A method for EEG signal identity recognition based on personalized federated incremental learning

Abstract

The present invention provides an EEG signal identity recognition method based on personalized federated incremental learning, comprising the following steps: first, a server-client based federated learning architecture is adopted, and each client initializes a private local model, a learning rate network, and a playback sample pool. Then, at each task increment, each client uses an incremental meta-learning method to continuously learn the local continuous task stream. Next, the present invention designs a personalized federated incremental learning strategy to achieve learning of heterogeneous data of different incremental tasks in different clients by globally sharing model parameters and locally retaining the learning rate network. Finally, the central server uses the global model trained during the communication process to perform identity recognition on the EEG signal data of different institutions. The personalized federated incremental learning method proposed in the present invention realizes the forward transfer of knowledge between clients during the federated communication process to generate a universal identity recognition template applicable to each end.

Description

Technical Field

[0001] The present invention relates to the field of federated incremental learning for electroencephalogram (EEG) signal identification. Specifically, the present invention relates to an EEG signal identification method for privacy protection and continuous learning in distributed task increment scenarios based on personalized federated incremental learning. Background Art

[0002] With the surging demand for highly reliable identity authentication systems and the rapid development of artificial intelligence (AI), biometric recognition (including facial recognition, speaker recognition, iris recognition, and EEG signal recognition) has made significant progress and widespread application in recent years. As people continue to generate data, they are also increasingly concerned about privacy and security. Especially when identity authentication systems involve personal private data, people demand even greater privacy and security. EEG signals, due to their ubiquity, permanence, collectability, and uniqueness, have become a highly secure biometric authentication feature. However, they still have certain limitations. On the one hand, the emergence of deep neural networks has enabled rapid and accurate recognition of complex biometric encodings such as EEG, achieving higher recognition accuracy and better generalization on large-scale datasets. However, with the widespread adoption of hardware devices and the increase in online data streams, traditional offline training of deep neural network models has become ineffective. For example, in real-world identity authentication systems, new users register their EEG signatures daily. Offline-trained models are unable to adapt to newly registered identity data, posing challenges for efficient and reliable identity authentication systems. Traditional training methods use the entire dataset for centralized training, resulting in offline models with the best recognition performance and generalization. However, as data volumes grow, this approach places significant pressure on storage and computational memory and is extremely time-consuming. Furthermore, due to privacy concerns or storage limitations, data from previous tasks is often not fully accessible. In this case, traditional neural networks are trained only on samples from new tasks, leading to a bias in the model toward new data and a decline in recognition performance on older data, a phenomenon known as catastrophic forgetting. Therefore, facing a continuously increasing stream of biological data, reliable authentication models must demonstrate continuous learning capabilities: the ability to learn consecutive tasks without forgetting previously trained tasks. On the other hand, real-world data comes from distributed edge devices. Centralizing training on large datasets can improve model accuracy and robustness, but this places high demands on the storage and computational capabilities of cloud devices, potentially introducing privacy risks and data leakage threats to edge devices. Training each distributed entity on its own data can protect the privacy of each edge device, but it significantly reduces the accuracy and generalization of each distributed model.

[0003] In real-world scenarios, both data increment and privacy issues need to be considered. Task data from each distributed authentication device arrives continuously, which undoubtedly presents challenges for EEG identity recognition. On the one hand, with each new task stream, each client needs to retrain a new model based on all training samples to accommodate all tasks, which results in memory usage and time waste. Especially when each edge device is lightweight, the continuous task stream increases the storage pressure on each device. On the other hand, when the data between clients is not independent and identically distributed, if training is performed only based on the most recent task data, the model gradient will be biased towards the new task, leading to catastrophic forgetting of past tasks. In summary, designing an EEG identity recognition method suitable for distributed task increment scenarios has guiding significance for biometric authentication systems in real-world scenarios. Summary of the Invention

[0004] To solve the above problems, the present invention provides an EEG signal identity recognition method based on personalized federated incremental learning, constructs a personalized federated incremental learning framework based on privacy protection, realizes personalized incremental learning by globally sharing meta-parameters and locally retaining learning rates, and each client adopts an adaptive incremental meta-learning method based on sample playback and task distillation for local updates, thereby avoiding forgetting old tasks while achieving adaptive learning of different tasks.

