CSI behavior identification method based on fairness adjustment federation

Abstract

The invention discloses a CSI behavior recognition method based on fairness adjustment federation, and relates to the technical field of wireless communication and machine learning, and the method comprises the steps: collecting a plurality of sample data generated by different user objects according to a fixed CSI behavior, building a federated network model, and obtaining each local model and a global model; obtaining model parameters corresponding to each local model and the global model according to the collected sample data in each source domain client; calculating a minimum empirical risk corresponding to each local model according to the model parameters to obtain a generalization gap of the model; setting the number of cycles to form a new global model; and distributing the new global model to each source domain client to form circulation, and obtaining a final trained federated network model. According to the method, federal learning and a behavior recognition system are combined, so that the privacy protection and data security requirements are met, the human body behavior recognition precision is improved, and the technical progress in the field related to human body behavior recognition is promoted.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless communications and machine learning, and in particular to a CSI behavior recognition method based on fairness adjustment federation. Background Art

[0002] With the rapid development of mobile internet and wireless communication technologies, intelligent human-computer interaction (HCI) technology has attracted widespread attention. Human behavior recognition plays a key role in multiple fields. Current HCI technologies include wearable sensor-based systems, computer vision-based systems, and wireless signal-based systems. Wearable sensors require users to carry them at all times, which is difficult to achieve in some situations, and the constant carrying of sensors can affect user comfort. Computer vision systems use cameras to capture user activity information, offering high recognition accuracy. However, cameras are expensive and susceptible to lighting conditions and non-line-of-sight obstacles, posing significant privacy risks. Wireless signal-based systems, particularly WiFi, have enormous potential for HCI applications due to their advantages of requiring no equipment, low cost, unrestricted lighting conditions, non-line-of-sight capabilities, and excellent privacy.

[0003] However, WiFi sensing systems face the challenge of high data requirements when training behavior recognition models. As the number of recognized actions increases, the amount of data required also increases. Data from a single client is often insufficient for model training, necessitating centralized training with data from multiple clients. However, centralized training can pose privacy risks and raise legal and ethical issues. Furthermore, the frequent occurrence of personal data leaks and the public's heightened awareness of privacy protection have made data collection more difficult, creating data silos and posing significant challenges to behavior recognition through centralized multi-client training. Summary of the Invention

[0004] The purpose of the present invention is to provide a CSI behavior identification method based on fairness adjustment federation to solve the problems raised in the prior art.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A CSI behavior recognition method based on fairness-adjusted federation includes the following steps:

[0007] Step S100: Collect sample data generated by various user objects according to fixed CSI behaviors. CSI behaviors are human behaviors, and each user object is treated as a source domain client. A federated network model is established, and local models trained in parallel based on the source domain clients in the federated network model are obtained, as well as a global model aggregated from the local models.

[0008] Step S200: Based on the collected sample data from each source domain client, a local model and a global model corresponding to each source domain client are trained using a deep learning network based on deconvolution, thereby obtaining model parameters corresponding to each local model and the global model;

[0009] Step S300: Calculate the minimum empirical risk corresponding to each local model based on the model parameters, and then obtain the generalization gap between the global model and each local model. Calculate the variance of the generalization gap between different local models and the global model, and then obtain the function formula for calculating the minimum empirical risk of the global model. Set the number of loops, and perform aggregation and fusion based on the model parameters and pre-set weights of each local model to form a new global model.

[0010] Step S400: Distribute the new global model to each source domain client, replacing the original local model of each source domain client, forming a cycle. Repeat the cycle process until the global model achieves satisfactory generalization performance and converges on all source domain clients, and obtains the final trained federated network model.

[0011] Furthermore, step S100 includes the following steps: obtaining T different indoor environments and M user objects with different physical characteristics, where the physical characteristics include height, weight, gender, and age, and setting a total of P CSI behaviors, and performing all CSI behaviors for each user object in each indoor environment; collecting data in which a certain user object m performs all CSI behaviors in each indoor environment and each CSI behavior is performed Q times as sample data of object m, and obtaining a total sample size of T×P×Q in the sample data corresponding to object m.

