Anti-dual isomerism personalized federal learning method for radar human body activity recognition

Abstract

The invention discloses an anti-dual heterogeneous personalized federal learning method for radar human body activity recognition, and belongs to the crossing field of radar perception, Internet of Things and distributed artificial intelligence. The method comprises the following steps: firstly, constructing structure priori by utilizing sparse graph representation of local data, and realizing personalized weight aggregation based on an internal association relationship between clients so as to eliminate model deviation caused by label distribution difference; meanwhile, a historical knowledge fusion mechanism based on learnable masks is introduced, dynamic matching is performed on a local feature space and a local feature space, and the phenomenon of training turn reduction caused by inconsistent feature distribution is effectively relieved. According to the method, on the premise of strictly protecting the privacy of original data of each radar node, the generalization ability and convergence stability of a human body activity identification model in a non-independent identically distributed scene are remarkably improved, and the dependence on large-scale centralized data is reduced; therefore, the actual deployment and application of the distributed intelligent radar sensing system in the fields of smart home, medical monitoring and the like are promoted.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of radar perception, Internet of Things and distributed artificial intelligence, and relates to radar human activity recognition technology based on federated learning. In particular, it relates to a personalized federated learning method that can simultaneously solve the problems of heterogeneous label distribution and heterogeneous feature distribution in non-independent and identically distributed data scenarios. Background Technology

[0002] Radar-based Human Activity Recognition (HAR) utilizes the micro-Doppler effect to capture the complex frequency shift characteristics generated by human limb movements, and has broad application prospects in IoT smart environments, health monitoring, and security. Compared with vision-based (camera) and wearable device-based solutions, radar sensing has inherent advantages such as non-contact measurement, immunity to lighting conditions, and the ability to protect visual privacy through penetrating detection.

[0003] However, building high-precision and highly generalizable deep learning recognition models typically relies on large-scale, diverse, and representative training data. In practical applications, radar systems are often deployed in a decentralized manner (e.g., in different homes or hospital rooms). Directly aggregating raw radar data from each node to a central server for training is not only costly in terms of transmission but also poses a serious risk of data privacy breaches, as radar echo data may contain sensitive personal behavioral information.

[0004] To resolve the conflict between data silos and privacy protection, Federated Learning (FL) has emerged as a distributed learning paradigm. It allows clients to train models on local data and only upload the model parameters to a server for aggregation, thus achieving collaborative training without sharing the original data. While Federated Learning addresses data privacy issues to some extent, its practical deployment in radar human activity recognition still faces significant challenges due to the presence of non-independent, identically distributed data. Existing technologies exhibit the following significant drawbacks: 1. Ineffective handling of heterogeneous label distribution: In real-world scenarios, different radar nodes operate in different environments or monitor different user groups, resulting in significant differences in the distribution of activity categories collected by each node (e.g., some nodes may only contain "walking" data, while others may contain "falling" data). Traditional federated learning methods (such as FedAvg) typically aim to optimize a globally universal model, employing parameter averaging or simple weighted aggregation strategies, ignoring the differences in local label distributions on each client. This "one-size-fits-all" aggregation approach leads to biases in the decision boundaries of the global model on specific clients, making it difficult to adapt to local, specific activity recognition tasks.

[0005] 2. Difficulty in Overcoming Heterogeneous Feature Distribution: Even for the same action category, due to differences in radar equipment hardware (such as antenna deployment location, waveform parameters, calibration drift, and other instrument-level factors) and the physiological characteristics of different users (such as body size and gait), the generated radar micro-Doppler spectrograms exhibit significant misalignment in the feature space. Existing research shows that this heterogeneity in feature distribution leads to "round descent" in federated learning. That is, after the global model parameters are distributed, the local model, in order to adapt to the global feature space, destroys the locally learned feature representations, similar to the "negative transfer" problem in meta-learning, which seriously hinders model convergence and performance improvement.

[0006] 3. Lack of a unified solution to dual heterogeneity: Existing research on personalized federated learning often focuses on improvements in a single aspect. For example, some methods only alleviate feature bias through local batch normalization (FedBN) or only handle label distribution differences through model similarity weighting. In the specific field of radar human activity recognition, there is currently a lack of a unified framework that can simultaneously address the dual heterogeneity of label and feature distributions.

