Physically-informed Bayesian feature optimization for semi-supervised industrial anomaly detection

Abstract In tackling frequent batch anomalies in manufacturing processes, this study introduces a novel semi-supervised in-process monitoring method that enables real-time anomaly detection during production rather than post-process inspection. By integrating physically-informed Bayesian feature optimization with semi-supervised learning, the proposed method captures batch defects under conditions of scarce anomaly samples. First, a Bayesian feature optimization strategy is proposed to extract critical indicators from redundant industrial signals, effectively bridging the gap between domain expertise and data-driven representations. Second, to address the semi-supervised challenge of scarce anomaly samples, an out-of-distribution detection model trained on normal samples is established, combined with conformal calibration to determine robust decision thresholds. The superiority of the proposed method is validated with real-world stamping and injection molding datasets, with ablation studies confirming robustness to hyperparameter variations. The results demonstrate the potential of the proposed method for anomaly monitoring in data-scarce industrial contexts.