Adaptive Multivariate Time-Series Anomaly Detection

Abstract

Existing Multivariate time-series anomaly detection methods aim to calculate the anomaly scores of observed sequences and learn a threshold to judge whether the input data is abnormal. However, they neglected the temporal covariate shift problem, which leads to the learned thresholds cannot be generalized in the test set, resulting in suboptimal detection performance in practical cases. We propose the Adaptive Multivariate Time-series Anomaly Detection framework in this paper, namely DATECT, to address the above challenging problems. Specifically, to enhance the robustness of anomaly measurement, DATECT adopts the dilated convolution based AutoEncoder to integrate both prediction errors and reconstruction errors into the output anomaly scores. Meanwhile, a novel Adaptive Window Normalization method is put forth to reduce the diversity of the distribution of anomaly scores in the test set, hence effectively improving the generalization capability of the detection model. Finally, to further reduce the side-effect of domain-specific dynamic noise, DATECT utilizes Non-parametric Scan Statistics to select the subsets of significantly abnormal signals and highlight the anomaly segments. Experiments on five datasets show that our method can significantly alleviate the performance drop caused by the temporal covariate shift problem, outperforms the baseline in terms of detection performance and generalization, averagely improving the F1-score by 8.66% and the F1∗-score (upper bound) by 1.18%.