Analysis of Optimizer Effects on CNN Model for Defect Identification in Welding Results of A 106 Grade B Carbon Steel Pipe

Welding is one of the sectors that play a crucial role in supporting the rapid development of infrastructure. However, in the process, many defects occur that sometimes escape inspection and will lead to damage in building construction. The purpose of conducting this research is to investigate the influence of optimization methods on Convolutional Neural Network (CNN) models in detecting the quality of welding results based on three conditions: Normal, Excessive Reinforcement, and Undercut. The area to be detected is a portion of the welding result with a circular welding path around a full A106 Grade B carbon steel pipe. The image capture process is carried out using a digital camera mounted on a tool for rotating around the pipe to follow the welding path. The captured image result of 3024 x 3024 pixels is first processed by converting RGB to grayscale. Afterwards, the image is resized to a smaller scale of 128 x 128 pixels to accelerate the training process. Finally, the training and testing processes are conducted using a CNN model. In the CNN model, there are several optimizers such as SGDM, Rmsrop, and Adam. In this study, a comparison is conducted to find the best optimizer in an effort to minimize the occurrence of overfitting as well as achieve optimal detection accuracy. For optimal detection, a dataset of 300 images is required, consisting of: 100 Normal images, 100 Excess Reinforcement images, and 100 Undercut images.