An Input Module of Deep Learning for the Analysis of Time Series with Unequal Length

Abstract

Deep learning, particularly deep neural networks, has received increasing interest in time series classification, and several deep learning methods have been proposed recently. However, most of these algorithms are designed for time series with equal length, while clustering of time series with unequal length is often encountered in real-world problems. This paper proposes an input module of deep learning, transforming time series with unequal length into a warping matrix processed by neural network for training. The trajectory warping matrix is generated by DTW algorithm according to the similarity difference of time series. The Gaussian blur iterative algorithm is introduced to converted from the warping matrix of any size to equal dimension. The effectiveness of the proposed input module combined with some advanced neural networks are assessed based on the CWRU dataset. Overall, the analysis shows that the input module assists the depth learning to classify time series with unequal length accurately.