CNN-Based Detection of Welding Crack Defects in Radiographic Non-Destructive Testing

In the industrial sector, the focus in recent years has been on enhancing production and minimizing human error. Therefore, engineers have used Non-Destructive Testing to evaluate material by combining artificial intelligence technologies with NDT. One of the field applications of NDT is the detection of welding defects. The use of neural networks can greatly enhance the accuracy of detecting defects in industrial welding. A convolutional neural network with triple classification for welding defects has been suggested. In the first step, the original images are cut down to 150 x 150 pixels. Subsequently, the images are categorized into three groups: Training, Testing, and Validation. In the triple classification experiment (Crack, other types of Defects, No Defects), the CNN model had 6 layers and 9,667 parameters. The model accuracy approached 92% after 800 epochs. The F1 factors of crack, other types of defects, and no defects were 100%, 91%, and 90%, respectively. The article provides methods used by CNN in detecting welding defects and highlights the potential to improve defect detection accuracy.