An online surface defect detection method for weld seam based on SAE model and background extraction method

Abstract

The surface defect detection method of weld seam based on line-structured light has the advantages of non-contact measurement, high accuracy, and strong anti-interference capability, which has received increasing attention. How to improve the efficiency of defect detection to meet the needs of online detection in actual industry, while ensuring detection accuracy remains a challenge. This study proposed a surface defect detection method for weld seam based on Stacked Auto Encoder (SAE) model and background extraction method. In the proposed method, the raw weld contour data obtained from the structured light sensor is preprocessed to reduce the influence of environmental noise and sensor movement. Then defect detection is divided into two steps: defect recognition and defect segmentation. The former applies the SAE model to identify defective areas in the entire weld seam to avoid analyzing defect free areas and improve efficiency, while the latter uses the background extraction method to segment defects from the contour of the weld seam containing defects to reduce the complexity of defect segmentation. The proposed method has been applied to the typical defect detection of aluminum alloy samples of high-speed rail vehicle bodies, such as surface pore, arc pits, overlap, undercut, and surface collapse. The results show that the accuracy of the defect recognition model in recognizing continuous weld defects exceeds 97 %. The segmentation error of typical weld seam defects is within 0.2 mm.