[0012] Furthermore, in step S200, the local model is trained based on the deconvolution deep learning network, including: processing the sample data of the user object through the deconvolution layer, converting the signal into a feature map with spatial encoding, and sending the generated feature map into a feature extractor with a residual block, and comprehensively processing the extracted features through the fully connected layer, thereby realizing the local training model of the deconvolution network.

[0013] Furthermore, calculating the minimum empirical risk corresponding to each local model according to the model parameters in step S300 includes the following steps:

[0014] Step S310: Transmit the model parameters corresponding to each local model and the global model to the server, and obtain the training set corresponding to the i-th local model as 1≤i≤M,N i represents the total number of samples corresponding to the i-th local model, represents the jth sample in the i-th local model, represents the label of the jth sample in the i-th local model. According to each sample in the training set, the minimum empirical risk corresponding to the i-th local model is obtained. The formula is: Among them, min is the minimum value, ε D represents the empirical risk, L is the loss function corresponding to the i-th local model, f is the local training model for implementing the deconvolution network, and θ is the model parameter of the i-th local model;

[0015] Add proximal term penalty regularization to the loss function L of the local model to iterate the global model parameters. Set the maximum number of iterations to r+1, and then adjust the minimum empirical risk formula corresponding to the i-th local model. The adjusted formula is: Among them, λ is the scale parameter that controls the penalty of proximal terms, θ r+1 represents the global model parameters at the r+1th iteration, θ r represents the global model parameters at the rth iteration, |||| 2 To find the square of the Euclidean distance; and then get the minimum empirical risk corresponding to all local models.

[0016] Furthermore, step S300 includes the following steps:

[0017] Step S320: The formula for calculating the generalization gap between the i-th local model and the global model is: in, is the generalization gap between the i-th local model and the global model, M is the number of local models, is the minimum empirical risk of the global model, is the minimum empirical risk of the i-th local model, a i Represents the preset weight of the i-th local model, 1≤i≤M, is the optimal solution of the local model of the i-th local model;

[0018] According to the generalization gap between the global model and each local model, the corresponding variance between all generalization gaps is calculated as: M is the number of local models, is the generalization gap corresponding to the i-th local model; and the objective function formula for calculating the minimum empirical risk of the global model is obtained as follows:

[0019] Where β is the variance coefficient, β∈(0,+∞).

[0020] Furthermore, step S300 includes the following steps:

[0021] Step S330: Based on the model parameter θ of the mth local model m, and the global model parameter θ, subtract the two to obtain the parameter difference Δθ corresponding to the mth local model = θ-θ m , according to the weight a corresponding to the model parameters of the mth local model m , adjust the parameter difference and get Δθ=(1-a m )θ m , the weighted parameters of all local models Added to the adjusted parameter difference, we get If you increase a m , the global model parameter θ is closer to the model parameter θ of the mth local model m , so that the local model training loss on the mth local model decreases, and in the next round of communication, the global model θ is used i As the initial weights of all local models.

[0022] Furthermore, step S300 includes the following steps:

[0023] Step S340: r represents the number of communications, E represents the number of training times on the local model, f represents the local model, D i represents the training set of the i-th local model, It represents the model parameters of the i-th local model when the local client trains E times and communicates with the server r times; each source domain client trains a local local model, and Local training sends local model parameters to the server and the corresponding generalization gap

[0024] After the client-side local training is completed, the global model and the local model are generalized and adjusted. The generalization adjustment strategy is expressed by the following formula: in, is the generalization gap of the i-th client after the r-th iteration, ε D (θ r ) is the empirical risk of the global model on the i-th client dataset after the i-th iteration, is the empirical risk of the local model on the i-th client dataset after the i-th iteration, and the value of i ranges from [1, M].

[0025] Furthermore, step S300 includes the following steps:

[0026] Step S350: Calculate the weight of the current i-th local model after replacement based on the generalization errors of all local models and the current weights. The formula is: After simplification, we get where μ is the average generalization gap, d r Is the server parameter, d r =(1-r / R)*d, d is the hyperparameter that controls the modification range, d∈(0,1), r is the current number of iterations, R is the maximum number of iterations, and the maximum range of each generalization adjustment is limited to d r Within the corresponding numerical range, after adjustment, the global model θ r+1 and the generalization weight a r Aggregation, after the end of the rth round of communication, the global model θ r+1 Sent to each local model as the initialization parameter for the next round.