[0005] The process of using the EEG signal identity recognition method based on personalized federated incremental learning provided by the present invention includes the following steps:

[0006] The method for identifying EEG signals based on personalized federated incremental learning according to claim 1 is characterized in that, in step S101, a server-client based federated learning architecture is adopted to evenly distribute the motor imagery EEG signals collected by each incremental task to all K clients (federated communication involves a total of K clients), and each client k (1≤k≤K) initializes a private local model Learning Rate Network and playback sample pool

[0007] First, the server-client architecture used is a federated learning framework with the server as the center and the clients as distributed nodes. Each client trains only based on local private data samples and does not share data with other clients or servers, but does share model parameters with the server.

[0008] Then, according to different data distribution scenarios, the collected motor imagery EEG signals are evenly distributed to all clients. For EEG signal data with M categories and n samples in each category, under the IID (independent and identically distributed) setting, the training samples of all classes are evenly distributed on all clients, and each client will process data with the same category but non-overlapping samples: each client k processes M categories and n samples. In the non-IID (non-independent and identically distributed) setting, all categories are evenly distributed across all clients, and the categories processed by each client are disjoint: the number of sample categories processed by each client k is The number of samples is

[0009] Finally, each client k is initialized, including the model parameters of the local meta-learning model Model parameters of the local learning rate network and a playback sample pool of a certain capacity

[0010] The method for EEG signal identification based on personalized federated incremental learning according to claim 1 is characterized in that in step S102: for the k-th client (1≤k≤K, the federated communication involves a total of K clients) described in step S101, its private task flow is set When the cth task communicates (1≤c≤C, there are C tasks communicating in the federated process), enter the current incremental task Sample data in, based on local models and the current incremental task Conduct incremental learning;

[0011] First, p and i represent the number of outer and inner loops respectively. represents the model parameters obtained by performing i internal trainings during the p-th external loop on the k-th client;

[0012] Then, in the inner loop of incremental meta-learning, meta-learning is combined with the replay of the most recent class mean sample; based on the current incremental task The data is used to perform meta-learning on the multi-task small sample scenario using a learnable learning rate network. During the c-th task communication (2≤c≤C), the local playback sample pool of each client k is Perform sample sampling to obtain the sampled old task samples Based on old task samples Calculate distillation loss; for the current incremental task Perform sample sampling to obtain a new task sample b={(X c ,Y c),2≤c≤C}, based on the new task sample b and the old task sample Calculate the classification loss; the goal of meta-learning internal training is

[0013]

[0014] Among them, X 1:c and Y 1:c They represent the training data and corresponding labels processed by the client during the first c task communications, X c and Y c They represent the training data and corresponding labels processed by the client during the c-th task communication, λ represents the relative weight of classification loss and distillation loss, and l meta represents the loss function of the meta-task, Indicates that the input is X 1:c , the model parameters are The output prediction result, l CE and l KD Represent the classification loss function and distillation loss function respectively; at the same time, the replay sample pool is updated according to the recent class mean sample replay rule;

[0015] Finally, in the outer loop of incremental meta-learning, the parameters of the learning rate network and meta-model are updated by gradient descent according to the meta-loss to find the learning rate and direction suitable for different tasks; for the next outer loop p+1 of k clients, the meta-parameters Gradient update is performed by the meta-loss based on classification loss and distillation loss:

[0016]

[0017] Where β represents the learning rate parameter for updating the meta-model parameters; for the next external cycle p+1 of the k-th client, the meta-loss is based on the learning rate network of the p-th external cycle Calculate the gradient and adaptively update the learning rate network:

[0018]

[0019] Among them, α hyperlr represents the learning rate parameter for network updates, is a learnable learning rate network for the k-th client p-th outer loop, the architecture and meta-model of this network The architecture is the same.

