Self-adaptive meta-federal learning method suitable for personnel action recognition in industrial scene

Abstract

The invention discloses a self-adaptive meta-federal learning method suitable for personnel action recognition in an industrial scene. The method comprises the following steps: acquiring action data of a to-be-recognized person; the to-be-recognized person action data is input to a self-adaptive meta-federated learning framework, an action recognition result is obtained, the self-adaptive meta-federated learning framework comprises a plurality of personalized recognition models, the personalized recognition models are formed through training of a total loss function, and the personalized recognition models are obtained through training of the total loss function; the personalized recognition model carries out local model training at a client by using local data and obtains local and local knowledge, obtains public accumulated knowledge from a personalized model of an adjacent client through knowledge distillation, and adaptively and dynamically fuses the public accumulated knowledge and the local and local knowledge; and finally, high performance and individuation of the model are realized for each client. According to the invention, the recognition performance of the personnel action recognition model can be improved without the participation of a central server.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial production site safety management, and in particular relates to an adaptive meta-federated learning method suitable for personnel action recognition in industrial scenarios. Background Art

[0002] In modern industrial environments, wearable sensor-based human motion recognition technology is becoming increasingly important. This technology is not affected by light, field of view, occlusion, etc., and can be applied to many aspects including safe production management, production efficiency improvement, and automation and robot collaboration. By monitoring the actions of operators in real time, potential unsafe behaviors can be discovered in time, major safety accidents can be prevented, and practical cases can be provided for safety training. At the same time, human motion recognition can also help optimize workflows, improve production efficiency, and ensure product quality. In recent years, with the development of artificial intelligence technology, especially the breakthroughs in deep learning in image recognition and behavior analysis, the performance of human motion recognition methods based on deep neural networks has been significantly improved. In order to train deep learning models with better performance, more high-quality data is inevitably needed. However, due to the data barriers between various industrial enterprises, scientific research institutes, and colleges and universities, the application process of deep learning models in the industrial field has been seriously affected.

[0003] Federated learning is an emerging distributed machine learning paradigm. Traditional federated learning effectively solves the problem of data privacy protection by allowing edge devices to perform model training locally and only upload parameter updates to the central server. However, in practical applications, the data distribution between different companies and factories may vary greatly, which makes model training extremely challenging and requires a lot of communication overhead for model transmission. Therefore, there is an urgent need for an adaptive meta-federated learning method suitable for human action recognition in industrial scenarios. Summary of the invention

[0004] To solve the above technical problems, the present invention proposes an adaptive meta-federated learning method suitable for personnel motion recognition in industrial scenarios, aiming to break down data barriers between enterprises while ensuring data privacy and security, and to improve the training quality and recognition performance of the personnel motion recognition model without the need for the participation of a central server.

[0005] The present invention provides an adaptive meta-federated learning method suitable for personnel action recognition in industrial scenarios, comprising:

[0006] Obtaining action data of the person to be identified;

[0007] The action data of the person to be identified is input into an adaptive meta-federated learning framework to obtain action recognition results, wherein the adaptive meta-federated learning framework includes several personalized recognition models, and the personalized recognition models are trained by a total loss function. The personalized recognition models use local data on the client to perform local model training and obtain local local knowledge. At the same time, the public accumulated knowledge is obtained from the personalized models of adjacent clients through knowledge distillation, and the public accumulated knowledge and local local knowledge are adaptively and dynamically integrated.

[0008] Optionally, local model training using local data in each client includes:

[0009] Get the local current data set;

[0010] The local current data set is input into the initial local model, and the local model is trained by a back-transfer algorithm using a cross-entropy loss function to obtain a local model that can reflect the local knowledge of the current client.

[0011] Optionally, the cross entropy loss function is:

[0012]

[0013] Among them, L e is the cross entropy loss function, is the probability that sample i belongs to the jth class, y ij is the true label of sample i belonging to the jth class, y ij =1 means it belongs to, y ij =0 means it does not belong, and k is the number of categories in the data set.

