A two-stage adversarial Transformer based approach for multivariate industrial time series anomaly detection

Sensors in complex industrial systems generate multivariate time series data, frequently leading to diverse abnormal patterns that pose challenges for detection. The existing multivariate abnormal detection methods may encounter difficulties when applied to datasets with low dimensions or sparse relationships between variables. To address these issues, this study proposes a two-stage adversarial Transformer-based anomaly detection method. On the one hand, an autoregressive temporal convolutional network component is embedded before the multi-head attention module to capture features encompassing long-term and local information. Besides, this component utilizes a trainable neural network instead of the vanilla Transformer's absolute position encoding, resulting in enhanced position information. On the other hand, the proposed two-stage adversarial learning strategy allows the model to effectively learn intricate multivariate data patterns via constraining latent space, thereby enhancing anomaly detection performance. Our method achieves F1 scores of 0.9679, 0.7947, and 0.6452 on a real-world dataset and two public industrial sensor datasets, demonstrating superior overall anomaly detection performance compared to recent advanced works.