Towards Time-Series Key Points Detection Through Self-supervised Learning and Probability Compensation

Key points detection is crucial for signal analysis by marking the identification points of specific events. Deep learning methods have been introduced into key points detection tasks due to their significant representation learning ability. However, in contrast to common time series classification and prediction tasks, the target key points correspond to significantly different time-series patterns and account for an extremely small proportion in a whole sample. Consequently, existing end-to-end methods for key points detection encounter two major problems: specificity and sparsity. Thus, in this work, we address these issues by proposing a probability compensated self-supervised learning framework named ProCSS. Our ProCSS consists of two major components: 1) a pretext task module pretraining an encoder based on self-supervised learning to capture effective time-series representations with a higher generalization ability; 2) a joint loss function providing both dynamic focal adaptation and probability compensation by extreme value theory. Extensive experiments using both real-world and benchmark datasets are conducted. The results indicate that our method outperforms our rival methods for time-series key points detection.