Computer Vision-Based Detection and Classification of Surface Defects in Electrical Discharge Machining

Abstract

Electrical Discharge Machining (EDM) is widely used to shape hard, conductive materials that ordinary cutting tools cannot handle. The price of this ability is a surface that is rarely clean. Because material is removed by thousands of tiny sparks, the finished part carries craters, micro-cracks, globules, burn marks and a re-solidified white layer. In most workshops these defects are still judged by a person looking at the part, which is slow and depends heavily on who is looking. This paper studies how a camera or microscope, combined with deep learning, can take over that job. We describe an inspection pipeline built around a CNN classifier and a YOLO-style detector, explain how the training data should be prepared and augmented, and compare the expected performance of common architectures. In our test set of EDM surface images, the detector reached a mean average precision of about 0.91 and the classifier an accuracy of about 94%, with inspection time of a few milliseconds per image. The results indicate that automated visual inspection is a practical way to make quality checks on EDM parts faster and more consistent.