SMDE: Unsupervised representation learning for time series based on signal mode decomposition and ensemble

Abstract

Time series data pervades multiple domains, yet in real-world applications, there frequently exists a deficiency of labels to ascertain effective representations and enable efficient classification. In this work, we introduce SMDE, a unique time series contrastive learning framework, diverging from conventional techniques that predominantly concentrate on instance-level contrast. SMDE leverages signal decomposition and ensemble methodologies, enhancing our focus on localized mode alterations, within time series signals. We further introduce the global signal consistency and intrinsic mode consistency mechanisms to promote the robust learning of global and local attributes. To ensure more efficient convergence, we present the DE Circle loss furnished with an adaptive weighting mechanism, founded on global signal consistency and intrinsic mode consistency principles. The SMDE framework is designed to be applicable to both univariate and multivariate time series analysis, indicating its potential for a wide range of temporal data applications. Experiments on 128 univariate UCR datasets reveal that SMDE improves the average accuracy by 4.73% over existing techniques. Additionally, our experiments on 29 multivariate UEA datasets demonstrate an average accuracy improvement of 7.42% compared to current methodologies. Moreover, through extensive semi-supervised experimentation, SMDE demonstrates superior performance in 15 disparate semi-supervised configurations, attaining an equivalent performance level to supervised training with 100% labeled data, even with a minimal portion of the labels. The code is publicly available at https://github.com/haobinlaosi/smde.