Spatio-Temporal Graph Neural Networks for Anomaly Detection in Complex Industrial Processes.

Area of Science:

• Intelligent Manufacturing
• Cyber-Physical Production Systems (CPPSs)
• Process Monitoring

Background:

• CPPSs generate massive, complex data challenging real-time monitoring.
• Existing anomaly detection methods struggle with spatio-temporal dynamics, computational cost, and incipient fault detection.

Purpose of the Study:

• To propose a robust anomaly detection method for CPPSs.
• To address limitations of current methods in handling complex data and detecting early faults.

Main Methods:

• Developed a Spatio-Temporal Variational Graph Statistical Attention Autoencoder (ST-VGSAE).
• Employed an Adaptive Lifting Wavelet Module for multi-scale temporal decomposition and noise suppression.
• Integrated a spatio-temporal Token statistical self-attention mechanism for reduced computational cost and enhanced anomaly discriminability.

Main Results:

• The ST-VGSAE model demonstrated superior performance on the Tennessee Eastman (TE) process dataset.
• Achieved significant improvements in Fault Detection Rate and False Alarm Rate compared to state-of-the-art methods.
• Exhibited enhanced noise robustness and real-time processing capabilities.

Conclusions:

• The proposed ST-VGSAE effectively addresses challenges in real-time process monitoring for CPPSs.
• The model offers a robust and computationally efficient solution for anomaly detection.
• Validated effectiveness through experiments on a benchmark industrial process dataset.