Real-Time Welding Defect Detection and Classification Using Artificial Intelligence and Its Implementation in Manufacturing Plants

Weld joint flaws can lead to part and assembly rejection, costly repairs, a significant reduction in performance under normal operating conditions, and, in the worst-case scenario, serious property, and life loss. In reality, flawless welding is nearly impossible, and in most situations, providing the relevant service functions is insufficient. However, early detection and segregation are still preferable. To test the quality of welded joints, several methodologies have been fabricated time and again. Non-destructive technology has essentially replaced the destructive methods of testing today, as this form of testing allows continuous and in-service inspection and is the basis for quality inspection in modern world. In our work, we use combination of digital image processing techniques and a well-established deep learning model for real-time detection of surface welding defects along with displaying the severity of each defect after their classifications. Moreover, the paper also shows how can this technology be used in an actual large-scale manufacturing plant in order to test the quality of welded joints. The objective is achieved using a hardware and software system that consists of a low-cost vision system for acquiring the image of a job, computer vision libraries, and a deep learning model for statistical analysis for quality monitoring.