Multivariate Time Series Anomaly Detection with Adaptive Transformer-CNN Architecture Fusing Adversarial Training
Junpeng He, Zhe Dong, Yaqing Huang · 2024
In the era of IoT, efficient real-time anomaly detection is critical for preventing industrial incidents and ensuring operational integrity. This paper introduces ATC-AFAT (Adaptive Transformer-CNN Architecture Fusing Adversarial Training), a novel unsupervised anomaly detection framework for multivariate time-series data, leveraging an attention-based sequence encoder with adversarial training and CNN discriminate model to tackle the challenges posed by the diverse and nonlinear nature of sensor data. Our method demonstrates significant improvements over existing approaches, reducing training time by 76.5% and enhancing interpretability, as validated on four diverse datasets. ATC-AFAT emerges as a robust solution for real-time anomaly detection in complex IoT environments.