Self-supervised Representation Learning for Time Series via Temporal Contrasting and Transformation
Qirong Chen, Menglei Zhao, Zhenguo Zhang · 2023
Learning representations in widely available time series is of great significance, and representations in time series often represent some characteristic of this phase and play an irreplaceable role in fault detection in real industrial tasks, prediction in financial models, and diagnosis in medical data, while learning appropriate representations from unlabeled time series data is a challenging task. Most of the current methods are supervised learning methods, however, the labelling of data is time-consuming and impractical, and how to obtain effective data representations from unlabeled time-series data has great application value, based on this we propose an unsupervised time series representation learning method via Temporal Contrasting and Transformation (TS-TCT) to learn sequence representations from data lacking labelling. First, the time series is transformed into two homology sequences by applying mixed Augmentation and fast Fourier transform. Secondly, we use a cross-prediction based approach to enhance the time representation learning effect. Finally, to further enhance the learning effect, we employ a contextual comparison method based on the temporal comparison. The aim is to maximize the similarity of the same sample contexts while reducing the similarity between different sample contexts. The performance on two time series datasets shows that the unsupervised learning method based on TST-CT has some performance improvement over the existing unsupervised learning methods, and it is worth mentioning that our method has comparable performance with the supervised method and shows excellent results on some datasets.