Koopman-GRU Model for Cap Sealing Pressure Anomaly Detection Using Filling-line Torque and Vision Data
Authors
Files
Abstract
Cap sealing quality directly affects leakage control, product stability and downstream packaging reliability in filling lines. This paper proposes a Koopman-GRU multi-source model for cap sealing pressure anomaly detection based on torque curves, rotation-angle sequences, visual cap-state images, cap-height measurements and station-quality tags. A synchronous sealing window is first constructed from torque and vision data; Koopman lifting then represents nonlinear sealing dynamics in a stable latent space; finally, GRU-based temporal modelling identifies abnormal sealing pressure, confidence and review priority. The simulated filling-line dataset contains 5,200 sealing windows, including normal caps, under-tightened caps, over-compressed seals, tilted caps, thread mismatch and uncertain visual cases. Compared with the rule-threshold baseline, the proposed model improves anomaly recall from 80.7% to 92.1%, increases macro-F1 to 0.895 and reduces manual inspection by 29.4%. The model remains stable under partial image blur, torque-sensor drift and short-term line-speed fluctuations. Calibrated evidence output supports station-level maintenance decisions and product- level rejection review under different production batches and inspection conditions. It also separates product-level rejection, station-level maintenance and low-confidence secondary review so that uncertain visual or torque evidence is not treated as a confirmed pressure defect. These results indicate that Koopman-GRU fusion can support reliable online cap inspection and traceable pressure-anomaly analysis in high-speed filling production, while low-confidence cases still require operator confirmation or secondary inspection.