[0020] The method for EEG signal identification based on personalized federated incremental learning according to claim 1 is characterized in that, in step S103: the personalized federated incremental learning framework is composed of multiple clients described in step S102 and a central server. After each client completes the local update of the current task, it locally retains the learning rate network and the playback sample pool, and globally shares the parameters of the metamodel, thereby achieving communication with the server and retaining the local personalized learning direction;

[0021] First, the central server aggregates the local metamodel parameters from all clients The amount of data processed by each client in the current round is d1,…,d k ,…,d K Then, the aggregated model parameters are federated and weighted averaged according to the amount of data to obtain the global model parameters Finally, the central server distributes the global model parameters Θ to all clients.

[0022] The method for EEG signal identity recognition based on personalized federated incremental learning according to claim 1 is characterized in that in step S104: the personalized federated incremental learning method designed in S103 is iteratively trained until R communication rounds to ensure model convergence; the personalized federated incremental learning method is used to perform distributed incremental learning on the EEG data sample to be identified to determine the user label corresponding to the EEG data sample, and the specific process is that the central server identifies the input test EEG signal x based on the trained global parameter Θ, and the obtained label y is the predicted identity information.

[0023] The beneficial effects of the present invention are: the EEG signal identification method based on personalized federated incremental learning of the present invention constructs a federated learning framework based on a server-client architecture, in which each client adopts an incremental meta-learning method to perform local incremental learning; then, a personalized federated communication strategy with globally shared model parameters and locally retained learning rates is designed, and each client is set to communicate only its meta-learning parameters with the server instead of communicating a private learning rate network. On the one hand, it can strengthen the protection of the local model learning method, and on the other hand, it can realize personalized learning of different optimizer learning rates of different clients. At the same time, the local privatization of the learning rate network also reduces the communication cost of transmission. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a structural diagram of an EEG signal identity recognition method based on personalized federated incremental learning in an embodiment of the present invention.

[0025] Figure 2 2 is a schematic diagram of the architecture of an incremental learning method based on adaptive meta-learning in a client according to an embodiment of the present invention.

[0026] Figure 3 2 is a schematic diagram of calculating the distillation loss and classification loss in the meta-task of an embodiment of the present invention.

[0027] Figure 4 Schematic diagram of the update process of the inner loop and outer loop of the incremental meta-learning method according to an embodiment of the present invention.

[0028] Figure 5 This is a diagram of the overall framework of EEG signal identity recognition using personalized federated incremental learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments: The method of the present invention is divided into four parts.

[0030] Part 1: Architecture Construction of Federated Learning

[0031] Part 2: Client performs local updates based on incremental meta-learning

[0032] Part III: Personalized Communication between Server and Client

[0033] Part 4: Server performs EEG signal identification

[0034] According to these four parts, the embodiment of the present invention is based on the EEG signal identity recognition method of personalized federated incremental learning, such as Figure 1 As shown, the following steps are included:

[0035] S101: Using a server-client based federated learning architecture, the motor imagery EEG signals collected by each incremental task are evenly distributed to all K clients (federated communication involves a total of K clients), and each client k (1≤k≤K) initializes a private local model Learning Rate Network and playback sample pool

[0036] First, the server-client architecture adopted is a federated learning framework with the server as the center and the client as the distributed node. Each client is trained only based on local private data samples, and does not share data with other clients or servers, but shares model parameters with the server.