[0007] In summary, the key challenge that current federated learning-based radar human activity recognition technology urgently needs to address is how to achieve personalized aggregation by leveraging the prior label distributions of different clients within a unified framework, while also eliminating the negative impact of feature space misalignment. Summary of the Invention

[0008] The purpose of this invention is to propose a personalized federated learning method for radar human activity recognition that is resistant to dual heterogeneity in order to solve the problems mentioned in the background art. By combining a class-aware structural similarity weighting mechanism with a learnable historical knowledge fusion strategy, the method utilizes structured prior information and historical model parameters to simultaneously solve the problems of heterogeneous label distribution and feature distribution caused by differences in user environment and device during the federated aggregation process.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A personalized federated learning method for radar human activity recognition that is resistant to dual heterogeneity includes the following steps: S1. Dataset Construction: Construct a radar human activity micro-Doppler dataset containing non-independent identically distributed (Non-IID) characteristics; S2. Network Structure Construction: A personalized federated learning network framework (pFedRadar) based on a client-server architecture is constructed. This framework includes: A client-side local extraction and classification network is used to extract high-dimensional features from radar spectrograms and perform action classification. The local structured prior construction module is used to process local feature vectors through clustering algorithms, construct and upload a structured prior that represents the local label distribution. The server-side personalized aggregation module is used to receive structured priors uploaded by each client, calculate the structural similarity between clients, and generate personalized aggregation weights by combining dynamic filtering mechanisms to synthesize a unique global model for each client. The learnable historical knowledge fusion module dynamically fuses the currently received global model features with the local historical model features retained from the previous round by embedding a learnable mask layer in the client network. S3. Loss Function Design: Construct a multi-objective joint loss function to guide model training. The loss function includes cross-entropy loss, sparsity regularization loss, entropy regularization loss, and mask focusing loss. S4. Model Training and Validation: The personalized federated learning network framework (pFedRadar) is trained and validated using the dataset constructed in S1, specifically including: The server initializes global model parameters and distributes them to each client. In each round of communication, each client receives the global model, uses the learnable historical knowledge fusion module to fuse it with the local historical model, and trains it using local data and a multi-objective joint loss function. After the client-side training is complete, the updated model parameters and the locally built structured priors are uploaded to the server. The server uses the personalized aggregation module to calculate personalized aggregation weights based on the uploaded structured priors, and generates and distributes the next round of personalized models. The classification performance of the final generated personalized model was evaluated using an independent test set.

[0010] Preferably, S1 specifically includes the following: Data acquisition is performed using a monostatic ultra-wideband (UWB) radar system, with the radar system deployed at a height aligned with the center of gravity of the human body. Collect radar echo data of multiple types of human activities, including at least boxing, crawling, creeping, jumping, running, and walking. The original echo signal is subjected to short-time Fourier transform (STFT) to generate a micro-Doppler spectrum, and a sliding window data enhancement strategy is applied to expand the sample. The client samples are divided based on the Dirichlet distribution to simulate the imbalance of label distribution, and each client is restricted to contain only a few action categories to simulate extreme data silos, thereby constructing a Non-IID data scenario.

[0011] Preferably, the local structured prior construction module specifically uses the FINCH clustering algorithm to process feature vectors, constructs a sparse connectivity graph by calculating intra-class cosine similarity, and uploads the mean vector of the sparse graph as a structured prior to the server.

[0012] Preferably, the server-side personalized aggregation module specifically performs the following operations: Calculate the cosine similarity between any two structured priors uploaded by clients; A Gaussian statistical model is constructed based on the similarity calculated in historical communication rounds, and a dynamic threshold is set accordingly to filter out noisy similarity connections. Based on the filtered similarity matrix, personalized weights for model aggregation are generated for each client.

[0013] Preferably, the multi-objective joint loss function is expressed as:

[0014] in, L CE Represents cross-entropy loss; L s This represents the sparsity regularization loss. L e This represents the entropy regularization loss; L m This indicates that the mask focuses on the loss; , This is the balance coefficient.

[0015] Preferably, the cross-entropy loss is used to measure the prediction accuracy of the model on the human activity classification task, in order to calculate the difference between the predicted label and the true label; The mask focusing loss introduces a gradient-guided regularization strategy, which measures the importance of feature dimensions by calculating the gradient magnitude of the task loss relative to the feature vector. The mask focusing loss forces the mask to assign higher weights to feature dimensions with large gradients (i.e., those important to the task), thereby achieving task-aware feature selection. The sparsity regularization loss uses the L1 norm to constrain the mask, causing the elements in the mask to tend towards 0; The entropy regularization loss applies entropy constraints to the mask, causing the mask value to become polarized (tend to 0 or 1), thus avoiding the mask value from remaining in an ambiguous state near the midpoint (0.5).

[0016] Preferably, in step S4, the hyperparameter settings for model training include: Each client uses a stochastic gradient descent (SGD) optimizer for parameter optimization; the learning rate is set to 0.001; the batch size for local training is set to 32; the number of local training epochs within each communication round is set to 5; and the total number of communication rounds is set to 100.