[0014] Optionally, obtaining the personalized model of the current client includes:

[0015] Input the local current data set into the student model and the teacher model, and obtain the accuracy of the teacher model and the student model, wherein the personalized model trained by the previous client is read as the teacher model, and the local model of the current client is used as the student model;

[0016] According to the accuracy, obtaining the optimal dynamic fusion ratio parameter of the current client;

[0017] The optimal dynamic fusion ratio parameter is used to adaptively and dynamically adjust the learning ratio of public accumulated knowledge and local partial knowledge to obtain a personalized model of the current client.

[0018] Optionally, the method for obtaining the optimal dynamic fusion ratio parameter of the current client is:

[0019]

[0020] Among them, λ i is the dynamic fusion ratio parameter of the current client, λ i-1 is the dynamic fusion ratio parameter of the previous client, d is the accuracy difference between the student model and the teacher model, β is the threshold, and α is the exponential smoothing coefficient.

[0021] Optionally, acquiring the public accumulated knowledge and local partial knowledge includes:

[0022] The cross entropy loss function is used to learn the local knowledge in the student model;

[0023] The common accumulated knowledge is learned from the teacher model using a boundary loss function based on the hidden layer neuron activations.

[0024] Optionally, the hidden layer neuron activation boundary loss function is:

[0025]

[0026] Among them, L d (I) is the hidden layer neuron activation boundary loss function, I is the input data, T is the function representation of the teacher model from the input layer to the hidden layer, S is the function representation of the student model from the input layer to the hidden layer, ⊙ is the element product of the vector, T(I)∈R M is the hidden layer neuron response vector of the teacher model when the input data is I, S(I)∈R M is the hidden layer neuron response vector of the student model, M is the number of neurons in the hidden layer, l is a vector with the same length as the number of neurons in the hidden layer, μ is the margin, σ(x)=max(0,x) is the ReLU activation function, and ρ(x) is the activation indicator function;

[0027]

[0028] Where x is the activation function input, ρ(x)=1 means that the neuron is activated, and ρ(x)=0 means that it is not activated.

[0029] Optionally, the total loss function is:

[0030] L i =(1-λ i )L e +λ i L d

[0031] Among them, L i is the total loss function of the personalized model of the i-th client, L e is the cross entropy loss function, L d is the activation margin loss of hidden layer neurons, λ iis the dynamic fusion ratio parameter.

[0032] Compared with the prior art, the present invention has the following advantages and technical effects:

[0033] 1. The adaptive meta-federated learning framework without a central server designed by the present invention realizes the effective migration of public accumulated knowledge between different clients through the knowledge distillation method based on the activation boundary of hidden layer neurons. Through this framework, the adaptive meta-federated learning framework performs personalized training on each client in turn in a cyclic manner, adopts an adaptive dynamic fusion strategy, and dynamically adjusts the learning ratio of public accumulated knowledge and local knowledge at different personalized model training stages, thereby training a personalized model for the current client, accumulating useful information from different clients without affecting privacy security, and effectively retaining local local knowledge, thereby enhancing the generalization performance of the model.

[0034] 2. The knowledge distillation method based on the activation boundary of the hidden layer neurons adopted in the present invention takes into account the activation boundaries of the model output layer and the intermediate layer at the same time, so as to capture more levels of feature representation, enabling the student model to learn richer and more subtle knowledge, thereby not only realizing knowledge transfer more comprehensively, but also effectively improving the generalization ability and training efficiency of the model and enhancing robustness.