[0037] Then, according to different data distribution scenarios, the collected motor imagery EEG signals are evenly distributed to all clients. For EEG signal data with M categories and n samples in each category, under the IID (independent and identically distributed) setting, the training samples of all classes are evenly distributed on all clients, and each client will process data with the same category but non-overlapping samples: each client k processes M categories and n samples. In the non-IID setting, all categories are evenly distributed across all clients, and the categories processed by each client are disjoint: the number of samples processed by each client k is The number of sample categories is

[0038] Finally, each client k is initialized, including the model parameters of the local meta-learning model Model parameters of the local learning rate network and a playback sample pool of a certain capacity

[0039] S102: At each task increment, each client initialized in step S101 performs incremental learning based on the local model and the current private incremental task. This is mainly achieved by the following three steps: 1) combining the inner loop of meta-learning with the playback of the most recent class mean sample; 2) meta-task distillation; 3) meta-learning outer loop based on a learnable learning rate network;

[0040] Figure 2 The figure shows the architecture diagram of the incremental learning method based on adaptive meta-learning local to each client.

[0041] The first step is to combine the inner loop of meta-learning with the replay of the nearest class mean sample. For the k-th client (1≤k≤K, federated communication involves a total of K clients), the private task flow When the cth task communicates (1≤c≤C, there are C tasks communicating in the federated process), the private task processed by client k is During the c-th task communication (1≤c≤C), each client k (1≤k≤K) based on its private task The samples are used to train the local incremental model, where the task Sampled from private dataset And includes a total categories. Assuming that the sample space has N category labels, then Private datasets middle represents sample data, It should be noted that for client k, the task sample categories processed by different task communication rounds do not overlap, that is, when i≠j For different clients p and q, the task sample categories processed by the same task communication round may overlap, that is, when p≠q

[0042] The incremental meta-learning training process can be divided into internal training and external training. For time c, this method is based on the current new task The new task sample b is sampled for internal training, based on b and from the sample pool Sample of old tasks Therefore, this method can ensure the alignment of gradients of new and old tasks while learning the new data distribution.

[0043] First, in the inner loop of incremental meta-learning, based on the data of the current new task, a learnable learning rate network is used to quickly adapt to the multi-task small sample scenario. p and i represent the number of outer and inner loops, respectively. Represents the model parameters obtained by performing i internal trainings in the p-th external loop on the k-th client.

[0044] Then, let b m =m c-1 ∪b is the number of samples from the sample pool Sample of old tasks Mixed data of new task sample b sampled by the current task, based on the model parameters obtained by internal training and mixed data b m Perform external meta-training to obtain local meta-parameters Θ c Therefore, at time c, the objective of formula (1) can be rewritten as:

[0045]

[0046] Among them, {X 1:c ,Y 1:c} is from b m Sample data of 1 to c tasks processed so far obtained by sampling.

[0047] The recent class mean sample replay method uses a fixed-capacity replay sample pool, which stores the sample data of the recent class mean in the old task for retrieval during the new task training process. It is the playback sample pool obtained by sampling the nearest class mean in the cth round of communication, and is updated to The recent class mean sample replay method includes two main stages: sample sampling and sample pool update. The sample sampling stage constructs a limited size replay sample m (the process is shown in Algorithm 1); the sample pool update stage rebuilds the fixed capacity replay sample pool at time c. (The process is shown in Algorithm 2).

[0048]

[0049]

[0050] The second step is meta-task distillation: As mentioned above, the b of meta-task sampling m is composed of the current task batch b and the sample pool Sample playback samples The present invention does not need to store the network parameters Θ corresponding to each task. c , but the prototype of the nearest class mean sample in the replay sample pool (the feature embedding extracted before the fully connected layer of the corresponding task) is used as the soft label. For each old task τ, the soft label in the replay sample pool is the classifier Θ when the task τ was just trained. τ Generated to ensure Θ τ The sample distribution of task τ can be learned most accurately.

