Gating Mechanism Based Feature Fusion Networks for Time Series Classification

With the development of digitization, more and more sensors or monitoring tools are used in all walks of life. These devices generate a large amount of time series data every day. It is of great significance to process these data in a timely and efficient manner. Deep neural networks have been proven to perform good feature extraction and modeling for time series, and good results have been achieved in time series classification problems through stacking of basic modules such as multilayer perceptrons, fully convolutional networks, and deep residual networks. However, the generalization of these methods is poor, and the extraction of pattern features is limited. At the same time, because these modules cannot extract temporal correlation features well, there is still room for improvement in classification accuracy. Based on the dual consideration of temporal correlation features and pattern features, we designed a time series classification method based on two-channel feature extraction based on gating mechanism fusion. Through extensive experiments on multiple datasets from different domains, our classification method proposed in this paper achieves better performance, and has significant advantages in classification accuracy and generalization.