Search

CN-122020541-A - Anti-double-isomerism personalized federal learning method for radar human activity recognition

CN122020541ACN 122020541 ACN122020541 ACN 122020541ACN-122020541-A

Abstract

The invention discloses an anti-double heterogeneous personalized federal learning method for radar human activity recognition, and belongs to the field of radar perception, internet of things and distributed artificial intelligence intersection. According to the invention, a sparse graph representation construction structure of local data is utilized to realize personalized weight aggregation based on an internal association relationship between clients so as to eliminate model deviation caused by label distribution difference, and meanwhile, a history knowledge fusion mechanism based on a learnable mask is introduced to dynamically align global and local feature spaces, so that the phenomenon of 'round drop' of training caused by inconsistent feature distribution is effectively relieved. According to the method, on the premise of strictly protecting the privacy of the original data of each radar node, the generalization capability and convergence stability of the human body activity recognition model under the non-independent co-distribution scene can be obviously improved, and the dependence on large-scale centralized data is reduced, so that 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.

Inventors

  • WANG JIAN
  • YANG YANG

Assignees

  • 天津大学

Dates

Publication Date
20260512
Application Date
20260129

Claims (10)

  1. 1. The anti-double heterogeneous personalized federal learning method for radar human activity recognition is characterized by comprising the following steps of: S1, constructing a data set, namely constructing a radar human body activity micro Doppler data set containing non-independent co-distribution characteristics; s2, building a network structure, namely building a personalized federal learning network framework based on a client-server architecture, wherein the framework comprises the following components: The client side local extraction and classification network is used for extracting high-dimensional features from the radar spectrogram and performing action classification; The local structured prior construction module is used for processing the local feature vector through a clustering algorithm, and constructing and uploading a structured prior representing the distribution of the local tags; The server personalized aggregation module is used for receiving the structured prior uploaded by each client, calculating the structural similarity between the clients, generating personalized aggregation weight by combining a dynamic filtering mechanism, and synthesizing a dedicated global model for each client; the learning history knowledge fusion module dynamically fuses the currently received global model features and the local history model features reserved in the previous round by embedding a learning mask layer in the client network; s3, designing a loss function, namely constructing a multi-target joint loss function guiding model training, wherein the loss function comprises cross entropy loss, sparsity regularization loss, entropy regularization loss and mask focusing loss; S4, training and verifying the model, namely training and verifying the personalized federal learning network framework by using the data set constructed in the S1, wherein the training and verifying comprises the following steps of: The server initializes global model parameters and distributes the parameters to each client; in each round of communication, each client receives a global model, fuses the global model with a local history model by utilizing a learnable history knowledge fusion module, and trains by using local data and a multi-target joint loss function; After the training of the client is completed, uploading the updated model parameters and the locally constructed structured priori to a server; The server calculates personalized aggregation weight based on the uploaded structured prior by using a personalized aggregation module, and generates and transmits a personalized model of the next round; and evaluating the classification performance of the finally generated personalized model by using the independent test set.
  2. 2. The radar human activity recognition-oriented anti-double heterogeneous personalized federal learning method according to claim 1, wherein S1 specifically comprises the following contents: Adopting a single-base ultra-wideband radar system to acquire data, and arranging the radar system to be highly aligned with the gravity center of a human body; collecting radar echo data of various human activities, including at least boxing, creeping, crawling, jumping, running and walking; performing short-time Fourier transform on the original echo signal to generate a micro Doppler spectrogram, and expanding a sample by applying a sliding window data enhancement strategy; Dividing each client sample based on dirichlet distribution to simulate tag distribution imbalance, and limiting each client to only contain a few action categories to simulate extreme data islands so as to construct Non-IID data scenes.
  3. 3. The radar human activity recognition-oriented dual-isomerism resistant personalized federal learning method according to claim 1, wherein the local structured priori construction module specifically adopts FINCH clustering algorithm to process feature vectors, constructs a sparse connection graph by calculating intra-class cosine similarity, and uploads a mean vector of the sparse graph as a structured priori to a server side.
  4. 4. The radar human activity recognition-oriented dual-isomerism resistant personalized federal learning method according to claim 1, wherein the server personalized aggregation module specifically performs the following operations: Calculating cosine similarity between the structured priors uploaded by any two clients; establishing a Gaussian statistical model based on the calculated similarity in the historical communication turns, and setting a dynamic threshold value according to the Gaussian statistical model to filter noise similarity connection; based on the filtered similarity matrix, personalized weights for model aggregation are generated for each client.
  5. 5. The radar human activity recognition oriented anti-double heterogeneous personalized federal learning method according to claim 1, wherein the multi-objective joint loss function is expressed as: Wherein L CE represents a cross entropy loss, L s represents a sparsity regularization loss, L e represents an entropy regularization loss, and L m represents a mask focus loss; 、 Is a balance coefficient.
  6. 6. The radar human activity recognition-oriented anti-double heterogeneous personalized federal learning method according to claim 5, wherein the cross entropy loss is used for measuring prediction accuracy of a model on a human activity classification task to calculate a difference between a prediction tag and a real tag; The mask focusing loss introduces a regularization strategy of gradient guidance, and the importance of the feature dimension is measured by calculating the gradient amplitude of the task loss relative to the feature vector; the mask focusing loss forces the mask to give higher weight to the feature dimension with large gradient so as to realize task-aware feature selection; The sparsity regularization loss adopts L1 norms to restrict the mask, so that elements in the mask tend to be 0; the entropy regularization loss applies entropy constraint to the mask, promotes two polarizations of the mask value, and avoids a fuzzy state where the mask value stays near a midpoint.
  7. 7. The radar human activity recognition oriented anti-double heterogeneous personalized federal learning method according to claim 1, wherein in S4, the super-parameter setting of training of the model comprises: the clients adopt a random gradient descent optimizer to perform parameter optimization, the learning rate is set to be 0.001, the batch size of the local training is set to be 32, the number of the local training rounds in each communication round is set to be 5, and the total communication round is set to be 100 rounds.
  8. 8. The radar human activity recognition-oriented dual-heterogeneous personalized federal learning method according to claim 7, wherein in S4, the model verification adopts an independent test set to evaluate the performance of the generated personalized model, the data set of each client is divided into a training set and a test set according to the ratio of 8:2, the class distribution between the training set and the test set is kept consistent, and the classification accuracy is used as a core evaluation index.
  9. 9. A computer device comprising a processor and a memory, wherein the memory stores 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 radar human activity recognition-oriented dual heterogeneous personalized federal learning method of any of claims 1-8.
  10. 10. A computer readable storage medium having stored therein at least one instruction, at least one program, code set, or instruction set that is loaded and executed by a processor to implement the radar human activity identification-oriented dual heterogeneous personalized federal learning method of any of claims 1-8.