[0017] Preferably, in step S4, the model validation uses an independent test set to evaluate the performance of the generated personalized model. The dataset for each client is divided into a training set and a test set in an 8:2 ratio, and the class distribution between the training set and the test set is kept consistent. The classification accuracy is used as the core evaluation metric.

[0018] The present invention further protects a computer device, the computer device including a processor and a memory, the memory storing at least one instruction, at least one program, code set or instruction set, the instruction, program, code set or instruction set being loaded and executed by the processor to implement the above-mentioned anti-dual heterogeneous personalized federated learning method for radar human activity recognition.

[0019] The present invention further protects a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the instruction, program, code set, or instruction set is loaded and executed by a processor to implement the above-mentioned anti-dual heterogeneous personalized federated learning method for radar human activity recognition.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention proposes a personalized federated learning method for radar human activity recognition that is resistant to dual heterogeneity, which can achieve efficient recognition of radar human activity in non-independent and identically distributed (Non-IID) scenarios. The method proposed in this invention can effectively overcome the problems of heterogeneous label distribution and feature distribution among radar nodes while strictly protecting the privacy of users' original data, and generate a personalized recognition model that can accurately adapt to the local environment and user characteristics.

[0021] (2) The proposed anti-dual heterogeneous personalized federated learning framework (pFedRadar) effectively improves the model's adaptability and training stability in complex heterogeneous environments by introducing a class-aware structural similarity weighting mechanism and a learnable historical knowledge fusion strategy. Compared with traditional federated learning methods (such as FedAvg and FedBN), this invention eliminates model bias during global aggregation by mining structured prior relationships between clients; at the same time, it significantly alleviates the "round descent" phenomenon during training by aligning the feature space using historical knowledge. Under the challenges of extremely imbalanced label distribution and misaligned feature space, this invention shows significant advantages in action recognition accuracy and convergence speed, ensuring efficient collaborative learning on decentralized data.

[0022] (3) This invention not only has significant application value in distributed radar sensing and IoT intelligent monitoring, but can also be extended to other edge computing scenarios that are sensitive to data privacy and have significant individual differences. By achieving personalized collaborative training that relies solely on model parameters and structural priors, this invention significantly reduces the dependence on large-scale centralized labeled data, solves the problem of radar data silos, and provides a new technological path with high robustness and privacy security for the large-scale deployment of smart home, telemedicine, and intelligent security systems in real environments, with broad application prospects. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings involved in the embodiments are now briefly described. Obviously, the drawings in the following description are merely illustrative of some embodiments of the present invention. For those skilled in the art, other forms of drawings can be constructed based on these drawings without creative effort.

[0024] Figure 1 This is the overall architecture diagram of the anti-dual heterogeneous personalized federated learning framework (pFedRadar) proposed in Embodiment 1 of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] This invention proposes a personalized federated learning method for radar-based human activity recognition that resists dual heterogeneity. First, it constructs a structural prior using sparse graph representations of local data, enabling personalized weight aggregation based on the inherent relationships between clients to eliminate model bias caused by differences in label distribution. Simultaneously, it introduces a historical knowledge fusion mechanism based on learnable masks to dynamically align global and local feature spaces, effectively mitigating the "round descent" phenomenon caused by inconsistent feature distributions. This method significantly improves the generalization ability and convergence stability of human activity recognition models in non-independent, identically distributed scenarios while strictly protecting the privacy of the original data of each radar node. It also reduces dependence on large-scale centralized data, thereby promoting the practical deployment and application of distributed intelligent radar sensing systems in fields such as smart homes and medical monitoring.

[0027] The following section further describes the proposed anti-dual heterogeneous personalized federated learning method for radar human activity recognition, in conjunction with the accompanying drawings and relevant examples. The specific details are as follows.

[0028] Example 1: This embodiment proposes a personalized federated learning method for radar human activity recognition that is resistant to dual heterogeneity, specifically including: 1. Dataset Construction This invention constructs a radar human activity micro-Doppler spectrum dataset containing non-independent identically distributed (Non-IID) characteristics for training and validating the proposed federated learning model. Specifically, it includes the following: (1) Radar data acquisition system: Data acquisition was performed using a monostatic ultra-wideband (UWB) radar system. The core of the system was the PulsOn 440 (P440) radar module, operating in the 3.1 to 4.8 GHz frequency band, with a pulse repetition frequency (PRF) of 368 Hz and a coherent processing interval (CPI) of 0.2 seconds. The radar was deployed at a height of 1.2 meters to align with the center of gravity of the personnel.

