Research On Human Activity Recognition Algorithms Based On Time Series Data From Sensors

At present,researches based on human activity recognition(HAR)have achieved many results and been used in various aspects such as industry,medical and health,sports competition,safety certification,human-computer interaction,intelligent applications and so on.It has important theories value and practical application value.This paper mainly focuses on the human activity recognition in sensors,and discusses it from the perspective of Time Series Data(TSD)to improve the performance and accuracy of the algorithm.The study of Human Activity Recognition based on Time Series Data from sensors generally extracts feature values and uses classifiers to complete activity recognition.Therefore,the selection of algorithm for data feature extraction and recognition are the important questions in this field.However,the existing feature extraction methods and recognition methods are still inadequate.So this paper will propose the following two improved models:First,an improved data segmentation and feature extraction algorithm named BU_DSW(Bottom-Up Dynamic Sliding Window algorithm)is proposed.For time series data,data segmentation is important to extract effective features.Compared to the Sliding Window Algorithm,which is most commonly used in Human Activity Recognition,the new BU_DSW algorithm combines the bottom-up algorithm BU and the dynamic sliding window algorithm DSW(Dynamic Sliding Window Algorithm)to dynamically segment time series data according to the distribution law,instead of blindly dividing the data Split into equallength subsequences.At the same time of segmentation,the time domain features and frequency domain features of the segmented subsequence are all extracted using window sliding,which are fused with the features extracted by the classifier to obtain a set of composite feature vectors.Secondly,this paper proposes an improved multi-scale convolutional neural network algorithm named MCNN,which combines the advantages of the CNN model and the Inception structure,and can extract a greater number and variety of data features and improve recognition accuracy.The human activity recognition data collected from wearable sensors are usually discrete time-series signal data,and people's daily behavior activities are diverse.This makes the activity recognition data very complex,and it is difficult for traditional CNN models to achieve good generalization and robustness.The multi-scale mining models can get more abundant information.In the experimental parts,this paper uses the UCI HAR dataset and WISDM dataset to compare with Convolutional Neural Network(CNN),Recurrent Neural Network(RNN),Long Short-Term Memory(LSTM),Bi-directional Long Short-Term Memory(BLSTM),Multi-Layer Perceptron(MLP)algorithms,and verify the algorithm by using accuracy rate,recall rate,loss rate,F1 score,etc.The final result shows that the two improved schemes in this paper can improve the algorithm effect.