Multivariate Time Series Data Transformation for Convolutional Neural Network
Chao-Lung Yang, Chen-Yi Yang, Zhixuan Chen, Nai‐Wei Lo · 2019
This paper proposes a novel framework to encode time series data into two-dimensional (2-D) images, and aggregate the images into one single image to solve multiple time series classification problem. In this research, Gramian Angular Summation Field (GASF) and Gramian Angular Difference Field (GADF) were applied to encode time series into images. The proposed image aggregation method which appends multiple images into a single image is suggested. After transformation and aggregation, the 2-D images passed through a convolutional neural network (CNN), which is outstanding in solving computer vision problems, for classification. An open wafer dataset was used to validate the proposed method. The preliminary results of the experiments find that encoding time series data into images and aggregating the images by the appending method are helpful on increasing classification accuracy. The statistical test also showed that the proposed image appending is “order-free” on the sequences of 2-D images.