[0027] Furthermore, step S400 includes the following steps: embedding control variables carrying gradient information in the global model and the local model, adjusting the update path of the local model, and ensuring that the update path of the local model is biased towards the optimization direction of the global model.

[0028] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention provides a CSI behavior recognition method based on fairness-adjusted federation, including: collecting a number of sample data generated by different user objects according to fixed CSI behaviors, establishing a federated network model, and obtaining each local model and a global model; obtaining model parameters corresponding to each local model and the global model based on the sample data collected from each source domain client; calculating the minimum empirical risk corresponding to each local model based on the model parameters to obtain the generalization gap of the model; setting the number of loops to form a new global model; distributing the new global model to each source domain client to form a loop to obtain the final trained federated network model. By combining federated learning with a behavior recognition system, the present invention can enable different data owners to collaboratively train models without exchanging data, thereby meeting privacy protection and data security requirements, and improving the accuracy of human behavior recognition while protecting user privacy, thereby promoting technological progress in the field of human behavior recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Schematic diagram of a flow chart of a CSI behavior identification method based on fairness adjustment federation of the present invention;

[0030] Figure 2 A client training network diagram in a CSI behavior recognition method based on fairness adjustment federation according to the present invention;

[0031] Figure 3 A schematic diagram of client drift in a CSI behavior identification method based on fairness adjustment federation of the present invention;

[0032] Figure 4A schematic diagram of a control variable correction principle used in a CSI behavior recognition method based on fairness adjustment federation according to the present invention;

[0033] Figure 5 This is a floor plan of the experimental environment in a CSI behavior recognition method based on fairness adjustment federation of the present invention;

[0034] Figure 6 This is a structural diagram of a data acquisition device in a CSI behavior recognition method based on fairness adjustment federation of the present invention. DETAILED DESCRIPTION

[0035] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0036] Example: Figure 1 As shown, the present invention provides a CSI behavior recognition method and technical solution based on fairness adjustment federation, including the following steps:

[0037] In order to better explain the present invention, the symbols corresponding to the term names are unified here: i represents the i-th local model / client, M represents the number of local models, j represents the j-th sample, a represents the weight of the local model, ε D represents empirical risk, θ represents model parameters, G D Indicates the generalization gap. The specific interpretation should be determined based on the actual situation in the plan.

[0038] Step S100: Collect sample data generated by various user objects according to fixed CSI behaviors. CSI behaviors are human behaviors, and each user object is treated as a source domain client. A federated network model is established, and local models trained in parallel based on the source domain clients in the federated network model are obtained, as well as a global model aggregated from the local models.

[0039] Obtain T different indoor environments and M user objects with different physical characteristics, including height, weight, gender, and age. Set a total of P CSI behaviors, and perform all CSI behaviors for each user object in each indoor environment. Collect data from a user object m performing all CSI behaviors in each indoor environment, with each CSI behavior performed Q times, as the sample data of object m. The total number of samples in the sample data corresponding to object m is T × P × Q.

[0040] Step S200: Based on the collected sample data from each source domain client, a local model and a global model corresponding to each source domain client are trained using a deep learning network based on deconvolution, thereby obtaining model parameters corresponding to each local model and the global model;

[0041] Based on the deconvolution deep learning network, the training of the local model includes: processing the sample data of the user object through the deconvolution layer, converting the signal into a feature map with spatial encoding, and sending the generated feature map to the feature extractor with residual blocks, and performing comprehensive processing on the extracted features through the fully connected layer, thereby realizing the local training model of the deconvolution network. Figure 2 As shown in Figure 1, the deep learning model based on the deconvolutional network includes the generation stage, the feature learning stage and the task stage.

[0042] In summary, the local training model for implementing the deconvolution network can be summarized as: f(x)=T(F(G(x))), where x is the sample on the client, f is the local training model, G is the deconvolution module, F is the feature learning module, and T is the fully connected layer module.

