Dependency Graph Learning for Anomaly Detection in Dynamic Smart-Manufacturing Sensor Data

Abstract

Smart manufacturing systems generate multivariate sensor streams from robotic arms, CNC machines, conveyor units, and environmental controllers. These signals are often non-stationary because of tool wear, production switching, maintenance operations, and workload changes. Conventional anomaly detection methods usually assume stable temporal patterns and may fail when normal operating conditions drift over time. This study proposes an adaptive dependency graph learning model for non-stationary multivariate sensor anomaly detection in smart manufacturing systems. The proposed method first constructs dynamic sensor-dependency graphs using sliding-window correlation and attention-based temporal encoding. A drift-aware normalization module is then used to separate long-term operating changes from short-term abnormal fluctuations. Finally, a graph-temporal reconstruction network estimates normal sensor behavior and detects anomalies through adaptive residual thresholds. Experiments are conducted on a smart manufacturing dataset collected from 42 production lines, 318 industrial sensors, and 11 machine categories over 96 operating days. The dataset contains 74.6 million timestamped sensor records, including normal production cycles, tool degradation, cooling-system instability, abnormal vibration, and actuator faults. The proposed method reduces average detection delay from 18.7 minutes to 6.4 minutes compared with a static LSTM autoencoder. The false alarm rate decreases to 1.72 events per 1,000 operating hours, and the Matthews correlation coefficient reaches 0.917 under mixed production modes. The model processes 12,800 sensor points per second on an edge workstation, with a memory footprint of 1.36 GB during online inference. Ablation analysis shows that removing dynamic dependency graphs increases detection delay by 9.2 minutes, while removing drift-aware normalization causes frequent false alarms after production switching. The results demonstrate that adaptive spatiotemporal dependency modeling can improve anomaly detection under non-stationary industrial sensor conditions.