A Deep Learning-Based Machine Vision System for Online Monitoring and Quality Evaluation During Multi-Layer Multi-Pass Welding

Abstract

Multi-layer multi-pass welding plays an important role in manufacturing industries such as nuclear power plants, pressure vessel manufacturing, and ship building. However, distortion or welding defects are still challenges; therefore, welding monitoring and quality control are essential tasks for the dynamic adjustment of execution during welding. The aim was to propose a machine vision system for monitoring and surface quality evaluation during multi-pass welding using a line scanner and infrared camera sensors. The cross-section modelling based on the line scanner data enabled the measurement of distortion and dynamic control of the welding plan. Lack of fusion, porosity, and burn-through defects were intentionally generated by controlling welding parameters to construct a defect inspection dataset. To reduce the influence of material surface colour, the proposed normal map approach combined with a deep learning approach was applied for inspecting the surface defects on each layer, achieving a mean average precision of 0.88. In addition to monitoring the temperature of the weld pool, a burn-through defect detection algorithm was introduced to track welding status. The whole system was integrated into a graphical user interface to visualize the welding progress. This work provides a solid foundation for monitoring and potential for the further development of the automatic adaptive welding system in multi-layer multi-pass welding.