[0035] 3. The adaptive dynamic fusion strategy designed by the present invention dynamically fuses public accumulated knowledge with local knowledge by effectively utilizing the calculated adaptive dynamic fusion ratio parameters, and obtains a personalized model for each client. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0037] Figure 1 is a flow chart of an adaptive meta-federated learning method for human action recognition in industrial scenarios according to an embodiment of the present invention;

[0038] Figure 2 It is a schematic diagram of the structure of the student model and the teacher model for each client in an embodiment of the present invention;

[0039] Figure 3 It is a schematic diagram of action types of an operator action data set in an industrial production scenario according to an embodiment of the present invention. DETAILED DESCRIPTION

[0040] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0041] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0042] The present invention provides an adaptive meta-federated learning method suitable for human action recognition in industrial scenarios, such as Figure 1 As shown, including:

[0043] Obtaining action data of the person to be identified;

[0044] The action data of the person to be identified is input into the adaptive meta-federated learning framework to obtain the action recognition result. The adaptive meta-federated learning framework includes several personalized recognition models. The personalized recognition model is trained through the total loss function. The personalized recognition model uses local data on the client to perform local model training and obtain local local knowledge. At the same time, it obtains public accumulated knowledge from the personalized models of adjacent clients through knowledge distillation, and adaptively and dynamically integrates public accumulated knowledge and local local knowledge.

[0045] Specifically, step S1: using multiple wearable sensors including accelerometers, gyroscopes, and magnetometers in each client to collect operator motion data as local private data sets; further dividing the collected local private data sets into a local training set, a validation set, and a test set according to a certain ratio;

[0046] Step S2: construct an adaptive meta-federated learning framework by uniting multiple clients, in which the personalized model training of each client includes two stages: local model learning and adaptive dynamic fusion;

[0047] Step S3: In the local model learning phase, a local model based on a deep neural network is initialized in each client. Then the local training set data is input into the local model, and the cross entropy loss function is used to train the local model through the back propagation algorithm. After the training is completed, a local model that can reflect the local knowledge of the current client is obtained;

[0048] Step S4: In the adaptive dynamic fusion stage, the personalized model trained by the previous client is read as the teacher model, the local model of the current client is used as the student model, the local verification set of the current client is input into the teacher model and the student model, and the optimal dynamic fusion ratio parameter suitable for the current client is calculated according to the accuracy of the teacher model and the student model on the local verification set;

[0049] The student model uses the cross entropy loss function to learn local knowledge, and uses the knowledge distillation method based on the hidden layer neuron activation boundary to learn the public accumulated knowledge from the teacher model. The learning ratio of public accumulated knowledge and local knowledge is dynamically adjusted through the dynamic fusion ratio parameter, and the back propagation algorithm is used to optimize and update the student model parameters, so as to train and obtain the personalized model of the current client.

[0050] Step S5: Save the current client personalized model and send it to the next client;

[0051] Step S6: Return to step S3 until the personalized model of each client converges, and each client obtains the final personalized model.

[0052] Step S7: Each client inputs the test data into the final personalized model to obtain the personnel action recognition results of each client.

[0053] Furthermore, using local data to perform local model training in each client includes:

[0054] Get the local current data set;

[0055] The local current data set is input into the initial local model, and the cross entropy loss function is used to train the local model through the back propagation algorithm to obtain a local model that can reflect the local knowledge of the current client.

[0056] Furthermore, the cross entropy loss function is:

[0057]

[0058] Among them, L e is the cross entropy loss function, is the probability that sample i belongs to the jth class, y ij is the true label of sample i belonging to the jth class, y ij =1 means it belongs to, y ij =0 means it does not belong, and k is the number of categories in the data set.

[0059] Furthermore, obtaining the personalized model of the current client includes:

[0060] Input the local current data set into the student model and the teacher model to obtain the accuracy of the teacher model and the student model. The personalized model trained by the previous client is read as the teacher model, and the local model of the current client is used as the student model.

[0061] According to the accuracy, the optimal dynamic fusion ratio parameter of the current client is obtained;

[0062] The optimal dynamic fusion ratio parameter is used to adaptively and dynamically adjust the learning ratio of public accumulated knowledge and local local knowledge to obtain a personalized model for the current client.

[0063] Furthermore, the method for obtaining the optimal dynamic fusion ratio parameter of the current client is:

[0064]

[0065] Among them, λ i is the dynamic fusion ratio parameter of the current client, λ i-1 is the dynamic fusion ratio parameter of the previous client, d is the accuracy difference between the student model and the teacher model, β is the threshold, and α is the exponential smoothing coefficient.

