Research On Human Activity Recognition With Non-IID Data

Human activity recognition is an important research direction in ubiquitous computing,aimed at sensing the sensor data signals and analyzing and recognizing the current activity status of the human body.However,in practical scenarios,the sensor data collected from different users and environments is non-independent and identically distributed.This can cause human activity recognition models to be unable to effectively recognize activities.In order to solve this problem,cross-domain and personalized human activity recognition models are used to adaptively train models for different domain data to improve model accuracy.However,current cross-domain and personalized human activity recognition methods still face many challenges.Firstly,personalized human activity recognition lacks personalized feature extraction and protection of user privacy information in the feature extraction stage.Secondly,the cross-domain human activity recognition model does not consider the source domain selection and its reasons,leading to negative transfer and lack of interpretability.Thirdly,personalized activity recognition has a “cold start” problem for new users due to the heterogeneity of data.Fourthly,different data heterogeneity in different user contexts in personalized activity recognition leads to performance degradation.Fifthly,the low effciency of federated learning models and high data labeling costs in cross-domain activity recognition.To address these challenges,this thesis conducts research on cross-domain human activity recognition based on various domain adaptive models.The main research contents of this thesis are as follows:Firstly,a federated multi-task attention framework is proposed to address the heterogeneous data and privacy issues.The framework can effectively extract global shared features and individual unique features based on personalized attention mechanisms.In addition,the framework can effectively protect user privacy through localized training and parameter transfer by federated training.Experimental results on four public benchmark datasets demonstrate the superiority of the framework.Secondly,a meta-transfer model for cross-user human activity recognition is proposed to address the problem of source domain selection and interpretability.The model uses a linear weighted scoring function based on interpretable meta-features to rank and select source domains.During the training phase,the model uses a meta-learning framework and Bayesian optimization to train the domain selection module parameters.Experimental results show that the model can effectively improve performance and provide interpretability analysis based on meta-features.In addition,the model performs well on small sample datasets and datasets with missing data.Thirdly,a diversity-aware activity recognition model based on the federated meta-learning framework is proposed to address the“cold start” problem.Firstly,the model clusters similar users.Then it uses a global feature extraction network to extract shared features between all users and uses an attention module in each distributed network to extract specific features for each group of users.The framework applies meta-learning to the federated learning architecture to effciently train personalized networks for new users.Experimental results on real-world datasets show that the model has better generalization ability.Fourthly,a diversity-aware activity recognition model based on hierarchical multi-task learning is proposed to consider the heterogeneity under different contexts.Firstly,the model uses a multi-task learning model to model the differences and similarities between users by treating activity recognition of different individuals as multiple related tasks.Secondly,to model the correlation between activity and context,the model further treats each user’ s context recognition as an auxiliary task for activity recognition.In this way,the implicit relationship between context and activity can be extracted.Experimental results show that using context information and multi-task learning can effectively improve model performance in crossdomain activity recognition tasks.Finally,a clustering federated learning method is proposed to address the low effciency of federated learning.The method uses client-side local batch normalization to handle nonindependent and identically distributed data from the perspective of feature shift.Then it uses a method of calculating model parameter similarity to achieve the unsupervised classification of clients.In this way,model robustness and convergence speed can be improved during training.Experimental results on three real datasets demonstrate that the method has better performance and higher effciency compared to classical federated learning methods.