Classification of Weld Defects Based on Computer Vision System Data and Deep Learning

Abstract

Visual and dimensional inspection is widely used in the search for external weld defects that locate on the surface of welded joints. Such inspections are carried out by trained specialists whereas computer vision-based quality control systems require the significant investments. In addition, the data about the object of scanning that can be obtained during the operation of computer vision systems irretrievably disappear, or remain in the memory of the system without their processing and any analysis to find probable patterns of welding defects. This study utilized the capabilities of the iRVision system, which is built into a Fanuc welding robot, to obtain the data about the weld joint after robotic welding. Such data could be accumulated with the help of the 3D Laser Scanning subsystem’s laser sensor. Pre-processing of the collected data was carried out. And the artificial neural networks, using deep learning, were subsequently developed for the classification of the data of welds into defective and non-defective. The results of the most effective of the developed neural network models are an accuracy of 0.92, a precision of 0.90, and a recall of 0.81 on the test sample. The presented results allow us to automate the quality control and visual inspection process of welded joints based on the data from the iRVision system. The research has shown that the perspective direction of the project is to improve the model to search for the small size defects of < 1.0 mm which were not classified during the training stage.