FDNet: A Novel Multivariate Time Series Classification Model Through Fusing Feature and Difference

Abstract

The classification problem of multivariate time series has been widely used in many fields, but current methods still cannot achieve high accuracy. In this paper, a novel multivariate time series classification model named FDNet (Feature and Difference encoding fused Network) is proposed. A structure that comprehensively considers global feature extraction and local multi-scale feature extraction is used in the feature encoding, which can better represent the closeness between the sequence and true label. In the difference encoding, distance difference and shape difference are combined to further enhance the FDNet’s discriminative power. Also, a representative sequence selection algorithm is investigated to speed up the calculation of difference encoding, as well as a k-fold filtering algorithm based on density masking to enhance the accuracy of shape difference. In the comparative experiments on 23 public datasets, FDNet outperforms other baseline methods and state-of-the-art methods in terms of average ranking, average accuracy, and the number of wins/ties, which verifies FDNet has high accuracy and good generalization.