[0029] (2) Motion capture: Five participants were recruited and performed six specific actions at a distance of 5 meters from the radar: boxing, crawling, creeping, jumping, running, and walking. Each action was repeated to ensure data diversity.

[0030] (3) Signal preprocessing and spectrum generation: The acquired raw radar echo signals were subjected to a short-time Fourier transform (STFT) to generate a micro-Doppler spectrogram. A 0.1-second Hamming window was used, and a sliding window data augmentation strategy (window length 0.1 seconds, overlap rate 90%) was applied to expand the sample size. The final generated spectrograms had uniform resolution and were used as input data for the model.

[0031] (4) Non-IID scenario simulation: To verify the performance of federated learning, two non-independent and identically distributed scenarios were constructed: Dirichlet Non-IID: Utilizing the Dirichlet distribution (parameters) The samples from each client were divided to simulate different degrees of label distribution imbalance.

[0032] Pathological Non-IID: This is an extreme heterogeneous setup where each client contains only a very small number (e.g., 2) of action categories, simulating extreme data silos.

[0033] 2. Build the network structure This invention constructs a personalized federated learning network framework (pFedRadar) based on a client-server architecture, which mainly includes the following functional modules: (1) Local feature extraction and classification network: Each client uses a unified deep neural network as its infrastructure, including a feature extractor and a classifier, used to extract high-dimensional features from radar spectrograms and perform action classification. (2) Local structured prior construction module: To address the issue of label heterogeneity, this module processes feature vectors locally on the client side using the FINCH clustering algorithm. It constructs a sparse connectivity graph by calculating intra-class cosine similarity and uses the mean vector of the sparse graph as a structural prior. This module can extract the label distribution manifold features of the local data without revealing the original data.

[0034] (3) Server-side personalized aggregation module: The server-side deployment employs a class-aware structural similarity weighting mechanism. This module receives structured priors uploaded by each client, calculates the structural similarity between clients, and uses a dynamic filtering mechanism (a Gaussian statistical model based on historical similarity) to filter out noisy connections. Finally, personalized aggregation weights are generated based on the filtered similarity, synthesizing a unique global model for each client.

[0035] (4) Learnable historical knowledge integration module: To address feature heterogeneity, a learnable mask layer is embedded in the client network. This module introduces a trainable mask vector. This mask is used to dynamically fuse currently received global model features with local historical model features retained from the previous round during local training. It acts as a soft gating mechanism, adaptively determining the retention ratio of each dimension in the feature space.

[0036] 3. Loss Function Design To optimize the gating behavior of learnable masks and improve the effectiveness of feature fusion while maintaining classification accuracy, this invention designs a multi-objective joint loss function, which includes the following four parts: (1) Cross-entropy loss ): It is used to measure the prediction accuracy of the model on the human activity classification task and calculate the difference between the predicted label and the true label.

[0037] (2) Mask Focus Loss ): A gradient-guided regularization strategy is introduced. The importance of the feature dimension is measured by calculating the gradient magnitude of the task loss relative to the feature vector. This loss term forces the mask... Higher weights are assigned to feature dimensions with larger gradients (i.e., those important to the task) to achieve task-aware feature selection.

[0038] (3) Sparsity Regularization Loss ): Using L1 norm pair mask Constraints are applied to cause most elements in the mask to tend towards 0. This helps the model retain only the key feature dimensions and reduces redundant information interference.

[0039] (4) Entropy Regularization Loss ): (5) Mask Applying entropy constraints encourages the mask values ​​to become polarized (tending towards 0 or 1), preventing the mask values ​​from remaining in an ambiguous state around 0.5. This enhances the decision-making certainty and interpretability of the gating mechanism.

[0040] In summary, the total loss function ( )for: The ultimate local optimization objective is the weighted sum of the four losses mentioned above:

[0041] in, L CE Represents cross-entropy loss; L s This represents the sparsity regularization loss. L e This represents the entropy regularization loss; L m This indicates that the mask focuses on the loss; , This is the balance coefficient.

[0042] 4. Model Training and Validation This invention designs a complete model training and validation process to ensure the convergence and generalization performance of the pFedRadar federated learning framework under dual heterogeneous data distributions. During the training phase, the entire process follows an iterative interaction pattern between the server and clients. The server first initializes the global model parameters and distributes them to each client. In each round of communication, after receiving the global model, the client does not directly overwrite its local parameters but treats it as global knowledge and fuses it with the local model (historical knowledge) retained from the previous round. Specifically, the client uses local radar micro-Doppler spectrum data for several epochs of local training. After the input samples pass through the feature extractor, a learnable mask vector is used to dynamically weight and fuse the global features and local historical features. Subsequently, the total loss function, including cross-entropy loss, mask focusing loss, sparsity regularization loss, and entropy regularization loss, is calculated, and the model parameters and mask parameters are updated using the absolute value of the gradient through backpropagation. After local training is completed, the client uploads the updated model parameters and the constructed structured prior to the server. The server-side utilizes a class-aware structural similarity weighting mechanism, combined with a dynamic filtering strategy based on historical statistics, to calculate personalized aggregation weights and generate the next round of personalized models.