[0043] Step S300: Calculate the minimum empirical risk corresponding to each local model based on the model parameters, and then obtain the generalization gap between the global model and each local model. Calculate the variance of the generalization gap between different local models and the global model, and then obtain the function formula for calculating the minimum empirical risk of the global model. Set the number of loops, and perform aggregation and fusion based on the model parameters and pre-set weights of each local model to form a new global model.

[0044] Step S310: Transmit the model parameters corresponding to each local model and the global model to the server, and obtain the training set corresponding to the i-th local model as 1≤i≤M,N i represents the total number of samples corresponding to the i-th local model, represents the jth sample in the i-th local model, represents the label of the jth sample in the i-th local model. According to each sample in the training set, the minimum empirical risk corresponding to the i-th local model is obtained. The formula is: (The goal of this model is to minimize the empirical risk of all local models to obtain a global model with superior performance), where min is the minimum value, ε D represents the empirical risk, L is the loss function corresponding to the i-th local model, f is the local training model for implementing the deconvolution network, and θ is the model parameter of the i-th local model;

[0045] Since the global model is often prone to conflict with the local model, the local training process may overfit the local data distribution of each client, thereby reducing the generalization performance of the global model. For this reason, the present invention adds proximal term penalty regularization to the loss function of the local model to speed up the convergence speed and prevent statistical heterogeneity caused by excessive updating of the local model.

[0046] Add proximal term penalty regularization to the loss function L of the local model to iterate the global model parameters. Set the maximum number of iterations to r+1, and then adjust the minimum empirical risk formula corresponding to the i-th local model. The adjusted formula is: Among them, λ is the scale parameter that controls the penalty of proximal terms, θ r+1 represents the global model parameters at the r+1th iteration, θ r represents the global model parameters at the rth iteration, |||| 2 To find the square of the Euclidean distance; and then get the minimum empirical risk corresponding to all local models.

[0047] Here, the proximal term penalty regularization is divided by 2 to simplify the expression in the subsequent gradient calculation.

[0048] Step S320: The formula for calculating the generalization gap between the i-th local model and the global model is: in, is the generalization gap between the i-th local model and the global model, M is the number of local models, is the minimum empirical risk of the global model, is the minimum empirical risk of the i-th local model, a i Represents the preset weight of the i-th local model, 1≤i≤M, is the optimal solution of the local model of the i-th local model;

[0049] This paper designs a new federated domain generalization global objective that takes into account the variance of the generalization gap between local clients and the global model to ensure that the optimal global model is fair across all clients. Specifically:

[0050] According to the generalization gap between the global model and each local model, the corresponding variance between all generalization gaps is calculated as: M is the number of local models, is the generalization gap corresponding to the i-th local model; and the objective function formula for calculating the minimum empirical risk of the global model is obtained as follows:

[0051] Where β is the variance coefficient, β∈(0,+∞).

[0052] Among them, β controls the balance between reducing the global loss function and enhancing the fairness of the generalization gap. If β = 0, the model degenerates into the traditional federated averaging (Fedavg) algorithm, whose weight is related to the number of client datasets; when β = ∞, the generalization gap between all domains is equal, that is, the weight occupied by each client is the same and remains unchanged.

[0053] In the framework of federated learning, the local model weight a and model parameter θ in the objective function formula i Cannot be optimized at the same time, there is a certain relationship between a and the corresponding generalization gap, changing the weight a will affect the generalization gap. Therefore, the present invention divides the model optimization into two stages, the local training stage trains the optimal model parameters θ i ,The goal of the aggregation phase is to adjust the weight of each client according to the ,generalization gap to obtain a global model with higher generalization ,accuracy;

[0054] Step S330: Based on the model parameter θ of the mth local model m , and the global model parameter θ, subtract the two to obtain the parameter difference Δθ corresponding to the mth local model = θ-θ m , according to the weight a corresponding to the model parameters of the mth local model m , adjust the parameter difference and get Δθ=(1-a m )θ m , the weighted parameters of all local models Added to the adjusted parameter difference, we get If you increase a m , the global model parameter θ is closer to the model parameter θ of the mth local model m , so that the local model training loss on the mth local model decreases, and in the next round of communication, the global model θ is used i As the initial weights for all local models, this will make θ m and Δθ are closer, if a is reduced i , the situation is the opposite of the above.

