Measurement and inspection of electrical discharge machined steel surfaces using deep neural networks

We propose an industrial measurement and inspection system for steel workpieces eroded by electrical discharge machining, which uses deep neural networks for surface roughness estimation and defect detection. Specifically, a convolutional neural network (CNN) is used as a regressor in order to obtain steel surface roughness and a CNN based on spatial pooling pyramid is applied for defect classification. In addition, a new method for the region of interest selection based on morphological reconstruction and mean shift filtering is proposed for defect detection and localization. The regressor and classifier based on deep neural networks proposed here outperform state-of-the-art methods using handcrafted feature extraction. We achieve a mean absolute percentage error of 7.32% on roughness estimation; on defect detection, our approach yields an accuracy of 97.26% and an area under the ROC curve metric of 99.09%.