Towards a Zero-Defect in Welding: An Exploration of Factors to Improve the Training Data for Image Classification

Embracing zero-defect manufacturing in the era of Industry 4.0 in the welding industry requires understanding and addressing challenges linked to the methodologies and conditions governing the generation of datasets. We comprehensively explore the dataset composition of welding experiments across various works in the literature, aiming to discover patterns, variations, and similarities that affect the performance of defect detection methods. Our exploration revealed that, while considerable effort is dedicated to creating precise defect detection models, more attention needs to be paid to the conditions for composing an efficient training dataset considering the constraints of the real-time weld defect detection implementation. Therefore, in this study, we investigated the effect of the variation in the number of trials and the frame rate on defect detection by performing several experiments. We performed experiments using the Tungsten Inert Gas (TIG) welding Aluminium 5083 public dataset. Our findings show that a high frame rate does not always enhance the model’s performance for defect detection. Additionally, we attained 78% accuracy with 16 trials (eight non-defective, eight. defective) for defect detection with our dataset. To enhance the performance of our defect detection setup by 11%, we would need to double the number of trials, presenting challenges in generating defective trials within an industrial context. The findings contribute valuable insights for enhancing the effectiveness of real-time defect detection in welding processes.