Abstract: To solve the problems such as the difficulty to learn the representative stochastic variables of data and poor generalization ability for traditional time series anomaly detection methods, a score-based generative model was proposed. To detect anomalies for time series data in complex Cyber-Physical Systems, a regression model based anomaly detection method was devised to capture the intrinsic temporal pattern of the multivariate time series data. Considering the stochastic of the time series generation process, the gradient information was estimated based on the denoising score matching method. Using the estimated gradient information, an efficient anomaly scoring method was devised to improve the accuracy of the time series anomaly detection task. Experiments on Pooled Server Metrics (PSM) dataset and Secure Water Treatment (SWaT) dataset showed that the proposed method can achieve F1 score of 96% and 90.18% respectively, boosting accuracy by more than 1.02% and 1.01% compared than best baseline. Furthermore, the ablation experiments and case study proved the effectiveness of each module of the proposed method.

Key words: Cyber Physical System, time series, anomaly detection, score-based generative model