination according to the average value of each group of data points in each described combination second similarity value;

A combined dimensionality reduction loss value of each combination is determined according to the second similarity value of each combination, and a dimensionality reduction loss value corresponding to each feature mapping model is generated according to each combination dimensionality reduction loss value.

Further, the update module 30 also includes:

a judging unit, configured to judge whether the total loss value is less than a preset value within a preset number of times; The loss value is less than the preset value;

The execution unit is configured to execute the step of receiving the update parameters and updating the brain activity recognition model if the values are not less than the preset value within the preset times.

The specific implementation of the federated migration-based EEG data processing device of the present invention is basically the same as the above embodiments of the federated migration-based EEG data processing method, and will not be repeated here.

In addition, the embodiment of the present invention also proposes a medium.

A federated migration-based EEG data processing program is stored on the medium, and when the federated migration-based EEG data processing program is executed by the processor, the steps of the federated migration-based EEG data processing method described above are implemented.

The medium of the present invention may be a computer-readable storage medium, and its specific implementation manners are basically the same as the embodiments of the federated migration-based EEG data processing method described above, which will not be repeated here.

Embodiments of the present invention have been described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementations, and the above-mentioned specific implementations are only illustrative, rather than restrictive, and those of ordinary skill in the art will Under the enlightenment of the present invention, without departing from the gist of the present invention and the scope of protection of the claims, many forms can also be made, and any equivalent structure or equivalent process transformation made by using the description and drawings of the present invention, or Directly or indirectly used in other related technical fields, these all belong to the protection of the present invention.

Claims (9)

