GenAD: General Representations of Multivariate Time Series for Anomaly Detection

Xiaolei Hua, Su Wang, Lin Zhu, Dong Zhou, Junlan Feng, Yiting Wang, Chao Deng, Shuo Wang, Mingtao Mei

Keywords: Anomaly Detection, Multivariate Time Series, General Representations

Abstract: Anomaly Detection(AD) for multivariate time series is an active area in machine learning, with critical applications in Information Technology system management, Spacecraft Health monitoring, Multi-Robot Systems detection, etc.. However, due to complex correlations and various temporal patterns of large-scale multivariate time series, a general unsupervised anomaly detection model with higher F1-score and Timeliness remains a challenging task. In this paper, We propose a General representations of multivariate time series for Anomaly Detection(GenAD). First, we apply Time-Series Attention to represent the various temporal patterns of each time series. Second, we employ Multi-Correlation Attention to represent the complex correlations of multivariate time series. With the above innovations, GenAD improves F1-scores of AD by 0.3% to 5% over state-of-the-art model in public datasets, while detecting anomalies more rapidly in anomaly segments. Moreover, we propose a general pre-training algorithm on large-scale multivariate time series, which can be easily transferred to a specific AD tasks with only a few fine-tuning steps. Extensive experiments show that GenAD is able to outperform state-of-the-art model with only 10% of the training data.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

One-sentence Summary: we propose GenAD, a General rep-resentations of multivariate time series for Anomaly Detection, and GenAD outperforms state-of-the-art model on public datasets with only 10% of the training data.

Reviewed Version (pdf): https://openreview.net/references/pdf?id=uHvf8naBtW