Robot-Aided Quality Inspection of Plastic Injection Molding Parts Using an AI Anomaly Detection Approach in an Industrial Environment

Abstract

While artificial intelligence has shown promising results for quality inspection, it often requires large training datasets, which are impractical for industrial applications. Defective parts are heavily underrepresented in normal production, which makes it a poorly posed problem for supervised learning approaches. In this work, this issue is tackled by combining an automated inspection procedure with an anomaly detection approach for defect detection. A 7-DoF robotic manipulator was used to automate part handling in front of an industrial optical camera sensor. The captured images were used to train a PaDiM anomaly detection network to reconstruct a normal image of the part. The results show that various defects can be detected with defect detection rates up to 100% while maintaining approximately 91% specificity using a small dataset of 117 parts.