Physics-Informed AI for Online Monitoring and Diagnosis of Precision Forming

In precision forming, batch anomalies are often invisible until final inspection — by which time entire lots have already been scrapped. This project writes forming mechanics into the algorithm: each production cycle signal is segmented according to the actual die motion phases, probabilistic priors tell the model which segments to watch, and a detector trained on normal parts only learns the boundary of "normal," classifying every part as pass, anomaly, or review-needed within the production takt time. Validated on two real production lines, the method achieves 99.32% and 94.92% accuracy under zero anomaly-sample training.

1. 1

   Segment Signals by Die Motion Phases

   Averaging over the full signal washes out localized anomalies, so the signal is first segmented by forming mechanics. Following the actual motion sequence of the die — clamping, pressurizing, forming, holding, and retraction — each cycle is split into several phases, and time- and frequency-domain statistics are extracted independently from each phase. Different forming processes have different phase structures, but the segment-and-extract approach is universal.

2. 2

   Encode Process Knowledge as Probabilistic Priors

   Experienced operators know which phase is most likely to go wrong, but that intuition is opaque to an algorithm. The solution assigns a probabilistic prior to every candidate feature: high-risk phases receive high prior weight, low-risk phases get low weight, and uncertain phases are left at a neutral value for the data to decide. The priors steer feature selection toward physically meaningful segments from the start, rather than relying on statistics alone.

3. 3

   Train on Normal Parts Only

   High-yield lines produce few defective parts—not enough to learn a reliable decision boundary. The detection stage takes the opposite approach: an out-of-distribution (OOD) model is trained on normal samples only, learning a boundary that encloses the normal region, with the anomaly score defined as the distance to that boundary. Conformal calibration then constrains both the miss rate and the false-alarm rate, yielding two thresholds that sort parts into pass, anomaly, and review-needed.

4. 4

   Trace Anomaly Scores to Process Steps

   The anomaly score is not a black-box output: feature-selection results and interpretability analysis jointly show that selected key features concentrate in the phases where mechanical change is most intense, and anomalous samples exhibit deviation patterns consistent with the underlying process mechanics. The higher the score, the larger the deviation, letting engineers pinpoint the offending process step directly.