一种融合降噪自编码和对比学习的时间序列表示学习框架

In time series classification tasks, the lack of labeled data presents a challenge to traditional supervised deep learning methods, and self supervised representation learning can fully utilize unlabeled data. However, self-supervised representation learning for time series has not been fully studied, and existing methods have at least one of the following three shortcomings: high representation dimension, insufficient linear separability, and inadequate accuracy in downstream classification tasks. We propose a novel time series representation learning framework that integrates denoising auto-encoding with contrastive learning to address these issues. The proposed method fuses the input reconstruction task of denoising autoencoders and the contrastive learning task based on TWIST loss into a unified framework which can jointly learn low-dimensional and highly linearly separable representations of time series. The effectiveness of the proposed method has been validated through experimental results obtained from real datasets across three distinct fields. Specifically, compared to the baseline method, the proposed method demonstrated an accuracy improvement ranging from 1.19% to 5.27% and a macroscopic F1 score improvement of 1.41% to 2.07% in linear classification tasks. In semi supervised classification tasks, even with only 10% labeled data, our method exhibits performance comparable to supervised models trained with all labeled data.