[0051] like Figure 3 As shown in the figure, this is the method for calculating the distillation loss of the old sample and the classification loss of the new and old samples in the meta-task. When calculating the distillation loss, let the playback sample The number of class labels is |m|. Sample data in (where label Let the output of the previous old classifier and the output of the current new classifier be and O |m| (x)=[o 1 (x),…,o |m| (x)]. The distillation loss can then be expressed as:

[0052]

[0053] in T represents the temperature scale. When calculating the classification loss, let the number of category labels of the new task sample be |n|. Mixed data of |m| replay samples in and |n| samples in current task b (where label Let the output of the current new classifier be O |m|+|n| (x)=[o 1 (x),…,o |m|(x),o |m|+1 (x),…,o |m|+|n| (x)]. The cross entropy classification loss can then be expressed as:

[0054]

[0055] in, If the true category y of the sample is k, it takes 1, otherwise it takes 0, p k (x) represents the probability of the classifier outputting the kth category (such as the logits value of the softmax layer).

[0056] Therefore, formula (4) can be rewritten as:

[0057]

[0058] Among them, l CE is in b m The cross entropy loss function for correct classification, l KD is The distillation loss function used for network regularization is shown above. The meta-parameters of the next outer loop p+1 are updated by the meta-loss based on the classification loss and the distillation loss:

[0059]

[0060] Where β represents the learning rate parameter for updating the meta-model parameters.

[0061] The third step is the meta-learning outer loop based on the learnable learning rate network: at the p+1th outer loop of the kth client, the meta-loss is based on the learning rate network of the pth outer loop. Calculate the gradient and adaptively update the learning rate network:

[0062]

[0063] in, is a learnable learning rate network for the p-th outer loop, the architecture and meta-model of the network The architecture is the same.

[0064] Figure 4 Described and The update process, and All are updated by gradient descent, where c represents the number of tasks being processed, i and p represent the number of inner and outer loops, respectively. The goal is to adaptively adjust the learning rate and direction of model updates while reducing dependence on learning rate initialization.

[0065]

[0066]

[0067] S103: The client updated in step S102 retains the local rate network and playback sample pool, and shares the local meta-learning model parameters with the server; the central server obtains a global model by aggregating the shared model parameters of the clients updated in step S102; the central server distributes the global model obtained by federated averaging to all clients;

[0068] like Figure 5 As shown, each client k maintains its own private playback sample pool locally and the learnable learning rate network parameter Communicate its meta-learned model parameters Θ to the central server k The central server aggregates the meta-learning model parameters uploaded by each participating client and distributes them to each client as global model parameters. Unlike traditional federated systems that share the optimizer's learning rate during communication or reinitialize the optimizer's learning rate during each round of communication, in the framework constructed in this chapter, each client k only initializes the learning rate network during the first task flow. During subsequent federated communication, task flows are maintained and trained as they arrive. This approach enables local, personalized learning. This approach allows clients to individually learn their local data distribution and achieves forward knowledge transfer between clients during federated communication, while also reducing communication costs. Algorithm 4 describes the specific process of this framework. The personalized framework designed enables global sharing of meta-learning model parameters between clients, while each client locally maintains adaptive learning rate network parameters.

[0069]

[0070]

[0071] S104: The server performs identity recognition of the EEG signal based on the global model after federated averaging obtained in step S103; uses the personalized federated incremental learning method to perform distributed incremental learning on the EEG data sample to be recognized, and determines the user label corresponding to the EEG data sample. The specific process is that the central server recognizes the input test EEG signal x based on the trained global parameter Θ, and the obtained label y is the predicted identity information.

[0072] Experimental design

[0073] The training model of the present invention was tested on the large-scale standard EEG Motor Movement / Imagery Dataset. In this example, the dataset contains over 1,500 1-2 minute EEG signal recordings from 109 healthy subjects, sampled at 160 Hz. Each subject performed a different movement / imagery task, with 64-channel EEG signals recorded using the BCI2000 system. Each subject underwent 14 experiments: two 1-minute baseline exercises (the first with eyes open, the second with eyes closed), and three 2-minute exercises in each of the following four tasks:

[0074] Task 1: A target appears on the left or right side of the screen. The subject opens and closes the corresponding fist until the target disappears. The subject then relaxes.