[0055] Step S340: r represents the number of communications, E represents the number of training times on the local model, f represents the local model, D i represents the training set of the i-th local model, It represents the model parameters of the i-th local model when the local client trains E times and communicates with the server r times; each source domain client trains a local local model, and Local training sends local model parameters to the server and the corresponding generalization gap

[0056] After the client-side local training is completed, the global model and the local model are generalized and adjusted. The generalization adjustment strategy is expressed by the following formula: in, is the generalization gap of the i-th client after the r-th iteration, ε D (θ r ) is the empirical risk of the global model on the i-th client dataset after the i-th iteration, is the empirical risk of the local model on the i-th client dataset after the i-th iteration, and the value of i ranges from [1, M].

[0057] The purpose of generalization adjustment is to optimize the generalization performance of the global model across all clients, ensuring that the model performs well on all clients. By adjusting the weights of local models, the generalization gap between the global and local models can be reduced, thereby improving the overall performance and fairness of the model.

[0058] Step S350: Calculate the weight of the current i-th local model after replacement based on the generalization errors of all local models and the current weights. The formula is: After simplification, we get where μ is the average generalization gap, d r Is the server parameter, d r =(1-r / R)*d, d is the hyperparameter that controls the modification range, d∈(0,1), r is the current number of iterations, R is the maximum number of iterations, and the maximum range of each generalization adjustment is limited to d r Within the corresponding numerical range, after adjustment, the global model θ r+1 and the generalization weight a r Aggregation, after the end of the rth round of communication, the global model θ r+1 Sent to each local model as the initialization parameter for the next round.

[0059] The principle of dynamically adjusting the weight in step S350 is to consider the generalization error G of each client. D (θ r ) and the average generalization error μ, dynamically adjust the weight of each client. By controlling the modification amplitude, that is, by introducing the hyperparameter d r =(1-r / R)*d controls the amplitude of weight adjustment. As the number of iterations r increases, d r Gradually decrease, so that larger weight adjustments are made in the early stage of training and smaller adjustments are made in the later stage of training to improve the stability of the model.

[0060] The weights are then normalized by dividing each client's weight by the sum of all client weights so that the sum of all client weights is 1.

[0061] In the model of the present invention, the larger the generalization gap of a source domain client, the worse the generalization effect of the global model on that client. In addition, federated learning based on the generalization gap can improve domain generalization capabilities on the basis of strengthening local training methods, with low computational cost and no additional privacy risks. In summary, this method can better protect the privacy of each client by dynamically adjusting the weights based on the generalization gap during training, preventing the model from being biased towards clients with larger sample sizes.

[0062] Step S400: Distribute the new global model to each source domain client, replacing the original local model of each source domain client, forming a cycle. Repeat the cycle process until the global model achieves satisfactory generalization performance and converges on all source domain clients, and obtains the final trained federated network model.

[0063] Step S400 includes the following steps: embedding control variables carrying gradient information in the global model and the local model, adjusting the update path of the local model, and ensuring that the update path of the local model is biased towards the optimization direction of the global model.

[0064] To solve the client drift problem, control variables with gradient information are introduced into the global model and the local model. Client drift means that the update direction of the local model deviates from the optimization target of the global model. The present invention embeds control variables with gradient information into the global and local models. These control variables can adjust the update path of the local model to ensure that its update does not deviate from the optimization direction of the global model, thus avoiding the impact of client drift on the overall accuracy of the model. Figure 3 and Figure 4 shown.

[0065] The proposed combination of federated learning and a WiFi signal-based behavior recognition system offers numerous advantages. As a distributed machine learning architecture, federated learning enables different data owners to collaboratively train models without exchanging data, thus meeting privacy protection and data security requirements. This combination effectively avoids privacy leaks caused by centralized training data, providing new ideas and solutions for the development of behavior recognition technology.

