Deep Learning Approach for automatic detection of split defects on sheet metal stamping parts

Abstract

Sheet metal stamping processes are used primarily for high-volume products produced for a range of sectors, from white goods manufacturing to the automotive and aerospace sectors. However, the process is susceptible to defects. Due to the numerous potential defects that may arise in the stamping product, human inspectors are often deployed for their detection. However, they are unreliable and expensive, especially when operating at production speeds equivalent to the stamping rate. This study investigate CNN-based automatic inspection for stamping defects. The study carried out two sets of experiments. All the Experiments yielded high classification accuracy, recall and precision demonstrating the viability of the CNN method for defect detection in the sheet metal stamping process. Additionally, this study revealed that in limited data confounding factors can be a challenge. The second experiment further explored the impact of small neck defects, harsh lighting and reflections on defect detection. The observations indicated that the model struggled to identify defects occluded by reflections, particularly small neck defects.