Weld defect detection with convolutional neural network: an application of deep learning

In the present era of Industry 4.0, the manufacturing sector has expressed its high approval of automated camera-based weld defect detection in micro, small, and medium enterprises. However, the current classification system often inaccurately detects defective or non-defective parts. This study aims to evaluate and enhance the accuracy of welding operations by implementing deep learning convolutional neural networks (DCNN). The methodology of DCNN facilitates accurately identifying missing or incomplete weld processes on a safety–critical automotive subassembly metallic component. The DCNN method is applied to this welding image data set of 10,000 digitalized OK and Not-OK images as the training set. Afterward, the model was built, evaluated, and optimized to detect welding flaws. The model’s output was tested and results show that DCNN can predict defects with high accuracy. Data augmentation improves accuracy up to 99.01% after training over 9600 images. Thus, the proposed DCNN system could help quickly detect visual defects with precision. The proposed research can benefit firms facing similar quality issues.