Cross-Stage Defect Prediction from Heterogeneous Sensor Data in Multi-Step Manufacturing Processes

Abstract

Defects observed at the end of a manufacturing process may originate from small deviations occurring much earlier in production, making end-stage inspection insufficient for identifying their causes. This paper investigates cross-stage defect prediction using heterogeneous sensor measurements collected from multiple production stages. The proposed method organizes sensor observations according to process stage and learns separate representations for operating conditions, equipment signals, and intermediate quality measurements before integrating them for final defect prediction. A stage-level attention mechanism is introduced to estimate the contribution of different production stages to each prediction, allowing the model to distinguish persistent process deviations from isolated sensor fluctuations. Experiments on multi-stage production data demonstrate higher defect detection accuracy and earlier prediction capability than models that concatenate all sensor variables into a single input representation. Stage-removal experiments further identify several upstream process stages that provide predictive information well before final inspection. The approach provides both early quality warnings and stage-level evidence that can support subsequent process investigation.