[0075] Task 2: A target appears on the left or right side of the screen. The subject imagines opening and closing the corresponding fist until the target disappears. The subject then relaxes.

[0076] Task 3: A target appears at the top or bottom of the screen. The subject opens or closes their fists (if the target is at the top) or their feet (if the target is at the bottom) until the target disappears. The subject then relaxes.

[0077] Task 4: A target appears at the top or bottom of the screen. The subject imagines opening or closing their fists (if the target is at the top) or their feet (if the target is at the bottom) until the target disappears. The subject then relaxes.

[0078] The eyes open are referred to as EO (Eye Open), the eyes closed are referred to as EC (Eye Close), the movement state is referred to as PHY (Physical), and the imagined movement state is referred to as IMA (Image).

[0079] This paper uses inter-task datasets to train the identity recognition model: EO and EC resting state data are used for training and testing. In the federated incremental scenario, the number of clients is set to 5 and 10, the number of communication rounds is set to 1, 2, 5, 10, and 20, and the local playback sample pool size for each client is 109. Experiments are designed under both IID and non-IID settings:

[0080] IID setting, that is, independent and identically distributed setting. The training samples of all classes are evenly distributed on all clients, and each client will process the same category. Specifically, each client will process the task sequence with the same category label as other clients in a different order, and different clients process different category samples. In this setting, the dataset is divided into 11 tasks consisting of 10 non-overlapping incremental categories (the last task contains 9 classes). When each task is incremented, the samples of all incremental categories of the task are evenly distributed to all clients;

[0081] Non-IID setting: This is a non-IID setting. All classes are evenly distributed across all clients, and each client processes disjoint classes. Specifically, each client processes a sequence of tasks containing different class labels, and different clients process different classes. In this setting, when there are 5 clients in the federated framework, the 109 classes are divided into 9 tasks with 5 increments and 16 tasks with 4 increments; when there are 10 clients, the 109 classes are divided into 11 tasks with 3 increments and 19 tasks with 4 increments.

[0082] Experimental results

[0083] Tables 1 and 2 show the global server performance and local client performance of the present invention under the IID setting and non-IID setting, respectively, with 5 clients, 10 clients, and 1, 2, 5, 10, and 20 communication rounds. From the perspective of data distribution, the proposed method of the present invention is effective for federated incremental scenarios in both the IID setting and the non-IID setting; in particular, the performance in the non-IID setting is better with more communication rounds, indicating that when the data is not independent and identically distributed, the method proposed in this chapter can enable each client to learn more migration knowledge from other clients in more communication rounds. From the perspective of the number of clients and communication rounds, the present invention is effective under different parameter settings, and multi-scenario experiments can verify the generalization and robustness of the present invention.

[0084] Table 1. Federated incremental performance under different clients and different communication rounds in the IID setting. Evaluated by server-side global performance (%) and client-side local performance (%). Each experiment takes the average of three random seeds.

[0085]

[0086] Table 2: Non-IID setup: Federated incremental performance under different clients and different communication rounds. Evaluated by server-side global performance (%) and client-side local performance (%). Each experiment takes the average of three random seeds.

[0087]

[0088]