[0066] In the federated learning framework, multiple clients and an aggregation server work together, and the server designs a model to achieve model training under privacy protection. This approach not only solves the privacy issue of data collection, but also adapts to the growing privacy protection regulations and user privacy awareness, and opens up a new path for the development of intelligent human-computer interaction technology. Through federated learning, we can improve the accuracy of human behavior recognition while protecting user privacy, and promote technological progress in fields such as health monitoring, smart home, disaster relief and security monitoring.

[0067] In WiFi-based behavior recognition applications, the generalization ability of behavior recognition models is challenged by differences in WiFi data distribution between different clients, known as domain shift. Federated learning-based WiFi behavior recognition models effectively address this issue when target domain data is available, ensuring client data security. However, in real-world scenarios, domain shift may pose a significant challenge, especially when the server cannot directly access the target client data or the target client data is unavailable.

[0068] The solution proposed in this paper, through privacy-preserving measures within a federated learning framework, not only addresses privacy concerns surrounding data collection but also adapts to growing privacy regulations and user privacy awareness, paving a new path for the development of intelligent human-computer interaction technology. Federated learning can improve the accuracy of human behavior recognition while protecting user privacy, driving technological advancements in areas such as health monitoring, smart homes, disaster relief, and security surveillance.

[0069] The following embodiments are provided, specifically including: 1. deployment of the experimental environment; 2. collection of experimental data; 3. division of sample data; 4. comparison of experimental results.

[0070] 1. Deployment of the experimental environment:

[0071] Three indoor environments with large differences were selected for the experiment, marked as environment A, B and C, as shown in the attached figure. Figure 5 As shown, from left to right are Environments A, B, and C. Environment A is a small activity area equipped with several network devices and office supplies, including but not limited to 3D printing equipment, small machinery, and various miscellaneous items. Environment B is a medium-sized conference room equipped with several furniture items, including but not limited to wooden tables, chairs, and conference tables. Environment C is a larger open indoor area, measuring 16m x 32m, containing several buildings and equipment, including but not limited to glass rooms and machinery. During the data collection process, several staff members frequently entered and exited Environments A, B, and C.

[0072] like Figure 6 As shown, in this embodiment, a behavior recognition system based on a common commercial Intel 5300 wireless device is built. This system is used to collect the data required in this embodiment and test the model performance, including offline experimental data collection, model training, and real-time recognition. The system platform mainly consists of software and hardware. The hardware mainly includes two computers with Ubuntu 12.04 LTS Linux system, one as the CSI signal transmitter and the other as the signal receiver. It also includes an external boost antenna and a built-in Intel 5300 wireless network card. The external boost antenna is used to enhance the quality of the transmission signal, so that the signal receiver receives higher-quality data. The software part is divided into data collection and online real-time recognition interfaces. The data collection interface displays the channel waveform change diagram when collecting data, and the online real-time interface uses the trained model to recognize the behavior of the person in real time and displays it in real time on the host computer interface.

[0073] 2. Collection of experimental data:

[0074] Data were collected from seven users of different heights, weights, body types, and genders in environments A, B, and C. Each user was numbered P1, P2, ..., P7. The data collected was the user's behavior type data. There are seven types of behavior type data: "empty", "jump", "pick", "run", "sit", "walk", and "wave". "Empty" means that there is no user in the environment for corresponding data collection. To ensure data diversity and representativeness, each participant collected data 101 times for each behavior. Since each user collected data in environments A, B, and C, the total sample size for each user is 3 × 7 × 101 = 2121.

[0075] like Figure 6 As shown in the figure, during data collection, a transmitter (Tx) and a receiver (Rx) were set up in each environment. The boost antennas were 1.2 meters above the ground, and the distance between the Tx and Rx antennas was 3.5 meters. To optimize signal reception, three boost antennas were added to both the Tx and Rx terminals, for a total of six external antennas. An Intel 5300 wireless network card was used to extract information from 30 subcarriers of the Wi-Fi signal, thereby obtaining CSI signal characteristics encompassing 30 channels. The sampling rate was set to 500 Hz, with a sampling period of 4 seconds. Each sample contained 500 × 4 = 2000 data packets.

