Time Series Classification via Enhanced Temporal Representation Learning

Abstract

Due to the booming of time series (a temporal data sequence, including continuous recorded values and timestamps), time series classification has been considered as one of the most challenging studies in time series data mining, attracting great interest from industry and academia over the last decades. However, the current time series classification methods have their own flaws. On the one hand, most existing methods focus on using traditional machine learning models for achieving classification accuracy, while ignoring the advantages of deep representation learning; on the other hand, the current deep neural network models only rely on one single deep learning model, and hence fail to improve performance effectively. In this paper, we propose an end-to-end representation learning model for time series classification. Concretely, we first utilize 1-D temporal convolution to obtain the feature representations. Secondly, we separately adopt the residual network and bidirectional long short-term memory network to achieve temporal representation reinforcement. Finally, we adopt the multi-layer perception network for final classification. Experimental results on open source benchmark datasets have justified the superiority of our model.