Time Series Anomaly Detection Based on Score Generative Model

To solve the shortcomings of traditional time series anomaly detection models in the representative stochastic variables of data and poor model generalization ability, a score-based generative model was proposed. For time-series data monitoring in complex cyber-physical systems, a multidimensional time series anomaly detection framework is designed. This framework utilizes regression models to capture the inherent temporal patterns within the data. Considering the randomness of the time series generation process, we employed a denoising score matching method to estimate gradient information. Then, 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 dataset and secure water treatment dataset showed that the proposed method can achieve F1 score of 96% and 90. 18% respectively, by more than 1. 02% and 1. 01% respectively higher than the highest F1 scores obtained via baseline models. The results of the ablation experiments and case analyses demonstrate that the noise index module and the signature matrix module can improve the model蒺s capability of feature extraction, and the proposed model achieves an F1 score above 0. 8 for anomaly thresholds within the range of [0. 386, 0. 8). © 2024 Beijing University of Posts and Telecommunications. All rights reserved.