[0066] Furthermore, obtaining public accumulated knowledge and local partial knowledge includes:

[0067] The cross entropy loss function is used to learn the local knowledge in the student model;

[0068] The common accumulated knowledge is learned from the teacher model using a boundary loss function based on the hidden layer neuron activations.

[0069] Furthermore, the hidden layer neuron activation boundary loss function is:

[0070]

[0071] Among them, L d (I) is the hidden layer neuron activation boundary loss function, I is the input data, T is the function representation of the teacher model from the input layer to the hidden layer, S is the function representation of the student model from the input layer to the hidden layer, ⊙ is the element product of the vector, M is the number of neurons in the hidden layer, l is a vector with the same length as the number of neurons in the hidden layer, T(I)∈R M is the hidden layer neuron response vector of the teacher model when the input data is I, S(I)∈R M is the hidden layer neuron response vector of the student model, μ is the margin, σ(x)=max(0,x) is the ReLU activation function. ρ(x) is the activation indicator function, which can be expressed as:

[0072]

[0073] Where x is the activation function input, ρ(x)=1 means that the neuron is activated, and ρ(x)=0 means that it is not activated.

[0074] The i-th neuron response vector of the teacher model and the student model is denoted as t i and i , the above hidden layer neuron activation boundary loss function is for si The derivative of is:

[0075]

[0076] The back-propagation algorithm is used to update the student model parameters by minimizing the hidden layer neuron activation boundary function of the teacher model and the student model, thereby learning the common accumulated knowledge from the teacher model without the need for a central server and global model aggregation.

[0077] Furthermore, the total loss function of the current client personalization model is:

[0078] L i =(1-λ i )L e +λ i L d

[0079] Among them, L i is the total loss function of the personalized model of the i-th client, L e is the cross entropy loss function, L d is the activation margin loss of hidden layer neurons, λ i is the dynamic fusion ratio parameter.

[0080] In order to solve the problems of existing technologies that rely on servers, client data distribution differences, and the presence of many minority classes and missing classes. This embodiment proposes an adaptive meta-federation learning (AMFL) method suitable for personnel motion recognition in industrial scenarios. A manufacturing enterprise is equivalent to a client in an adaptive meta-federation learning framework. First, in each client, multiple wearable sensors are matched through a web page or a mobile terminal, and worn on the four limbs of the operator. The operator's motion data is collected in real time to form a local private data set for each client. After data preprocessing, it is input into the local model for training. After training, the local model parameters are updated, and then the personalized model trained by the current client is sent to the next client as the teacher model of the next client to participate in the personalized model training of the next client, and finally a personnel motion recognition model that better meets its own needs is trained for each client. The method includes:

[0081] Step S1: Set the number of clients to 20. A certain number of clients form a federation. Each federation is regarded as a meta-distribution. In the embodiment, an adaptive meta-federation learning framework is constructed by combining four federations. In this framework, the personalized model trained by the previous client will be used as the teacher model of the next client. The inheritance of public accumulated knowledge and model transmission are achieved between clients through knowledge distillation. The training of personalized models for each client includes two stages: local model learning and adaptive dynamic fusion, and finally a personalized model is implemented for each client;

[0082] In each client, a local model is initialized based on a two-dimensional convolutional neural network. The local model and the teacher model are as follows: Figure 2 As shown in the figure, they are composed of two convolutional layers, two batch normalization layers, two pooling layers and two fully connected layers. The collected personnel action data is input into the convolutional layer, and the local features of the input data are extracted through the convolutional layer; each small batch of data is normalized through the batch normalization layer to make the input distribution of each layer more stable; the spatial dimension of the feature map is further reduced through the pooling layer to reduce the amount of calculation; finally, the feature map extracted previously is converted into the final classification result through the fully connected layer.

