Time Series Classification Based on Multi-scale Dynamic Convolutional Features and Distance Features

Abstract

Time series classification is one of the most important and challenging problems in the field of data mining. This paper presents a time series classification model named FusionNet, which is based on multi-scale dynamic convolutional features and distance features. The main contribution of this paper includes three aspects: firstly, based on the multi-scale dynamic convolution operation, we propose a Multi-Scale Dynamic Convolution Network (MSDCNet). It uses multi-scale dynamic convolution to dynamically adjust the convolutional kernels for different input data and extract the features of time series. Secondly, by calculating the distance between the prototypes and the embedding vectors, we construct a Prototype Network (PrototypeNet) to extract the distance features of time series. At the same time, we design a distance loss to ensure to calculate the effective distance features. Finally, we fuse the multi-scale dynamic convolution features with the distance features to obtain the fusing features for classification. Experimental results on 44 UCR datasets show that the proposed FusionNet achieves better results on multiple datasets than the previous model, demonstrating its effectiveness.