Research On Time Series Classification Method Based On Convolutional Neural Network

With the rapid growth of time series data obtained by sensors,time series classification(TSC)has become an essential task in many real-world applications.TSC tasks can be divided into two categories: non-real-time TSC and real-time TSC.No matter what kind of task it is,convolutional neural networks are widely used for their excellent classification performance.The features of time series are naturally multi-scale.Therefore,and it is very important to select appropriate multi-scale features for classification.However,the current TSC methods based on convolutional neural networks tend to ignore the multi-scale problem of time series,and the performance still needs to be improved.This thesis aims to build convolutional neural networks capable of extracting multi-scale features from time series to overcome these shortcomings.Firstly,the basic knowledge of TSC and convolutional neural networks is introduced.On this basis,the WDM-Res Net(Wavelet decomposition matrix and Res Net)method are proposed for non-real-time TSC tasks by combining traditional wavelet decomposition with convolutional neural networks.In addition,a parallel-feature extraction(PFE)module and a feature attention(FA)module are introduced to propose a VIMSNet(Visually induced motion sickness network)for real-time TSC tasks.Finally,the thesis evaluates WDM-Res Net and VIMSNet.The main research results of this thesis include:(1)A non-real-time time series classification method based on WDM-Res Net.The characteristic of this method is that the time series is converted into images by wavelet decomposition matrix(WDM).The generated images contain multi-scale temporal and frequency domain information hidden in the time series.In addition,in order to make full use of the label information of the time series,the method imposes the similarity constraint on the images generated based on WDM,so that the images of the same category are close to each other,and those of different categories are far away from each other,which is more conducive to classification.(2)A real-time time series classification method based on VIMSNet.The method introduces a PFE module into the model,which uses multiple parallel convolutional blocks for feature learning,not only to capture multi-scale features,but also to meet the demands of real-time classification tasks.In addition,the method constructs a FA module that selectively enhances discriminative features while suppressing less useful ones,hence improving classification accuracy.On the actually collected EEG(Electroencephalogram)time series dataset,VIMSNet achieves the best real-time classification performance of visually induced motion sickness(VIMS).Experiments show that WDM-Res Net improves the accuracy of non-real-time TSC on UCR datasets by acquiring multi-scale time domain and frequency domain information through imaging representation;VIMSNet improves real-time classification performance on real EEG time-series datasets by capturing multi-scale features.The research results in this thesis provide new ideas for TSC problems based on convolutional neural networks.