[0043] To ensure training stability and optimal convergence, this invention employs strict hyperparameter settings during implementation. Each client uses a stochastic gradient descent (SGD) optimizer for parameter optimization, with a learning rate set to 0.001. The batch size for local training is set to 32, the number of local epochs within each communication round is set to 5, and the total number of communication rounds is set to 100. Throughout training, the model architecture remains consistent, employing a deep neural network architecture suitable for radar spectrum classification. This parameter setting balances local computational overhead with global convergence speed, ensuring model stability when handling heterogeneous feature distributions and effectively suppressing the "round descent" phenomenon commonly found in traditional federated learning.

[0044] During the validation phase, this invention uses an independent test set to evaluate the performance of the generated personalized model. Regardless of whether the scenario is Dirichlet non-independent and identically distributed (i.i.d.) or pathological non-independent and identically distributed (i.i.d.) scenario, the dataset for each client is strictly divided into training and test sets in an 8:2 ratio, maintaining consistent class distribution between the training and test sets. The validation process primarily focuses on the model's classification accuracy on the test set. To comprehensively evaluate the effectiveness of this invention, pFedRadar was compared with seven mainstream federated learning methods, including FedAvg, FedBN, FedCAC, FedDWA, LG-Mix, pFedLA, and FedRadar.

[0045] Example 2: Based on Embodiment 1, but with some differences, in order to make the technical solution of the present invention clearer, the following description, in conjunction with the accompanying drawings, further describes the proposed anti-dual heterogeneous personalized federated learning method for radar human activity recognition, specifically including the following: S1, Dataset Construction This invention constructs a radar human activity micro-Doppler spectrum dataset with dual heterogeneous characteristics for training and validating the proposed pFedRadar federated learning model, specifically including: (1) Radar Data Acquisition System Setup: A small monostatic radar detection system was constructed, based on the PulsOn 440 (P440) ultra-wideband radar module manufactured by TimeDomain. This system includes two LB-8180-SF directional horn antennas (one for transmitting and one for receiving), a laptop computer for data processing, and a power bank. The P440 module, as an ultra-wideband coherent wireless transceiver, operates in the 3.1–4.8 GHz frequency band, with a center frequency of 4.3 GHz, a pulse scan frequency (PRF) set to 368 Hz, and a transmit power of 50 μW. The module connects to the computer via a USB interface, and the RF port connects to the antenna via a 50Ω SMA coaxial cable to achieve real-time data communication and control.

[0046] (2) Experimental Environment and Motion Data Acquisition: The experiment was conducted in a static indoor environment. The radar module and antenna were placed on a horizontal table about 1.2 m above the ground (simulating the height of the human body's center of gravity). Subjects performed designated actions within the detection range about 3–5 m directly in front of the antenna. To ensure the diversity and representativeness of the data, five subjects were recruited and performed six typical human actions: walking, running, jumping, boxing, crawling, and creeping. Each action was repeated multiple times in both the "facing the radar" and "away from the radar" directions to cover motion characteristics from different perspectives.

[0047] (3) Signal preprocessing and spectrum generation: In order to eliminate environmental noise and extract high-quality time-frequency features, the following processing steps are performed on the acquired raw radar echo signal: 3.1) Background clutter suppression: The Average Background Subtraction (ABS) method is used to calculate the average background signal over a period of time and subtract it from the original signal in order to eliminate static background clutter and improve the signal-to-noise ratio.

[0048] 3.2) Data augmentation: A sliding window-based slicing augmentation method is adopted, in which a time window is slid across the time-range image and a high overlap rate (e.g., 90%) is set between windows to divide the continuous signal into independent analysis units.

[0049] 3.3) Time-Frequency Transformation: Short-Time Fourier Transform (STFT) is used as the core tool to convert the non-stationary echo signal into a micro-Doppler spectrum. The window length is set to 0.2 s, and the Hamming window function is used while maintaining a 90% overlap rate.

[0050] 3.4) Standardization: The generated micro-Doppler spectrograms are uniformly adjusted to 120×120 pixels and normalized to the [0, 1] interval to eliminate the influence of amplitude differences, and finally form standard samples that can be used as input to neural networks.