[0083] Step S2: In the local model learning phase, the divided local training set data is input into the local model for training using the cross entropy loss function. The cross entropy loss function is used to measure the difference between the probability distribution predicted by the model and the true label. By minimizing the cross entropy loss, the model's learning process can be more inclined to generate predicted values ​​that are closer to the true label. The calculation formula is:

[0084]

[0085] Where, L e is the cross entropy loss function, is the probability that sample i belongs to the jth class, y ij is the true label of sample i belonging to the jth class, y ij =1 means it belongs to, y ij =0 means it does not belong, and k is the number of categories in the data set.

[0086] After completing the local model training, local knowledge that can reflect local characteristics is obtained, and the accuracy of human action recognition on the local verification set can be further obtained;

[0087] Step S3: In the adaptive dynamic fusion stage, the personalized model trained by the previous client is read as the teacher model, and the local model of the current client is used as the student model. The student model uses the cross entropy loss function to learn local knowledge, and uses the knowledge distillation method based on the hidden layer neuron activation boundary to learn the public accumulated knowledge from the teacher model. The model hidden layer neuron activation boundary loss function is used for the student model to learn the public accumulated knowledge from the teacher model. The hidden layer neuron activation boundary loss function is as follows:

[0088]

[0089] Where, L d (I) is the hidden layer neuron activation boundary loss function, I is the input data, T is the function representation of the teacher model from the input layer to the hidden layer, S is the function representation of the student model from the input layer to the hidden layer; σ(x) = max(0, x) is the ReLU activation function, the ⊙ symbol represents the element product of the vector, M is the number of hidden layer neurons, l is a vector with the same length as the number of hidden layer neurons, in which all component values ​​are equal to 1; T(I)∈R M is the hidden layer neuron response vector of the teacher model when the input data is I, S(I)∈R M is the hidden layer neuron response vector of the student model; in order to leave a margin for the neuron response of the student model to obtain stability, a margin μ is introduced, and ρ(x) is the activation indicator function, which can be expressed as:

[0090]

[0091] Where x is the activation function input, ρ(x) = 1 means the neuron is activated, and ρ(x) = 0 means it is not activated. The response vectors of the i-th neuron in the teacher model and the student model are t i and i , the above hidden layer neuron activation boundary loss function is for s i The derivative of is:

[0092]

[0093] The back-propagation algorithm is used to obtain the common accumulated knowledge of the teacher model by minimizing the activation boundary of the hidden layer neurons of the teacher model and the student model.

[0094] Input the current client local verification set into the student model and the teacher model respectively, and obtain the accuracy of the student model and the teacher model on the verification set respectively. and And calculate the difference:

[0095]

[0096] By comparing the difference d with the threshold β, an appropriate exponential smoothing formula is selected to determine the dynamic fusion ratio parameter λ i :

[0097]

[0098] In the formula, λ i is the dynamic fusion ratio parameter of the current client, λ i-1 is the dynamic fusion ratio parameter of the previous client, d is the accuracy difference between the student model and the teacher model; β is the threshold, and α is the exponential smoothing coefficient, which can be set manually according to the specific task and data set. i The total training loss function of the personalized model trained on the i-th client is as follows:

[0099] L i =(1-λ i )L e +λ i L d

[0100] Where, L e is the above cross entropy loss function, L d To activate the boundary loss. In the process of personalized model training, the total training loss is minimized, and the back propagation algorithm is used to update the student model neural network parameters. The training finally obtains the current client personalized model.

[0101] Step S4: Save the current client personalized model and send it to the next client;

[0102] Step S5: Return to step S2 until the personalized model of each client converges, and each client obtains the final personalized model.

[0103] Step S6: Each client inputs the test data into the final personalized model to obtain the personnel action recognition results of each client.