Description

Anti-double-isomerism personalized federal learning method for radar human activity recognition Technical Field The invention belongs to the crossing field of radar perception, internet of things and distributed artificial intelligence, relates to a radar human body activity recognition technology based on federal learning, and particularly relates to a personalized federal learning method capable of simultaneously solving the problems of tag distribution isomerism and characteristic distribution isomerism under a non-independent co-distributed data scene. Background Human body activity recognition (Human Activity Recognition, HAR) based on radar utilizes Micro-Doppler Effect (Micro-Doppler Effect) to capture complex frequency shift characteristics generated by human body limb movement, and has wide application prospects in the fields of intelligent environments of the Internet of things, health monitoring and security. Compared to solutions based on vision (cameras) and wearable devices, radar perception has the natural advantages of non-contact measurement, immunity from lighting conditions, and the ability to protect visual privacy through penetration detection. However, building highly accurate and generalizing deep learning recognition models typically relies on extensive, diverse and representative training data. In practical applications, radar systems are usually deployed in a decentralised manner (e.g. in different homes or hospital rooms), and directly summarizing the raw radar data of each node to a central server for training is not only costly to transmit, but also faces serious risk of data privacy disclosure, because radar echo data may contain sensitive personal behavior information. To address the contradiction between data islands and privacy protection, federal learning (FEDERATED LEARNING, FL) has emerged as a distributed learning paradigm. The method allows each client to train the model on the local data, and only uploads the model parameters to the server for aggregation, so that collaborative training is realized on the premise of not sharing the original data. Although federal learning solves the data privacy problem to a certain extent, in actual deployment of radar human activity recognition, the following significant drawbacks exist in the prior art: 1. The tag distribution is not effectively heterogeneous, namely, in an actual scene, the environments where different radar nodes are located or the monitored user groups are different, so that the distribution of the activity categories collected by each node is huge in difference (for example, some nodes only contain walking data, and other nodes possibly contain falling data). Conventional federal learning methods (e.g., fedAvg) are generally directed to optimizing a global generic model, employing parameter averaging or simple weighted aggregation strategies, ignoring differences in the local label distribution of each client. This "cut-in-one" aggregation approach can lead to decision boundary deviations of the global model on a particular client that are difficult to accommodate for the local specific activity recognition task. 2. The characteristic distribution isomerism is difficult to overcome, even if the characteristics are the same, due to the hardware difference of radar equipment (such as the antenna due to the deployment position, waveform parameters, calibration drift and other instrument level factors) and the physiological characteristics (such as body type and gait) difference of different users, the generated radar micro Doppler spectrogram has obvious misalignment phenomenon in the characteristic space. The prior research shows that the isomerism of the feature distribution can cause the phenomenon of round drop in the federal learning process, namely, after the global model parameters are issued, the local model destroys the locally learned feature representation in order to adapt to the global feature space, and is similar to the problem of 'negative migration' in meta learning, thereby seriously impeding the convergence and the performance improvement of the model. 3. The lack of a unified solution to dual isomerism-existing personalized federal learning studies tend to focus on improvements in a single aspect. For example, some approaches only mitigate feature bias through local batch normalization (FedBN), or only handle tag distribution differences through model similarity weighting. In the specific field of radar human activity recognition, a unified framework capable of simultaneously solving the dual isomerism of tag distribution and feature distribution is not yet available. In summary, how to utilize tag distribution priors of different clients to realize personalized aggregation and eliminate negative effects caused by misalignment of feature spaces in a unified framework is a key problem to be solved in the current radar human activity recognition technology based on federal learning. Disclosure of