A Time Series Classification Algorithm Based On Multi-scale Separable Convolution And Self-attention

Time series classification is widely used in various fields such as behavior recognition,disease diagnosis,speech recognition,environmental monitoring,and industrial production.The difference between time series classification and general classification tasks lies in the fact that its features are ordered.However,current time series classification algorithms have certain limitations: On the one hand,traditional algorithms have high computational complexity,require laborious feature engineering,and demand researchers to possess domain-specific background knowledge.On the other hand,convolution-based deep learning algorithms are limited by their receptive field,which can only extract local features of the sequence and fail to capture the interdependent relationships between all time positions globally.The accuracy and generalization of these models still remain some space for improvement.To address these issues,this thesis proposes an end-to-end neural network structure that simultaneously extracts local information of the sequence at multiple scales and global temporal correlations,achieved mainly through a plug-in module called MSCMA.The module consists of two parallel branches: one branch uses depthwise separable convolution to improve the Inception Time module for extracting sub-sequence information at multiple scales;the other branch uses multihead self-attention with added positional encoding to extract global temporal correlations of the sequence.This paper conducts experiments on six single/multivariate time series datasets from UCR/UEA with seven comparative algorithms,and uses Friedman rank test and Hollander-Wolfe post hoc test to compare the model accuracy,verifying the significant improvement of the MSCMA model.This paper also provides a comparison of algorithm complexity,with the time complexity of the MSCMA module being (9)8)(9(8),which is competitive among the compared algorithms.Finally,this thesis addresses the weak interpretability of deep learning algorithms by conducting multidimensional interpretability research on Lightning2 data set.First,it uses class activation mapping to identify the subsequence that contributes most to the classification decision of the model;second,it uses Gramian angular field to encode one-dimensional convolutional kernel as an image and visualize the multi-scale convolutional kernel weights of the network to study their capture of different patterns of sequence information;finally,it calculates and visualizes the attention weight matrix to analyze the temporal correlations of different patterns of sequence information extracted by the multi-head self-attention.The interpretability analysis from these three perspectives all verify the effectiveness of the proposed structure in this thesis.