Performance evaluation of CNN-based crack detection for electrical discharge machined steel surfaces

Abstract

Electrical discharge machining (EDM) is a non-traditional machining technique that is frequently employed on hard materials. In today's industrial practice, it has been the most prevalent non-traditional material removal procedure. It allows you to process challenging materials and construct complicated forms with excellent accuracy. In the context of image pre-processing approaches to discover defective or non-defective machining pieces through EDM, the Centre of attention of this article is utilizing convolutional neural networks (CNN) to create a machinery piece defective detection approach. A total of 180 datasets of varying sizes produced from public datasets, the proposed CNN model is assessed and compared to pre-trained networks, namely the VGG-16, VGG-19, ResNet-50, and ResNet-101 models. The assessment took into crack detection outcomes, and classification parameters such as accuracy, precision, recall, and F1-score. Also, we have given the receiver operating characteristic (ROC) curve, precision–recall curve, and confusion matrices for each model for the required classification to predict the defective cracks, and also, we included the histograms to find the probabilities of defective and non-defective cracks. The suggested model can discriminate between pictures that are cracked and those that are not. According to the findings, VGG-16 has the best accuracy out of all of these models. In comparison to prior conventional procedures, the proposed EDM crack defective detecting methodology gives great accuracy.VGG-16 attains 93% accuracy, 93% of F1-score, 94% of precision, 93% of recall, 93% of specificity, as well as 97% of AUC, testing findings suggest that our strategy is capable of incredible performances.