[0104] In order to prove the effectiveness and superiority of the method of this embodiment, this embodiment uses the following data sets for verification. Including: PAMAP2, UCI HAR, Opportunity and a self-built operator action data set (named WACID) containing 14 common industrial production scenarios. The WACID data set construction process fixes four six-axis wearable sensors on the limbs of the operator in each client. The IMU output data of each limb includes a three-axis acceleration signal and a three-axis gyroscope signal. The sampling rate of the sensor is 148Hz, and the total acquisition time is about half an hour. The specific wearing position of the sensor is as follows: Figure 2As shown in Figure 2, the dataset includes 14 common activity types of five operators. The collected actions are as follows: Figure 3 As shown, it includes sitting, going up stairs, going down stairs, leaning, climbing up the ladder, climbing down the ladder, pulling the isolation belt, plugging the blind plate, turning the valve to the left, turning the valve to the right, running, walking, turning on the high-voltage cabinet, and turning off the power. Then, a comparative experiment was conducted on these four datasets with the current advanced federated learning methods. The comparative methods are shown in Table 1, including three federated learning methods: FedAvg, FedProx, and FedBN.

[0105] Table 1

[0106]

[0107] Experiment execution strategy:

[0108] In this experiment, the local model of the technical solution uses a two-dimensional convolutional neural network as the basic skeleton, such as Figure 2 As shown in the figure, two convolutional layers, two batch normalization layers, two pooling layers, and two fully connected layers are used. The layers before the fully connected layers are regarded as feature extractors, and all fully connected layers are regarded as classifiers. The specific execution steps are as follows:

[0109] (1) The four data sets are divided into local training set, validation set and test set in a ratio of 4:3:3, and then the training data is passed into the local model for feature extraction. The feature vector is then input into the fully connected layer, and regression analysis is performed in the fully connected layer. The cross entropy loss function is used to calculate the loss between the predicted output and the actual label. The local model training uses the stochastic gradient descent optimization (SGD) algorithm to calculate the gradient of the model parameters based on the current loss. This is achieved through the back propagation algorithm, which starts from the output layer and calculates the gradient of the loss function with respect to the parameters of each layer layer by layer, and then uses the calculated gradient to update the model parameters. The update formula is as follows:

[0110]

[0111] In the formula, θ t+1 is the updated model parameter, θ t is the parameter at the current moment, η is the learning rate, L e is the cross entropy loss function, is the loss function with respect to the parameter θ t The gradient on the current sample. The local model learns to obtain local knowledge. The key hyperparameter settings of the local model are shown in Table 2:

[0112] Table 2

[0113]

[0114] The test data set is used to perform a performance test on the local final personalized model to obtain the recognition accuracy of the final personalized model of each client for all action types in the test data.

[0115] (2) Read the personalized model trained by the previous client as the teacher model, use the local model of the current client as the student model, input the local verification set of the current client into the teacher model and the student model, and calculate the optimal dynamic fusion ratio parameter suitable for the current client based on the accuracy of the teacher model and the student model on the local verification set;

[0116] (3) The student model uses the cross entropy loss function to learn local knowledge, and uses the knowledge distillation method based on the activation boundary of hidden layer neurons to learn the public accumulated knowledge from the teacher model; the learning ratio of public accumulated knowledge and local knowledge is dynamically adjusted through the dynamic fusion ratio parameter, and the back propagation algorithm is used to optimize and update the student model parameters;

[0117] (4) The training sets of the four data sets are respectively input into the FedAvg, FedProx and FedBN algorithms with the same parameter settings for model training, and the test sets of the four data sets are respectively used to evaluate the model performance, and the personnel action classification results of different algorithms are obtained. The average accuracy results of the method proposed in this embodiment and other methods are shown in Table 3. The average accuracy of the adaptive meta-federation learning (AMFL) method proposed in this embodiment on the four data sets is better than that of the other three methods.

[0118] Table 3

[0119]

[0120] The AMFL method adopted in this embodiment provides a federated learning framework that can transmit models without the participation of a central server, and can more effectively identify human movements. Through knowledge distillation, different clients can better learn and inherit public accumulated knowledge, which is crucial to the efficiency of model training and the construction of a meta-federal framework. Through an adaptive dynamic fusion strategy, the AMFL method can adaptively and dynamically fuse public accumulated knowledge and local knowledge, allowing the model to learn more robust feature representations in action recognition tasks, effectively improving the generalization ability and accuracy of the model. Through the constructed adaptive meta-federal framework, personalized training of local models of different clients can be achieved while ensuring data security and without the need for a central server, greatly reducing communication costs and improving model training efficiency.

