Contrastive Learning for Time Series Classification with Stronger Augmentations

How to efficiently represent time series data through representation learning is of great importance, as it enables the automatic extraction of features without extensive manual labeling. This capability is crucial in various fields and is particularly useful in situations with high-dimensional and complex data. In this paper, we propose a novel contrastive learning framework for time series classification, which leverages both strong and weak augmentation strategies to enhance the learning process. We combine time domain and frequency domain augmentation techniques to form a strong augmentation strategy, allowing our model to capture more comprehensive features from time series data. We use a one-dimensional convolutional neural network (CNN) as the backbone for the encoder, our model achieved an accuracy of 83.81% on the Sleep-EDF dataset and 97.73% on the Epilepsy dataset. These results demonstrate the effectiveness of combining strong and weak augmentations for contrastive learning in time series classification tasks.