[0076] 3. Sample data division:

[0077] Four users (P1, P2, P5, P6) are randomly selected from seven users, and any three of them (P2, P5, P6) are used as source domain clients. The sample data of each user is divided into training set and test set according to the ratio of 8:2, and saved separately in the client of federated learning. They are not shared with others to protect privacy. Then another user (P1) is used as the target client, and its action data is used as input data and input into the CSI behavior recognition model of the fairness adjustment federation of this scheme to obtain the behavior recognition result of the target client.

[0078] 4. Comparison of experimental results;

[0079] Taking P2, P5, and P6 as source domain clients and P1 as target client, the results of the comparative experiment are shown in Table 1:

[0080] Table 1 Comparative experiment

[0081]

[0082] Central means centralized training among clients, Federated means training using a federated approach, RSC and AM represent different algorithms in the centralized training method Central, Fedavg, Fedprox, and FedSAM represent different algorithms in the Federated training method, and Ours represents training using the fairness-adjusted federated approach of this solution.

[0083] As can be seen from Table 1, the model algorithm proposed in the present invention performs well in average accuracy and achieves the highest accuracy when tested under any setting. This fully proves that the model has strong generalization ability under the federated learning framework. It also shows that the strategy of combining the adjustment of client weights with the improvement of local training is more effective than the generalization model formed by relying solely on local training to improve model accuracy. On the other hand, compared with the centrally trained RSC and AM domain generalization algorithms, the average accuracy of the model of the present invention is 83.51%-72.54%=10.97% and 83.51%-72.08%=11.43% higher, respectively. This further proves that even under the framework of federated learning, the model can still perform good generalization work and form a behavior recognition model with strong generalization ability on unknown clients.

[0084] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A CSI behavior recognition method based on fairness adjustment federation, characterized by: The following steps are involved: Step S100: Collect sample data generated by multiple user objects according to fixed CSI behaviors, where the CSI behaviors are human behaviors, and treat each user object as a source domain client. Establish a federated network model, obtain local models trained in parallel based on each source domain client in the federated network model, and obtain a global model aggregated from the local models. Step S200: Based on the collected sample data from each source domain client, a local model corresponding to each source domain client and the global model are trained using a deep learning network based on deconvolution, thereby obtaining model parameters corresponding to each local model and the global model; Step S300: Calculate the minimum empirical risk corresponding to each local model based on the model parameters, and then obtain the generalization gap between the global model and each local model. Calculate the variance of the generalization gap between different local models and the global model, and then obtain a function formula for calculating the minimum empirical risk of the global model. Set the number of cycles, summarize and fuse the model parameters of each local model and the pre-set weights to form a new global model; Step S400: Distribute the new global model to each source domain client, replacing the original local model of each source domain client, forming a cycle. Repeat the cycle process until the global model achieves satisfactory generalization performance and converges on all source domain clients, and obtains the final trained federated network model.

2. The CSI behavior identification method based on fairness adjustment federation according to claim 1, characterized in that: Step S100 includes the following steps: obtaining T different indoor environments and M user objects with different physical characteristics, where the physical characteristics include height, weight, gender, and age, and setting a total of P CSI behaviors, wherein each user object performs all CSI behaviors in each indoor environment; collecting data in which a user object m performs all CSI behaviors in each indoor environment, and each CSI behavior is performed Q times, as sample data of object m, to obtain a total of T×P×Q samples in the sample data corresponding to the object m.

3. The CSI behavior identification method based on fairness adjustment federation according to claim 2, characterized in that: In step S200, the local model is trained based on the deep learning network based on deconvolution, including: processing the sample data of the user object through the deconvolution layer, converting the signal into a feature map with spatial encoding, and sending the generated feature map to the feature extractor with the residual block, and performing comprehensive processing on the extracted features through the fully connected layer, thereby realizing the local training model of the deconvolution network.

