Advanced ANN Techniques for Precise Detection and Classification of Welding Defects

Abstract

The implementation of the artificial neural network (ANN) algorithm for detecting and classifying welding defects is detailed in this study. A total of 558 welding workpiece images were processed using techniques such as resizing, auto-orientation, flipping, rotation, and annotation, ultimately expanding the dataset to 1,288 images. Feature extraction identified 24 traits across 12,000 data points, which were then condensed to 5,735 data points for the ANN model. The model employed 100 hidden layers, the ReLU activation function, and the L-BFGS-B solver, running for 200 iterations. The configuration achieved near-perfect results, with metrics such as the area under the curve (AUC), classification accuracy, and F1 score averaging a precision of 0.97. These outcomes demonstrate the ANN model's high efficacy in detecting and classifying welding defects, underscoring its potential application for quality assurance in the welding industry. Further investigation into specific defect types, including porosity, spatter, cracks, and undercuts, could further improve detection accuracy.