[0051] (4) Simulation of Non-Independent Identical Distribution (Non-IID) Scenarios: In order to realistically simulate the data heterogeneity problem faced by federated learning in decentralized deployment, this invention designs a joint simulation scheme for two scenarios: heterogeneous feature distribution and heterogeneous label distribution. 4.1) Feature Distribution Heterogeneous Simulation: Data from 5 subjects were assigned to 5 different clients. Each client only had data from a specific subject. Due to physiological differences in body shape, gait, and movement habits among different subjects, the micro-Doppler features of the same movement category showed significant distribution shifts among different clients, thus simulating feature distribution heterogeneity.

[0052] 4.2) Heterogeneous Label Distribution Simulation: Based on the heterogeneity of features, the non-independent and identically distributed characteristics of label distribution are further simulated through the following two settings: 4.2.1) Dirichlet Non-IID Settings: Utilizes the Dirichlet Distribution to divide the sample categories for each client. Set the category distribution to follow... ,in This is a concentration parameter. When... When the value is small (e.g., 0.1), the distribution is extremely sparse, and the proportion of action categories possessed by different clients varies greatly; when... When the value is large, the distribution tends to be uniform. This setting simulates the differences in action frequency caused by different user preferences or environments.

[0053] 4.2.2) Pathological Non-IID Setting: Simulates extreme data silo scenarios. Each client is limited to accessing only a very small number of the six action classes (e.g., each client has access to only two action classes, and the classes may not overlap between different clients). This setting creates extreme label skew and is used to test the model's generalization ability in situations where there is a severe lack of samples for certain classes.

[0054] S2. Building the network structure This invention proposes a pFedRadar personalized federated learning framework, whose system workflow is as follows: Figure 1 As shown. This system mainly involves interaction and collaboration between the server and the client. The specific implementation steps are as follows: S2.1 Local Model Initialization: The server distributes the initialized global model parameters to each client. During the local training phase, each client... Based on local dataset Perform local optimization and establish local model parameters. For classification tasks, standard cross-entropy loss is used for training, and the parameters are ready to be uploaded after training.

[0055] S2.2 Structural Prior Construction: To address the heterogeneity of label distribution, the client constructs a structured prior representing the label distribution locally. Based on locally extracted feature vectors, the client utilizes sparse graph representation and nearest neighbor connectivity algorithms to construct a structured prior that compactly describes the local data distribution manifold. This data is used for subsequent uploading. This process does not require disclosing the original sample data.

[0056] S2.3 Parameter Upload: Each client will upload the updated local model parameters. and the constructed structured priors The data is packaged and uploaded to the central server. These structured priors are the key basis for subsequent personalized aggregation by the servers.

[0057] S2.4 Model Aggregation: After receiving the data, the server performs personalized aggregation: S2.4.1 Similarity Calculation: Calculation Client and Client similarity is obtained by using the Euclidean distance between structured priors on shared categories. .

[0058] S2.4.2 Dynamic Filtering: A Gaussian statistical model is established based on historical similarity data, and a threshold is set. Noisy connections are filtered out to obtain a refined similarity matrix.

[0059] S2.4.3 Weighted Aggregation: Calculate aggregation weights based on filtered similarity. For each client Generate personalized global models .

[0060] S2.4.5. The Personalized Model Distribution server aggregates the generated personalized model parameters. Accurately distribute back to the corresponding client Unlike traditional federated learning that distributes a unified global model, in this step, each client receives model parameters that are customized based on its similarity to other clients.

[0061] S2.5.6 Local Model Update: The client receives the updated model. Subsequently, instead of directly overwriting the local parameters, they are used as the basis for the next round of local training. In this step, a learnable historical knowledge fusion mechanism is introduced: the client retains the previous round's local model as "historical knowledge," utilizing learnable mask vectors... Dynamically fuse currently received model features (global knowledge) with historical model features (local knowledge) to align the feature space, solve the "round decline" problem caused by feature heterogeneity, and complete the update and optimization of the local model.

[0062] S3, Model Training and Validation This invention designs a rigorous model training and verification process, and verifies the superiority of the pFedRadar framework in handling the dual heterogeneity of radar human activity recognition through comparative experiments in two typical non-independent identically distributed (Non-IID) scenarios.

[0063] (1) Training process implementation details: The training process adopts an iterative model of server and client collaboration, and a total of There are several communication rounds. Each round of communication includes the following specific operations: 1.1) Parameter distribution: The server distributes the personalized model parameters generated in the previous round of aggregation to the corresponding clients.

[0064] 1.2) Local Optimization: After receiving the model, the client does not directly overwrite it, but instead fuses it as global knowledge with the locally stored historical models. The client uses a local dataset for training, setting the number of local training epochs to 5 and the batch size to 32. The optimizer used is stochastic gradient descent (SGD), with an initial learning rate set to 0.001. During backpropagation, the total loss function is simultaneously calculated. For model weights and mask parameters The gradient is calculated and both are updated synchronously.