1. a kind of EEG data processing method based on federation migration, it is characterized in that, described EEG data processing method based on federation migration comprises the following steps: Obtaining multiple sets of EEG data, and transmitting the multiple sets of EEG data to the local terminal respectively, and performing mapping processing on feature mapping models respectively corresponding to the multiple sets of EEG data, to generate standard EEG data; The standard EEG data is transmitted to the brain activity recognition model in the local terminal for processing, and the generated model parameters are transmitted to the preset public server, so that the preset public server can compare the model parameters and the information sent by at least one other terminal. Other model parameters are aggregated to generate update parameters; receiving the update parameters, updating the brain activity recognition model, and processing the standard EEG data obtained by remapping each feature mapping model based on the updated brain activity recognition model until the brain activity recognition model The generated total loss value is less than a preset value, and the processing of multiple sets of EEG data based on federated migration is completed; Wherein, after the step of processing the standard EEG data obtained by remapping each feature mapping model based on the updated brain activity recognition model, the method includes: Obtain the processing result generated by processing the standard EEG data obtained by the remapping process with the updated brain activity recognition model, and determine the classification loss value corresponding to the updated brain activity recognition model according to the processing result ; According to the standard EEG data processed by remapping, determine the dimensionality reduction loss value corresponding to each of the feature mapping models; According to the classification loss value and the dimensionality reduction loss value, determine the total loss value of the brain activity recognition model after updating, wherein the total loss value includes the classification of each EEG data by the brain activity recognition model Processing the classification loss value between the obtained predicted classification result and the actual classification result of each EEG data, and the dimensionality reduction loss value caused by the dimensionality reduction mapping of each EEG data by the feature mapping model. 2. the EEG data processing method based on federation migration as claimed in claim 1, is characterized in that, after described finishing is to the processing step of multiple groups of described EEG data based on federation migration, described method also comprises: When the EEG data to be processed is received, the feature mapping model is called to perform mapping processing; The EEG data to be processed by mapping is transmitted to the brain activity recognition model for classification processing, and the brain activity type corresponding to the EEG data to be processed is obtained. 3. the EEG data processing method based on federation migration as claimed in claim 1, is characterized in that, described multiple groups of described EEG data are respectively transmitted in the local terminal, corresponding to multiple groups of described EEG data respectively The feature mapping model is used for mapping processing, and the steps of generating standard EEG data include: Preprocessing the multiple sets of EEG data according to the data attributes of the multiple sets of EEG data; Transmitting the preprocessed multiple sets of EEG data to the local terminal respectively, and mapping the multiple sets of EEG data respectively according to the feature mapping models corresponding to the multiple sets of EEG data in the local terminal processing to generate the standard EEG data. 4. The EEG data processing method based on federation migration as claimed in claim 1, wherein, according to the processing result, the step of determining the classification loss value corresponding to the updated brain activity recognition model comprises: reading the reference result of the standard EEG data, and performing similarity calculation between the reference result and the processing result to generate a first similarity value; A classification loss value of the updated brain activity recognition model is determined according to the first similarity value. 5. the EEG data processing method based on federation migration as claimed in claim 1, is characterized in that, described according to the standard EEG data of remapping process, determine the dimensionality reduction loss value corresponding to each described feature mapping model Steps include: According to the corresponding relationship between the remapped standard EEG data and multiple sets of the EEG data, the remapped standard EEG data is divided into multiple sets of standard sub-data; performing two-two combinations on multiple sets of standard sub-data, and performing reproduction processing on each set of standard sub-data in each combination to generate two sets of data points in each combination; Perform mean value processing on each group of data points in each described combination, generate the average value of each group of data points in each described combination, and generate the average value of each described combination according to the average value of each group of data points in each described combination second similarity value; A combined dimensionality reduction loss value of each combination is determined according to the second similarity value of each combination, and a dimensionality reduction loss value corresponding to each feature mapping model is generated according to each combination dimensionality reduction loss value. 6. The EEG data processing method based on federation migration as claimed in claim 1, wherein, according to the classification loss value and the dimensionality reduction loss value, determine the updated brain activity recognition model After the step of total loss value, the method also includes: judging whether the total loss value is less than a preset value within a preset number of times, and if it is less than a preset value within a preset number of times, it is determined that the total loss value generated by the brain activity recognition model after updating is less than a preset value value; If it is not less than the preset value within the preset times, the step of receiving the update parameters and updating the brain activity recognition model is performed. 7. An EEG data processing device based on federation migration, characterized in that, the EEG data processing device based on federation migration comprises: The acquiring module is used to acquire multiple sets of EEG data, and transmit the multiple sets of EEG data to the local terminal respectively, perform mapping processing on feature mapping models corresponding to the multiple sets of EEG data, and generate standard EEG data data; A generating module, configured to transmit the standard EEG data to the brain activity recognition model in the local terminal for processing, and transmit the generated model parameters to a preset public server, so that the preset public server can compare the model parameters and at least Aggregate other model parameters sent by another terminal to generate updated parameters; An update module, configured to receive the update parameters, update the brain activity recognition model, and process the standard EEG data obtained by remapping each feature mapping model based on the updated brain activity recognition model until the The total loss value generated by the brain activity recognition model is less than a preset value, and the processing of multiple sets of EEG data based on federated migration is completed; Wherein, the EEG data processing device based on federation migration also includes: Obtain the processing result generated by processing the standard EEG data obtained by the remapping process with the updated brain activity recognition model, and determine the classification loss value corresponding to the updated brain activity recognition model according to the processing result ; According to the standard EEG data processed by remapping, determine the dimensionality reduction loss value corresponding to each of the feature mapping models; According to the classification loss value and the dimensionality reduction loss value, determine the total loss value of the brain activity recognition model after updating, wherein the total loss value includes the classification of each EEG data by the brain activity recognition model Processing the classification loss value between the obtained predicted classification result and the actual classification result of each EEG data, and the dimensionality reduction loss value caused by the dimensionality reduction mapping of each EEG data by the feature mapping model. 8. An EEG data processing device based on federation migration, characterized in that, the EEG data processing device based on federation migration includes a memory, a processor, and is stored on the memory and can run on the processor The EEG data processing program based on federation migration, when the EEG data processing program based on federation migration is executed by the processor, realize the EEG data based on federation migration as described in any one of claims 1-6 The steps of the processing method. 9. A medium, characterized in that the medium is stored with an EEG data processing program based on federated migration, and when the federated migration-based EEG data processing program is executed by a processor, it is implemented as in claims 1-6 The steps of any one of the federated migration-based EEG data processing methods.

Images