Research And Application Of Time Series Classification Algorithm Based On Human Activity Recognition

In recent years,with the continuous development of sensor technology and intensive study of activity recognition theory,sensor-based human activity recognition(HAR)has attracted extensive attention due to its excellent privacy,anti-interference and convenience.Deep learning avoids the time-consuming feature engineering stage of traditional methods in the way of end-to-end feature learning largely,which can be more accurate to represent human activity in complex scenes.At present,a large number of studies have applied deep learning to sensor-based HAR tasks.Although sensor-based HAR has made great progress,it still faces many difficulties and challenges.Firstly,HAR usually involves the records of multiple sensor channels,different sensor channels often contain different and complementary information,it becomes difficult to represent activity effectively by ignoring the potential correlation of sensor channels.Secondly,human activity has inter-class similarity and intra-class diversity,so the training process should not only ensure inter-class separation,but also ensure the intra-class compactness,which increases the difficulty of activity recognition;Finally,the essence of HAR is multi-dimensional time series signal modeling with long-term dependencies,so how to effectively learn temporal dependencies still needs to be explored.In order to solve the above problems,the specific research work is as follows:(1)This thesis proposes a human activity recognition algorithm based on two-stream Transformer.The two-stream structure sufficiently considers the correlation of sensor channels and time,and the self attention has global temporal information in the way of parallel computing,which solves the long-term temporal dependencies problem well.In addition,Gaussian range coding is used to replace the trigonometric function coding of the original Transformer,so as to preserve the temporal information and strengthen the interaction of time;The multi-scale convolution layer replaces the feedforward layer to capture the local features of different scales,so as to improve the feature extraction ability of model.(2)This thesis proposes a human activity recognition algorithm based on attention mechanism and center loss.Self-attention in the algorithm captures the potential correlation of sensor channels,so as to enrich the expression of convolutional features;The attention temporal encoder adds attention to the recurrent neural network to model the time context and further learn the temporal dependencies effectively;The cross entropy loss combined with the center loss can maximize the difference between classes and minimize the difference within classes,so as to obtain more discriminant features.In addition,this thesis also designs serial model structure and parallel model structure for feature extraction,the results show that parallel structure can capture the original temporal information of sensor data better and have better performance for activity recognition.For the above two algorithms,a large number of experiments are carried out on UCI HAR,WISDM and Opportunity datasets.The results show that the two algorithms all have good classification accuracy,especially the accuracy of WISDM reaches97.19% and 98.47% respectively,which can achieve good classification results for similar activities and activities in opposite order;In addition,through hyperparameter analysis and ablation research,the experimental results verify the effectiveness of the proposed algorithm on the improvement of the original algorithm,which shows that the proposed algorithm can further improve the ability of deep learning features;Finally,by comparing with many deep learning algorithms,the classification results of this algorithm are better than the comparative algorithms,which verifies the superiority of this algorithm in activity classification.