[0065] 1.3) Upload and Aggregation: After local training is complete, the client uploads the updated model parameters and the reconstructed structured prior to the server. The server uses the dynamically filtered structural similarity matrix to calculate personalized aggregation weights and generate the next round of personalized models.

[0066] (2) Validation Scheme Setup: To comprehensively evaluate model performance, two heterogeneous scenarios were set up: Dirichlet Non-IID (simulating different degrees of label distribution imbalance) and Pathological Non-IID (simulating extreme data silos). Dataset partitioning followed strict standards: the data for each client was divided into training and test sets in an 8:2 ratio, ensuring that the class distribution within the training and test sets remained consistent to avoid evaluation errors caused by distribution shifts. The evaluation metric used was classification accuracy, and our invention (Ours / pFedRadar) was compared with seven mainstream methods, including benchmark methods FedAvg and FedBN, as well as advanced personalized federated learning methods FedCAC, FedDWA, LG-Mix, pFedLA, and FedRadar.

[0067] (3) Experimental results and analysis: 3.1) Performance Analysis in Dirichlet Non-IID Scenarios Table 1 shows the parameters at different Dirichlet concentrations. Comparison of the average accuracy of each method on 5 clients at (0.1, 0.5, 1.0). The smaller the value, the more unbalanced the label distribution and the stronger the heterogeneity.

[0068] Table 1. Comparison of average accuracy of various methods in Dirichlet Non-IID scenarios.

[0069] Results Analysis: As shown in Table 1, the present invention (pFedRadar) performs well in all... The highest average accuracy was achieved under all settings. In extremely heterogeneous scenarios ( Under these conditions, traditional methods FedAvg and FedBN perform extremely poorly, with accuracies of only 48.18% and 50.54%, respectively, indicating that simple parameter averaging or feature normalization alone cannot cope with severe label shifts. In contrast, this invention achieves an extremely high accuracy of 96.08%, even surpassing the 94.81% accuracy of local training. This demonstrates that pFedRadar not only overcomes heterogeneity but also effectively utilizes the knowledge of other clients to enhance the local model through federated collaboration. Compared to FedDWA (95.30%), which also employs a personalized mechanism, this invention maintains its advantage, proving that a dual strategy combining "structured priors" and "historical knowledge fusion" is more effective than a single weight adjustment.

[0070] 3.2) Performance Analysis in Pathological Non-IID Scenarios This scenario simulates an extreme case of data silos, where each client has only two types of action data and the label distribution is extremely skewed.

[0071] Table 2. Comparison of average accuracy of various methods in pathological non-IID scenarios.

[0072] Results Analysis: As shown in Table 2, in this highly challenging scenario, FedAvg's performance significantly decreased to 68.12%, demonstrating the failure of the globally uniform model under extremely personalized requirements. Our invention (pFedRadar) achieved the highest average accuracy of 91.12%. This is attributed to our structured prior mechanism, which accurately captures potential relationships between clients even when client categories have little or no overlap, enabling effective knowledge transfer. Notably, our invention outperforms all comparable state-of-the-art methods (such as FedDWA's 90.06% and LG-Mix's 90.02%), fully demonstrating pFedRadar's robustness and superiority in handling extreme label and feature heterogeneity.

[0073] Through the above training and verification process, the pFedRadar framework of this invention achieves high-precision and high-stability personalized recognition under the severe challenges of dual heterogeneity in label distribution and feature distribution. It effectively solves the model bias and training "round decline" problems existing in distributed radar perception, and provides a new technical path for decentralized radar perception systems to solve the data silo problem and achieve collaborative intelligence under privacy protection.

[0074] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A personalized federated learning method for radar human activity recognition that is resistant to dual heterogeneity, characterized in that, Includes the following steps: S1. Dataset Construction: Construct a radar human activity micro-Doppler dataset containing non-independent and identically distributed characteristics; S2. Network Structure Construction: A personalized federated learning network framework based on a client-server architecture is constructed. This framework includes: A client-side local extraction and classification network is used to extract high-dimensional features from radar spectrograms and perform action classification. The local structured prior construction module is used to process local feature vectors through clustering algorithms, construct and upload structured priors representing the local label distribution; The server-side personalized aggregation module is used to receive structured priors uploaded by each client, calculate the structural similarity between clients, and generate personalized aggregation weights by combining dynamic filtering mechanisms to synthesize a unique global model for each client. The learnable historical knowledge fusion module dynamically fuses the currently received global model features with the local historical model features retained from the previous round by embedding a learnable mask layer in the client network. S3. Loss Function Design: Construct a multi-objective joint loss function to guide model training. The loss function includes cross-entropy loss, sparsity regularization loss, entropy regularization loss, and mask focusing loss. S4. Model Training and Validation: The personalized federated learning network framework is trained and validated using the dataset constructed in S1, specifically including: The server initializes global model parameters and distributes them to each client. In each round of communication, each client receives the global model, uses the learnable historical knowledge fusion module to fuse it with the local historical model, and trains it using local data and a multi-objective joint loss function. After the client-side training is complete, the updated model parameters and the locally built structured priors are uploaded to the server. The server uses the personalized aggregation module to calculate personalized aggregation weights based on the uploaded structured priors, and generates and distributes the next round of personalized models. The classification performance of the final generated personalized model was evaluated using an independent test set.

