Multivariate time series anomaly detection with adversarial transformer architecture in the Internet of Things

• Published in: Future generation computer systems 2023-07, Vol.144, p.244-255
• Main Authors: Zeng, Fanyu, Chen, Mengdong, Qian, Cheng, Wang, Yanyang, Zhou, Yijun, Tang, Wenzhong
• Format: Article
• Language: English
• Subjects: Anomaly detection; Internet of Things; Multivariate time series; Neural networks; Transformer; Unsupervised learning
• ISSN: 0167-739X; 1872-7115
• DOI: 10.1016/j.future.2023.02.015

Many real-world Internet of Things (IoT) systems contain various sensor devices. Operating the devices generates a large amount of multivariate time series data, which reflects the changing trends of the devices and the physical environment and provides data services for upstream applications. Given that the quality of IoT services usually depends on the accuracy and integrity of the data, we must guarantee the data’s accuracy. The massive sensor data can be continuously monitored to infer normal and abnormal behaviors through anomaly detection. Therefore, to ensure the stability of IoT infrastructure operation, anomaly detection of sensor data has high research value. In this paper, we propose a new multivariate time series anomaly detection structure that can effectively detect anomalies through an adversarial transformer structure. Additionally, the fused anomaly probability strategy can increase the discrimination between normal and abnormal; the reconstruction error of the first stage as the prior knowledge of the second stage can better detect anomalies. The evaluation experiments are conducted on four public datasets and achieve an average anomaly detection F1-Score higher than 0.89, which validates the effectiveness of our proposed method. •Adversarial two-stage training enhances data reconstruction for time series.•Fusion anomaly probability strategy amplifies the anomaly discrimination.•Reconstruction errors in the training phase can extract short-term trends better.