4. The CSI behavior identification method based on fairness adjustment federation according to claim 3, characterized in that: Calculating the minimum empirical risk corresponding to each local model according to the model parameters in step S300 includes the following steps: Step S310: Transmit the model parameters corresponding to each local model and the global model to the server, and obtain the training set corresponding to the i-th local model as 1≤i≤M,N i represents the total number of samples corresponding to the i-th local model, represents the jth sample in the i-th local model, represents the label of the jth sample in the i-th local model. According to each sample in the training set, the minimum empirical risk corresponding to the i-th local model is obtained. The formula is: Among them, min is the minimum value, ε D represents the empirical risk, L is the loss function corresponding to the i-th local model, f is the local training model for implementing the deconvolution network, and θ is the model parameter of the i-th local model; Add proximal term penalty regularization to the loss function L of the local model to iterate the global model parameters. Set the maximum number of iterations to r+1, and then adjust the minimum empirical risk formula corresponding to the i-th local model. The adjusted formula is: Among them, λ is the scale parameter that controls the penalty of proximal terms, θ r+1 represents the global model parameters at the r+1th iteration, θ r represents the global model parameters at the rth iteration, |||| 2 To find the square of the Euclidean distance; and then get the minimum empirical risk corresponding to all local models.

5. The CSI behavior identification method based on fairness adjustment federation according to claim 4, characterized in that: Step S300 includes the following steps: Step S320: The formula for calculating the generalization gap between the i-th local model and the global model is: in, is the generalization gap between the i-th local model and the global model, M is the number of local models, is the minimum empirical risk of the global model, is the minimum empirical risk of the i-th local model, a i Represents the preset weight of the i-th local model, 1≤i≤M, is the optimal solution of the local model of the i-th local model; According to the generalization gap between the global model and each local model, the corresponding variance between all generalization gaps is calculated as: M is the number of local models, is the generalization gap corresponding to the i-th local model; and the objective function formula for calculating the minimum empirical risk of the global model is obtained as follows: Where β is the variance coefficient, β∈(0,+∞).

6. The CSI behavior identification method based on fairness adjustment federation according to claim 5, characterized in that: Step S300 The following steps are involved: Step S330: Based on the model parameter θ of the mth local model m , and the global model parameter θ, subtract the two to obtain the parameter difference Δθ corresponding to the mth local model = θ-θ m , according to the weight a corresponding to the model parameters of the mth local model m , adjust the parameter difference and get Δθ=(1-a m )θ m , the weighted parameters of all local models Added to the adjusted parameter difference, we get If you increase a m , the global model parameter θ is closer to the model parameter θ of the mth local model m , so that the local model training loss on the mth local model decreases, and in the next round of communication, the global model θ is used i As the initial weights of all local models.

7. The CSI behavior identification method based on fairness adjustment federation according to claim 6, characterized in that: Step S300 includes the following steps: Step S340: r represents the number of communications, E represents the number of training times on the local model, f represents the local model, D i represents the training set of the i-th local model, It represents the model parameters of the i-th local model when the local client trains E times and communicates with the server r times; each source domain client trains a local local model, and Local training sends local model parameters to the server and the corresponding generalization gap After the client-side local training is completed, the global model and the local model are generalized and adjusted. The generalization adjustment strategy is expressed by the following formula: in, is the generalization gap of the i-th client after the r-th iteration, ε D (θ r ) is the empirical risk of the global model on the i-th client dataset after the i-th iteration, is the empirical risk of the local model on the i-th client dataset after the r-1th iteration, and the value of i ranges from [1, M].

8. The CSI behavior identification method based on fairness adjustment federation according to claim 7, characterized in that: Step S300 includes the following steps: Step S350: Calculate the weight of the current i-th local model after replacement based on the generalization errors of all local models and the current weights. The formula is: After simplification, we get where μ is the average generalization gap, d r Is the server parameter, d r =(1-r / R)*d, d is the hyperparameter that controls the modification range, d∈(0,1), r is the current number of iterations, R is the maximum number of iterations, and the maximum range of each generalization adjustment is limited to d r Within the corresponding numerical range, after adjustment, the global model θ r+1 and the generalization weight a r Aggregation, after the end of the rth round of communication, the global model θ r+1 Sent to each local model as the initialization parameter for the next round.

9. The CSI behavior identification method based on fairness adjustment federation according to claim 1, characterized in that: Step S400 includes the following steps: embedding control variables carrying gradient information in the global model and the local model, adjusting the update path of the local model, and ensuring that the update path of the local model is biased towards the optimization direction of the global model.

Images