Quality prediction and classifcation of resistance spot weld using artifcial neural network with open‑sourced, self‑executable andGUI‑based application tool Q‑Check

• Main Authors: SuhailaAbd Halim, SuhailaAbd Halim, Yupiter H. P. Manurung, Yupiter H. P. Manurung, MuhamadAiman Raziq, MuhamadAiman Raziq, ChengYee Low, ChengYee Low, Muhammad Saufy Rohmad, Muhammad Saufy Rohmad, John R. C. Dizon, John R. C. Dizon, Vladimir S. Kachinskyi, Vladimir S. Kachinskyi
• Format: Article
• Language: English
• Published: 2023
• Subjects: T Technology (General)
• Online Access: http://eprints.uthm.edu.my/10654/1/J15867_e7905ebedef880e4175e46a9cf254d31.pdf; http://eprints.uthm.edu.my/10654/; https://doi.org/10.1038/s41598-023-29906-0

Optimizing Resistance spot welding, often used as a time and cost-efective process in many industrial sectors, is very time-consuming due to the obscurity inherent within process with numerous interconnected welding parameters. Small changes in values will give efect to the quality of welds which actually can be easily analysed using application tool. Unfortunately, existing software to optimize the parameters are expensive, licensed and infexible which makes small industries and research centres refused to acquire. In this study, application tool using open-sourced and customized algorithm based on artifcial neural networks (ANN) was developed to enable better, fast, cheap and practical predictions of major parameters such as welding time, current and electrode force on tensile shear load bearing capacity (TSLBC) and weld quality classifcations (WQC). A supervised learning algorithm implemented in standard backpropagation neural network gradient descent (GD), stochastic gradient descent (SGD) and Levenberg–Marquardt (LM) was constructed using TensorFlow with Spyder IDE in python language. All the display and calculation processes are developed and compiled in the form of application tool of graphical user interface (GUI). Results showed that this low-cost application tool Q-Check based on ANN models can predict with 80% training and 20% test set on TSLBC with an accuracy of 87.220%, 92.865% and 93.670% for GD, SGD and LM algorithms respectively while on WQC 62.5% for GD and 75% for both SGD and LM. It is also expected that tool with fexible GUI can be widely used and enhanced by practitioner with minimum knowledge in the domain.