Multivariate Time Series Anomaly Detection in IIoT Using Integrated Transformer and Graph Neural Networks
Xinyi Yu
College of Engineering, Hangzhou Normal University, Hangzhou, Zhejiang, China.
Haoran Hu
College of Engineering, Hangzhou Normal University, Hangzhou, Zhejiang, China.
Jiaoping Wang
Jiande City Quality Measurement Monitoring Center, Hangzhou, Zhejiang, China.
Xinyao Wei
College of Engineering, Hangzhou Normal University, Hangzhou, Zhejiang, China.
Lidong Wang *
College of Engineering, Hangzhou Normal University, Hangzhou, Zhejiang, China.
*Author to whom correspondence should be addressed.
Abstract
In view of the high-dimensional, strongly coupled, and non-stationary characteristics of multivariate time series data in the Industrial Internet of Things (IIoT), as well as the limitations of traditional anomaly detection models (such as CNN-BiLSTM-SelfAttention) in capturing long-term temporal dependencies and modeling dynamic spatial correlations, this paper proposes an anomaly detection model (GTN) that integrates Transformer and Graph Neural Network (GNN), and enhances detection robustness by combining correlation analysis and extreme value theory. The specific steps are as follows: 1) Clean the raw data through the spectral residual algorithm to filter out random noise in the industrial environment; 2) Based on the cross-correlation function and dynamic time series graph construction, mine the lagged dependencies and spatial correlations between sensors, and divide the time series into related clusters to simplify the modeling complexity; 3) Improve GNN (introducing graph attention mechanism) to aggregate spatial features, and optimize Transformer (temporal subsequence self-attention + generative decoding) to capture long-term temporal dependencies; 4) Set adaptive thresholds based on prediction residuals and extreme value theory to achieve anomaly localization. Experiments on the SWaT and PSM industrial benchmark datasets show that the GTN model outperforms the traditional CNN-BiLSTM-SelfAttention model by 8.7% to 13.2% in F1-score and AUROC metrics, respectively. The recognition accuracy of equipment degradation and sensor coupling anomalies reaches 90.4%, and the false alarm rate is reduced by 11.5%, verifying the effectiveness and superiority of the model in industrial scenarios.
Keywords: Industrial internet of things, multivariate time series data, anomaly detection, transformer, graph neural network, correlation analysis.