Physics-Guided Patch Distribution Modeling for Unsupervised Pin Defect Detection of Electrical Connectors

Electrical connectors are critical components in electronic systems, and pin-related defects may lead to poor electrical contact, signal transmission failure, and product rejection. Automated inspection of connector pins is therefore essential for ensuring manufacturing quality and reliability. However, collecting and annotating sufficient defective samples remains challenging in industrial environments, limiting the applicability of conventional supervised learning methods. This paper proposes a physics-guided Patch Distribution Modeling (PGPDM) framework for unsupervised electrical connector pin defect detection. Trained exclusively on normal samples, it embeds connector structural priors via ROI masks to filter irrelevant background and focus feature learning on pin defect zones. Multiscale deep features from defect-free samples are utilized to build statistical distributions, and anomalies are detected by measuring feature distribution deviations at inference. Experimental results on a connector pin inspection dataset demonstrate that the proposed approach effectively highlights defective regions and improves detection performance compared with conventional feature-distribution-based anomaly detection methods. The proposed framework provides an accurate, annotation-efficient, and practically deployable solution for industrial connector inspection.