Time Series Classification Method Based on Multi-Scale Convolution with LSTM
Guangyuan Xu, Weiyuan Sun, Hanhan Xue, Xianzhou Feng · 2023
To solve the problem that both local and temporal features of time series can affect the classification accuracy, a Multi-scale Convolutional Neural Network with LSTM (LSTM-MCNN) is proposed. Firstly, in the full convolution part of the model, multi-size convolutions are used to extract local features of different scales from the input time series signals. Secondly, the SE (Squeeze-and-excitation) module was added after the convolutional layer to reduce the influence of useless channel information, so that valuable features of different sizes could be extracted. Finally, the local features obtained are combined with the temporal features extracted by the long short-term memory network to further mine the temporal feature information. The experimental results show that the accuracy of the proposed method on multiple standard time series datasets is higher than that of the compared model, which also proves that the proposed method has versatility and accuracy in the time series classification task.