Research On The Modeling Method For Multivariate Time Series Classification And Recognition Of Wearable Health Monitoring

In recent years,wearable devices have been rapidly developed due to their many advantages,and they are widely used in the field of health monitoring,where time series are the main form of data.However,a large number of redundant,irrelevant and even noisy features in multivariate time series data directly affect the performance of the classification model,so feature extraction and effective use is an important part of multivariate time series classification and recognition modelling.The paper focuses on wearable health monitoring multivariate time series and addresses key issues in classification and recognition modelling.It discusses the selection of effective region segments of time series for feature extraction.It also covers the utilization of temporal and spatial features,as well as interactions among multivariate time series features.The goal is to fully explore useful information in the features.The main research content includes:Firstly,this paper proposes a method for effectively selecting feature extraction region segments in multivariate time series classification and recognition modelling.The proposed method is based on class activation mapping and multi-channel convolutional neural network importance score calculation.The multi-channel convolutional neural network extracts features from the input data of multiple sensors.The class activation mapping method is used to calculate importance scores of time series at different moments for classification.Based on the importance scores,representative key features are extracted by selectively targeting feature extraction region segments from a large amount of data.Using the Sis Fall dataset as a validation example for wearable health monitoring,we selected feature extraction region segments based on importance scores.These segments were then summarised to create a segmentation strategy for inertial sensor data.This strategy is applied when the acceleration SMV is less than 9.217 m/s2 and the window length is 325 ms.The experimental results from various machine learning models demonstrate that the proposed method optimises the selection of effective feature extraction region segments for multivariate time series.Secondly,to address the issue of spatial-temporal feature fusion in multivariate time series classification modelling,we propose a method that utilises multivariate time series spatial-temporal feature superposition.This method combines features from long short-term memory neural networks and one-dimensional convolutional neural networks.The temporal and spatial features of physiological signals are extracted using the long and short-term memory neural network and the one-dimensional convolutional neural network,respectively.The fusion features are then obtained through Tanh activation function mapping and spatial-temporal feature superposition and summation.The WESAD dataset was used as a validation example for wearable health monitoring of stress and emotion.The binary and tertiary classification achieved an accuracy and F1 score of 94.9% and 94.98%,and 87.82% and 86.68%,respectively.The experimental results demonstrate that the proposed method of spatio-temporal feature fusion addresses the issue of spatio-temporal feature utilization in existing classification modelling methods based on long and short-term memory neural networks and convolutional neural networks.Thirdly,this paper proposes a method for addressing the issues of redundant or non-important features,as well as analyzing and exploiting relationships among multivariate time series features in classification and identification modelling.The proposed method integrates a gating mechanism,long short-term memory network,and graph attention neural network to analyze and exploit these relationships.The utilisation of the neural network for long short-term memory enables the extraction of temporal features from each variable of the multivariate time series,which are then used to construct the graph neural network as node features.The graph attention neural network assigns different aggregation weights to different nodes,and a gating mechanism is combined to further enhance the feature extraction effect,retaining important features and excluding redundant or irrelevant ones.The study used the WESAD dataset to validate wearable health monitoring for stress and affection.It found that electrical skin activity,and respiration sensors are crucial for the psychological stress classification and recognition model,based on the degree centrality of the nodes.The experimental results demonstrate that the proposed method validates the effectiveness of extracting features from multivariate time series and analyzing their inter-relationships based on graph structure.It also explains the roles played by each sensor in completing the classification task,provides a basis for sensor selection,and addresses the problem of utilizing the analysis of relationships between multivariate time series features in multi-sensor systems.This paper proposes solutions for three key issues in multivariate time series modelling: selecting effective region segments for feature extraction,using spatiotemporal feature fusion,and utilizing relationships between multivariate time series features.These solutions have the potential to be applied in the field of wearable health monitoring for multivariate time series classification modelling and provide relevant support to the development of wearable health monitoring systems.Algorithmic support can assist in developing wearable health monitoring systems,contributing to the advancement of intelligent medical systems.