Self-Supervised Multi-Transformation Learning for Time Series Anomaly Detection

Time series anomaly detection aims to find specific patterns in time series that do not conform to general rules, which is one of the important research directions in machine learning. However, most existing time series anomaly detection methods solely consider the original observed state space. In this paper, we investigate the different views of the original signals by reconstructing them from the different transformations to learn better comprehensive representations of normal patterns. Specifically, we propose a time series anomaly detection model based on self-supervised multi-transformation learning while jointly learning the noise and filter transformation of the normal time series and capturing the anomaly simultaneously in both transformation patterns. Firstly, the model randomly adds noise to the original signal to construct the noisy signal, and the Kalman filter is applied to the original signal to construct the filtered signal. The noisy and filtered signals are subsequently used to reconstruct the original signal, and the randomly added noise is classified. Finally, the model captures the two transformation patterns of normal time series signals and transforms the anomaly detection problem into the reconstruction metric problem of time series signals under noise and filter transformations. The effectiveness of the proposed method is verified on the UCR time series dataset.