[0121] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. An adaptive meta-federated learning method for human action recognition in industrial scenarios, characterized in that: include: Obtaining action data of the person to be identified; The action data of the person to be identified is input into an adaptive meta-federated learning framework to obtain action recognition results, wherein the adaptive meta-federated learning framework includes several personalized recognition models, and the personalized recognition models are trained by a total loss function. The personalized recognition models use local data on the client to perform local model training and obtain local local knowledge. At the same time, the public accumulated knowledge is obtained from the personalized models of adjacent clients through knowledge distillation, and the public accumulated knowledge and local local knowledge are adaptively and dynamically integrated.

2. According to claim 1, an adaptive meta-federated learning method suitable for human action recognition in industrial scenarios is characterized in that: Using local data to train a local model in each client includes: Get the local current data set; The local current data set is input into the initial local model, and the local model is trained by a back-transfer algorithm using a cross-entropy loss function to obtain a local model that can reflect the local knowledge of the current client.

3. The adaptive meta-federated learning method for human action recognition in industrial scenarios according to claim 2, characterized in that: The cross entropy loss function is: Among them, L e is the cross entropy loss function, is the probability that sample i belongs to the jth class, y ij is the true label of sample i belonging to the jth class, y ij =1 means it belongs to, y ij =0 means it does not belong, and k is the number of categories in the data set.

4. The adaptive meta-federated learning method for human action recognition in industrial scenarios according to claim 1, characterized in that: Obtaining the personalized model of the current client includes: Input the local current data set into the student model and the teacher model, and obtain the accuracy of the teacher model and the student model, wherein the personalized model trained by the previous client is read as the teacher model, and the local model of the current client is used as the student model; According to the accuracy, obtaining the optimal dynamic fusion ratio parameter of the current client; The optimal dynamic fusion ratio parameter is used to adaptively and dynamically adjust the learning ratio of public accumulated knowledge and local partial knowledge to obtain a personalized model of the current client.

5. The adaptive meta-federated learning method for human action recognition in industrial scenarios according to claim 4, characterized in that: The method to obtain the optimal dynamic fusion ratio parameter of the current client is: Among them, λ i is the dynamic fusion ratio parameter of the current client, λ i-1 is the dynamic fusion ratio parameter of the previous client, d is the accuracy difference between the student model and the teacher model, β is the threshold, and α is the exponential smoothing coefficient.

6. The adaptive meta-federated learning method for human action recognition in industrial scenarios according to claim 4, characterized in that: Acquiring the public accumulated knowledge and local partial knowledge includes: The cross entropy loss function is used to learn the local knowledge in the student model; The common accumulated knowledge is learned from the teacher model using a boundary loss function based on the hidden layer neuron activations.

7. The adaptive meta-federated learning method for human action recognition in industrial scenarios according to claim 6, characterized in that: The hidden layer neuron activation boundary loss function is: Among them, L d (I) is the hidden layer neuron activation boundary loss function, I is the input data, T is the function representation of the teacher model from the input layer to the hidden layer, S is the function representation of the student model from the input layer to the hidden layer, ⊙ is the element product of the vector, T(I)∈R M is the hidden layer neuron response vector of the teacher model when the input data is I, S(I)∈R M is the hidden layer neuron response vector of the student model, M is the number of neurons in the hidden layer, l is a vector with the same length as the number of neurons in the hidden layer, μ is the margin, σ(x)=max(0,x) is the ReLU activation function, and ρ(x) is the activation indicator function; Where x is the activation function input, ρ(x)=1 means that the neuron is activated, and ρ(x)=0 means that it is not activated.

8. The adaptive meta-federated learning method for human action recognition in industrial scenarios according to claim 1, characterized in that: The total loss function is: L i =(1-λ i )L e +λ i L d Among them, L i is the total loss function of the personalized model of the i-th client, L e is the cross entropy loss function, L d is the activation margin loss of hidden layer neurons, λ i is the dynamic fusion ratio parameter.

Images