Graph-Enhanced Spatiotemporal Learning for Early Anomaly Detection in Industrial Manufacturing Processes

Abstract

Early anomaly detection in industrial manufacturing processes is essential for improving production reliability, reducing operational disruptions, and maintaining consistent product quality. However, existing machine learning methods often struggle to capture complex dependencies among industrial sensors, dynamic process variations, and evolving abnormal operating patterns. This study proposes a graph-enhanced spatiotemporal learning framework for early anomaly detection in industrial manufacturing systems. The proposed framework integrates graph neural networks with temporal attention mechanisms to jointly model inter-sensor relationships and sequential process dynamics. An adaptive graph construction strategy is introduced to identify changing dependencies among process variables, while a multi-scale temporal feature extraction module captures both short-term fluctuations and long-term operational trends. To improve detection reliability, an uncertainty-aware anomaly scoring mechanism is incorporated to distinguish genuine process abnormalities from normal operating variations. The framework is designed for evaluation using the Tennessee Eastman Process (TEP) benchmark, which contains 52 monitored process variables and 21 predefined fault scenarios. Experimental comparisons will include Isolation Forest, LSTM, Autoencoder, and graph-based anomaly detection methods. Performance will be evaluated using F1-score, precision, recall, false alarm rate, and fault detection delay. Additional experiments will investigate robustness under sensor noise, missing measurements, and previously unseen fault patterns. The proposed framework aims to provide a scalable and interpretable anomaly detection solution for intelligent manufacturing and industrial process monitoring.