Early Time Series Classification Using TCN-Transformer
Huiling Chen, Aosheng Tian, Ye Zhang, Yuzi Liu
Abstract
Early time series classification is of great significance for time-sensitive applications such as fault detection and earthquake prediction. This task aims to classify time series with the least timestamps at desired accuracy. Recent deep learning methods usually used the Recurrent Neural Networks (RNNs) as the classification backbone and the exiting subnet for early quitting. However, the RNNs suffer from the ‘forgetting’ defect and insufficient local feature extraction. Besides, the balance between earliness and accuracy is not fully considered. In this paper, a framework named TCN-Transformer is proposed. To overcome the defects of RNNs, we combined Temporal Convolutional Network and Transformer to extract both local and global features. Then, a loss function is designed to ensure the classification performance, while focusing more on earlier features. The experimental results on ten univariate datasets.