Abstract: Existing time series contrastive learning methods have some problems,such as augmented sample construction methods rely too much on manual experience,insufficient generalization ability,positive samples are not defined in a general enough way,and coarse-grained representations of contrastive measures,resulting in weak overall time series representation.Therefore,a multi-granularity time series contrastive learning method based on time-frequency features(TSDC) is proposed.The seasonal-trend generation network generates temporal augmentation samples with stable variations in the time domain,and the multi-band fusion perturbation operation generates non-stable variations temporal augmentation samples in the frequency domain,and the two augmentation samples are learned through coarse-grained contrastive at the instance level and fine-grained contrastive at the dimension level,so that the model can be better adapted to different types of downstream time series tasks while obtaining better representation.Experiments on classification,prediction,and anomaly detection on multiple time series public datasets show that the representation obtained by the TSDC method outperforms typical baseline models for downstream tasks.

Key words: Time series, Representation learning, Contrastive learning, Data augmentation, Multi-granularity contrastive