Automated visual detection of external welding defect using embedded machine learning

Abstract

This project is designed to perform a quality check on a welded material's surface. Welding is the most useful process for joining materials in the manufacturing, automotive, and construction industries. As a result, in order to meet a client requirements, all welding works must be inspected, with the most basic method is nondestructive visual inspection testing. Surface inspection can be performed using nondestructive methods such as dye-penetration testing and magnetic particle inspection. However, those methods are expensive, take a long time to complete inspections, and require a complex procedure to operate. This project proposed an automated visual inspection system that relies on embedded machine learning. This system is made up of a camera, a microcontroller, and web server. The camera-based system will detect any defects on the welded material, which will then be processed by the microcontroller before being displayed on the web server through Wi-Fi connection. As a result, this system was split into two parts: software and hardware. The Arduino IDE was used to programme the system, and Edge Impulse was used to develop the embedding machine learning model. This system's hardware consists only ESP32- CAM module. As a result, it is possible to create an automated system that is user friendly and has simple operation procedure. Furthermore, a low-cost system with a short inspection time have been developed. A sharp and clear image with single type of defect appear on the workpiece help to provide best performance of defect classification. However, most likely same image captured, e.g. good welding and overlap defect, decrease the effectiveness of detection. The detection accuracy of this system can reach up to 95% with more training data provided, as for this project each defects detection accuracy are falls between 75 - 94 percent. In conclusion, the use of embedded machine learning in non-destructive testing is successful for the visual inspection method.