Convolutional Neural Network (CNN) Technology to Detect Welding Defects in Motorcycle Frames using Transfer Learning and Optimizer

The motorcycle frame is a component that supports the performance and safety of the vehicle. The design and material of the frame are carefully chosen by manufacturers to achieve an optimal balance between strength, weight, and maneuverability. The most crucial part of frame manufacturing is the welding process as it can influence the strength of the frame. This research discusses the classification of welding defects with categories: Incomplete Fusion, No Defect, Spatter, and Void using Convolutional Neural Network (CNN). The dataset used consists of 4734 images trained using the GoogleNet architecture and RMSprop optimizer. The accuracy obtained in this research is 92.04%, which is then compared with 5 previous research references that used VGG19, AlexNet, and ResNet-50 architectures. The accuracy of this research is better than the 5 existing references, with a sufficiently high accuracy value. This study can provide an easy solution for the inspection process in welding within industrial environments.