A CNN-based Unsupervised Learning System for Detecting Stator Pin Insertion Defects in Blower Motors for Industrial Environments

Abstract

This study proposes a system for detecting defective automobile blower motors based on deep one-class classification (DOCC). DOCC determines defects by learning only normal data to overcome the limited data environment experienced in the actual industrial sites where it is difficult to obtain defective image data associated with defective products. Furthermore, this study proposes two types of inspection systems with different camera configurations. To verify the hardware simplification and economic efficiency of the inspection system, the defect detection performance of a single ceiling camera configuration (inspection system A) and dual-side camera configurations (inspection system B) is compared and analyzed to solve the viewing angle problem caused by the lateral curvature. Based on the conducted experiments, the proposed inspection system shows that it can efficiently detect the defective stator pin insertion of the blower motor.