The Research Of Cross-user Activity Recognition Based On Deep Learning And Unsupervised Domain Adaptation

Human activity recognition has attracted much attention in recent years.It can be widely used in intelligent nursing,smart home and human-computer interaction.Wearable sensors have been widely studied and applied because of their low cost,portability and privacy protection.In the field of human activity recognition based on wearable sensors,the universality of existing models is usually poor.Since there are usually behavior differences between the target users and the users in the training set,the model fails to be well applied to different users.However,it is not realistic to collect a large amount of labeled data of target users to retrain the model.Aiming at the user personalization problem of activity recognition,this thesis proposes and implements a cross-user activity recognition model based on unsupervised deep domain adaptation called CUAS,and finally carries out multiple compared experiments on six datasets to verify the effectiveness of the proposed method.The main research work of this thesis includes:(1)Feature extraction method based on deep neural network.In order to extract the temporal features of sensor data more effectively,the Attribute Temporal Convolution and Segmentation Long Short-Term Memory Network(ATC-SLSTM)is proposed,which divides the temporal samples into several local segments and extracts the time correlation between segments by LSTM.The Attribute Temporal Convolutional Neural Network is proposed to extract features for each local segment.It uses multi-layer small convolution kernels to extract multi-granularity temporal features for each attribute.What's more,the attribute feature scaling mechanism is proposed to realize the dynamic weighting of each data attribute to emphasize the importance of different attributes.(2)Unsupervised domain adaptation method based on the feature distribution alignment.In order to solve the problem of behavior difference between target users and users in training dataset,this thesis uses ATC-SLSTM to map their behavior data to the same feature space,and then aligns the marginal distribution as well as the class conditional distribution in the feature space.In order to improve the alignment effect,Variance and Mean Discrepancy(VMD)is proposed to measure the distribution distance for marginal distribution alignment.In addition,an improved method of adding variance constraints and class weights to VMD is proposed to improve the alignment effect of class conditional distribution and pay more attention to the classes that are easy to be misclassified.(3)Self-training method for the cross-user activity recognition model CUAS.The quality of pseudo-labels of target user is the key to the class conditional distribution alignment,so it directly affects the performance of the model.In the self-training method proposed in this thesis,the minimum entropy is used to select optimal model for the next round of training in each round.Then,the model is used to predict pseudo-labels,and a confidence evaluation method based on Monte Carlo dropout is proposed to select samples with high quality pseudo-labels to train the model iteratively,taking into account the problem of class-imbalance.Experiments show that the CUAS model optimized by self-training method achieves the best effect for cross-user activity recognition problem compared with the other four methods.