Artificial vision inspection system for anomaly detection in metal stamped parts

Abstract

Detecting anomalies in the production process is an important necessity in large scale industrial manufacturing, and although there are methods to detect such anomalies, these often require humans, which is not ideal where consistency is required, or make use of tailored systems to detect a specific anomaly, demanding previous knowledge of the said anomaly. Therefore, we evaluated several autoencoder networks in two cases: general defects and specific metal stamping defects. Two models are proposed, myCAE and myCAE_optuna, which obtained scores of 0.91 and 0.93, respectively, in the SSIM tests, with the latter achieving a precision above 95% in five of the seven categories of anomalies tested.