Claims

1. A method for EEG signal identification based on personalized federated incremental learning, characterized in that: The following steps are involved: S101: The server-client architecture adopted is a federated learning framework with the server as the center and the clients as distributed nodes. Each client is trained only based on local private data samples and does not share data with other clients or servers, but shares model parameters with the server. Then, according to different data distribution scenarios, the collected motor imagery EEG signals are evenly distributed to all K clients. For EEG signal data with a total of M categories and n samples in each category, under the independent and identically distributed setting, the training samples of all classes are evenly distributed on all clients. Each client k, where 1≤k≤K, will process data with the same category but non-overlapping samples, that is, the number of sample categories processed by each client k is M and the number of samples is n. In the non-IID setting, all categories are evenly distributed across all clients, and the categories processed by each client are disjoint: the number of sample categories processed by each client k is The number of samples is Finally, each client k is initialized, including the model parameters of the local meta-learning model Model parameters of the local learning rate network and a playback sample pool of a certain capacity S102: At each task increment, each client initialized in step S101 performs incremental learning based on the local model and the current private incremental task; S103: The client updated in step S102 retains the local learning rate network and playback sample pool, and shares the local meta-learning model parameters with the server; the central server obtains a global model by aggregating the shared model parameters of the clients updated in step S102; the central server distributes the global model obtained by federated averaging to all clients; S104: Iteratively train the personalized federated incremental learning method designed in S103 until R communication rounds to ensure model convergence; use the personalized federated incremental learning method to perform distributed incremental learning on the EEG data sample to be identified to determine the user label corresponding to the EEG data sample. The specific process is that the central server identifies the input test EEG signal x based on the trained global parameter Θ, and the obtained label y is the predicted identity information.

2. The method for EEG signal identification based on personalized federated incremental learning according to claim 1, characterized in that: In step S102: for client k in step S101, where 1≤k≤K, and the federated communication involves a total of K clients, set its private task flow When the cth task communicates, where 1≤c≤C, the federated process has a total of C tasks communicating, and the current incremental task is entered Sample data in, based on local models and the current incremental task Conduct incremental learning; First, p and i represent the number of outer and inner loops respectively. represents the model parameters obtained by performing i internal trainings in the p-th external loop of the k-th client; Then, in the inner loop of incremental meta-learning, meta-learning is combined with the replay of the most recent class mean sample; based on the current incremental task The data of the local playback sample pool of each client k is replayed using a learnable learning rate network to perform meta-learning on the multi-task small sample scenario. In the c-th task communication, where 2≤c≤C, Perform sample sampling to obtain the sampled old task samples Based on old task samples Calculate distillation loss; for the current incremental task Perform sample sampling to obtain a new task sample b={(X c ,Y c ),2≤c≤C}, based on the new task sample b and the old task sample Calculate the classification loss; the goal of meta-learning internal training is Among them, X 1:c and Y 1:c They represent the training data and corresponding labels processed by the client during the first c task communications, X c and Y c They represent the training data and corresponding labels processed by the client during the c-th task communication, λ represents the relative weight of classification loss and distillation loss, represents the loss function of the meta-task, Indicates that the input is X 1:c , the model parameters are The output prediction result of and Represent the classification loss function and distillation loss function respectively; at the same time, the replay sample pool is updated according to the recent class mean sample replay rule; Finally, in the outer loop of incremental meta-learning, the parameters of the learning rate network and meta-model are updated by gradient descent according to the meta-loss to find the learning rate and direction suitable for different tasks; for the next outer loop p+1 of k clients, the meta-parameters Gradient update is performed by the meta-loss based on classification loss and distillation loss: Where β represents the learning rate parameter for updating the meta-model parameters; for the next external cycle p+1 of the k-th client, the meta-loss is based on the learning rate network of the p-th external cycle Calculate the gradient and adaptively update the learning rate network: Among them, α hyperlr represents the learning rate parameter for network updates, is a learnable learning rate network for the k-th client p-th outer loop, the architecture and meta-model of this network The architecture is the same.

3. The method for EEG signal identification based on personalized federated incremental learning according to claim 1, characterized in that: In step S103: the personalized federated incremental learning framework consists of multiple clients described in step S102 and a central server. After each client completes the local update of the current task, it locally retains the learning rate network and the playback sample pool, and globally shares the parameters of the meta-model, thereby achieving communication with the server and retaining the local personalized learning direction; First, the central server aggregates the local metamodel parameters from all clients The amount of data processed by each client in the current round is d1,…,d k ,…,d K Then, the aggregated model parameters are federated and weighted averaged according to the amount of data to obtain the global model parameters Finally, the central server distributes the global model parameters Θ to all clients.

Images