Time Series Classification Based on Multi-Dimensional Feature Fusion

Time series classification is a key problem in data mining, most of existing classification methods directly extract one-dimensional data from one-dimensional features, which cannot effectively express the inter-relation between different time points. Besides, some classification methods extract two-dimensional features through encoding raw one-dimensional data into two-dimensional images, and part of information is lost due to the difference of encoding methods. How to make full use of one-dimensional and two-dimensional features to extract valuable information and integrate them in an optimal fashion remains a promising challenge. In this paper, we propose a multi-scale convolutional network to extract one-dimensional features from time series for obtaining more feature information based on multi-scale convolution kernels. Two-dimensional features are constructed in terms of two-dimensional image coding based on ${G}$ ramian angular field, ${M}$ arkov transition field and ${R}$ ecurrence plot (GMR) methods. We develop a multi-dimensional feature fusion approach leveraging ${S}$ queeze-and- ${E}$ xcitation (SE) and ${S}$ elf- ${A}$ ttention (SA) mechanism to effective fusing one-dimensional multi-scale features and two-dimensional image features in terms of weight setting. We conduct experimental verification based on 84 complete data traces from a typical UCR dataset in the field. Experimental results show that the accuracy of our proposed approach improves by 3.35% compared with existing benchmark methods. The Grad ient-weighted ${C}$ lass ${A}$ ctivation ${M}$ apping (Grad-CAM) visualization analysis method is adopted, where our proposed approach extracts more accurate features and effectively distinguishes different time series data categories.