MSAnomaly: Time Series Anomaly Detection with Multi-scale Augmentation and Fusion

Tao Yin, Zhibin Zhang, Shikang Hou, Huan Zhao, Lijiao Zheng, Yifei Zhou, Jin Xie, Meng Yan

Abstract: In real-world industrial scenarios, unsupervised time series anomaly detection for massive multi-dimensional sensor data is a pressing research topic. While existing unsupervised anomaly detection methods have achieved significant progress in anomaly detection performance, they still face two limitations. First, existing methods primarily model and analyze data on a single time scale, ignoring the rich dependencies between features at different time scales. Second, traditional methods struggle to capture features across different time scales, failing to represent the temporal structure of the data comprehensively. To address these challenges, based on a multi-scale data augmentation approach and multi-scale Fusion block, we propose an unsupervised anomaly detection model, MSAnomaly, to improve the ability to learn sequential patterns at various time scales. Specifically, MSAnomaly transforms the original sequence into multiple time-scale sequences using a multi-scale data augmentation approach, fusing different time resolution features by multi-scale fusion block to model the time series effectively. Our proposed MSAnomaly enables effective model training with limited data. Extensive experiments demonstrate that MSAnomaly achieves state-of-the-art performance across multiple real-world benchmark datasets for anomaly detection.