Research And Application Of Semi-supervised Time Series Classification Algorithm For Human Activity Recognition

Time series data exist in almost all tasks that need human cognition.Any problem of classification using time series data can be considered as time series classification tasks,such as human activity recognition.In recent years,with the development of sensor and computer technology,sensor-based human activity recognition technology has been widely used in various fields,such as smart home,medical care,motion detection and so on.This technology has attracted more and more researchers’ attention.At present,it is not only an important research content of time series classification,but also an important research direction in the field of machine learning and human-computer interaction.However,the original data collected by the sensor has no label,and due to the poor readability of the original sensor data,manual labeling of data not only requires a lot of time,but also requires the tagger to have professional knowledge.Therefore,the lack of labeled data is an important reason for limiting the learning ability of human motion recognition model.On the premise of insufficient labeling of human motion data,this paper studies sensor-based human activity recognition,so that the activity recognition model can still accurately distinguish activity in the case of less labeled data.The main research contents are as follows:(1)Aiming at the problem that the features required to recognize different human actions are complex and changeable,and the distribution of labeled data features and unlabeled data features is different,a semi-supervised human activity recognition model robust to feature distribution is proposed.Firstly,the hybrid model of 2D convolution network and long-term and short-term memory network is used to learn the local and overall features of the original sensor data,so as to improve the learning ability of the model to the different scale features of the input sequence.Then,the feature classifier module is added,the gradient inversion layer is used to align the feature distribution of labeled data and unlabeled data,and the focal loss function is used to balance the loss proportion of the two kinds of data with large difference in the number of samples.(2)This paper presents a semi-supervised human activity recognition model based on SMATE network.Firstly,based on the general time series classification model SMATE,by cutting the cumbersome clustering module and adding the classification module of labeled data,the network structure is simplified and the use efficiency of the original data is improved,and calculate the spatio-temporal feature weight to fuse the two kind features.Secondly,the structure of the model time encoder is changed,and the dual stream long short-term memory network is used to learn the internal time dependence of time series data from both positive and negative directions.Then,according to the characteristics of human activity recognition data,add the channel attention module and change the spatio-temporal feature fusion mode,so that the model can give different attention to the spatio-temporal features of different channels and the time-domain features or spatio features themselves,so as to enhance the discrimination ability of the model to important features.Finally,the experimental analysis is carried out on the public human activity recognition datasets.The experimental results show that the proposed method can still achieve satisfactory results even if less than 10% of the labeled data is used for training.