Research On Time Series Classification Based On Deep Learning

Time series classification is one of the most challenging tasks in data mining.Any classification problem that needs to consider the order of data can be regarded as a time series classification problem.In recent years,the scale and popularity of time series data have been increasing,which is widely existing in human life.Therefore,how to further improve the classification performance of time series and how to better achieve the automatic classification of time series have become an urgent problem.Traditional time series classification methods rely on the professional knowledge of researchers,which leads to tedious and inefficient mining process and easy to lose the potential information in time series.The concise and efficient end-to-end deep learning method focuses on the laws of the data itself,which can easily and accurately dig out the important information hidden in the time series,so as to provide reliable and powerful support for human decision-making activities.Aiming at the problem that the ResNet,the most advanced univariate time series classification model,has too many parameters and is easy to overfit.This paper proposes an improved ASPP?ResNet model with ASPP structure for the classification of univariate time series.ASPP structure captures different types of feature information in time series in a multi-scale way,the internal structure adopts the dilated convolution,which not only reduces the interference of noise data,but also enables the model to have a larger receptive field without increasing the number of parameters,so that more abundant feature information in the sequence can be obtained.This paper conducts a comparative experiment in the latest UCR ? UEA data archive with 128 datasets.The results show that the ASPP?ResNet has fewer parameters and better classification performance.Finally,the visual technology of class activation map is used to analyze the decision-making process of ASPP?ResNet model and ResNet model.The results show that both use very similar feature subsequences in the time series as the basis for classification.Aiming at the problems of insufficient correlation mining between variable dimensions of multivariate time series and insufficient feature extraction by existing methods,this paper proposes an ASPPs structure for extracting spatial features of multivariate time series.ASPPs are composed of two forms of ASPP modules: The first ASPP structure first regards multi-dimensional sequence data as a single-channel image data format,and then uses a two-dimensional filter on the entire feature map to extract the common feature information among all dimensions of the time series.The second ASPP structure uses a multi-scale branch structure with multiple one-dimensional convolution kernels to extract aggregated feature information from variable dimensions and sequence spatial features.In the process of extracting sequence spatial features,Squeeze-and-Excitation module is also used to adaptively learn the interdependence of the feature sequence in variable dimensions,the final classification model consists of LSTM network and ASPPs.Comparative experiments on multiple datasets show that the classification performance of LSTM?ASPPs is better than many excellent time series classification models,and it has a more reliable classification effect.Moreover,the results of multiple ablation experiments show that the dilated convolution kernel used in this paper greatly enhances the performance of the classification model.