2. The anti-dual heterogeneous personalized federated learning method for radar human activity recognition according to claim 1, characterized in that, S1 specifically includes the following: Data acquisition is performed using a monostatic ultrawideband radar system, with the radar system deployed at a height aligned with the center of gravity of the human body. Collect radar echo data of various human activities, including at least boxing, crawling, jumping, running, and walking; The original echo signal is subjected to short-time Fourier transform to generate a micro-Doppler spectrum, and a sliding window data augmentation strategy is applied to expand the sample. The client samples are divided based on the Dirichlet distribution to simulate the imbalance of label distribution, and each client is restricted to contain only a few action categories to simulate extreme data silos, thereby constructing a Non-IID data scenario.

3. The anti-dual heterogeneous personalized federated learning method for radar human activity recognition according to claim 1, characterized in that, The local structured prior construction module specifically uses the FINCH clustering algorithm to process feature vectors, constructs a sparse connectivity graph by calculating intra-class cosine similarity, and uploads the mean vector of the sparse graph as a structured prior to the server.

4. The anti-dual heterogeneous personalized federated learning method for radar human activity recognition according to claim 1, characterized in that, The server-side personalized aggregation module specifically performs the following operations: Calculate the cosine similarity between any two structured priors uploaded by clients; A Gaussian statistical model is constructed based on the similarity calculated in historical communication rounds, and a dynamic threshold is set accordingly to filter out noisy similarity connections. Based on the filtered similarity matrix, personalized weights for model aggregation are generated for each client.

5. The anti-dual heterogeneous personalized federated learning method for radar human activity recognition according to claim 1, characterized in that, The multi-objective joint loss function is expressed as follows: in, L CE Represents cross-entropy loss; L s This represents the sparsity regularization loss. L e This represents the entropy regularization loss; L m This indicates that the mask focuses on the loss; , This is the balance coefficient.

6. The anti-dual heterogeneous personalized federated learning method for radar human activity recognition according to claim 5, characterized in that, The cross-entropy loss is used to measure the model's prediction accuracy on the human activity classification task, in order to calculate the difference between the predicted label and the true label. The mask focusing loss introduces a gradient-guided regularization strategy, which measures the importance of feature dimensions by calculating the gradient magnitude of the task loss relative to the feature vector. The mask focusing loss forces the mask to assign higher weights to feature dimensions with large gradients, thereby achieving task-aware feature selection. The sparsity regularization loss uses the L1 norm to constrain the mask, causing the elements in the mask to tend towards 0; The entropy regularization loss applies entropy constraints to the mask, causing the mask values ​​to become polarized and preventing the mask values ​​from remaining in an ambiguous state near the midpoint.

7. The anti-dual heterogeneous personalized federated learning method for radar human activity recognition according to claim 1, characterized in that, In S4, the hyperparameter settings for model training include: Each client uses a stochastic gradient descent optimizer for parameter optimization; the learning rate is set to 0.001; the batch size for local training is set to 32; the number of local training rounds within each communication round is set to 5; and the total number of communication rounds is set to 100.

8. The anti-dual heterogeneous personalized federated learning method for radar human activity recognition according to claim 7, characterized in that, In S4, the model validation uses an independent test set to evaluate the performance of the generated personalized model. The dataset for each client is divided into a training set and a test set in an 8:2 ratio, and the class distribution between the training set and the test set is kept consistent. The classification accuracy is used as the core evaluation metric.

9. A computer device, characterized in that, The computer device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set, or instruction set, and the instruction, program, code set, or instruction set is loaded and executed by the processor to implement the anti-dual heterogeneous personalized federated learning method for radar human activity recognition as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, at least one program, code set, or instruction set, which is loaded and executed by a processor to implement the anti-dual heterogeneous personalized federated learning method for radar human activity recognition as described in any one of claims 1-8.

Images