Pressure vessel-oriented visual inspection method based on deep learning.

Abstract

The detection of surface parameters of pressure vessel welds guarantees safe operation. To address the problems of low efficiency and poor accuracy of traditional manual inspection methods, a method for welding morphological parameters combined with vision and structured light is proposed in this study. First, a feature point extraction algorithm for weld parameters based on deep convolution was proposed. An accurate extraction method of weld image feature point coordinates was designed based on the combination of the loss function via seam undercut feature recognition and weld feature point extraction network structure. Second, a training data enhancement method based on the third-order non-uniform rational B-spline (NURBS) curve was proposed to reduce the amount of data collection for training. Finally, a pressure vessel measurement device was designed, and the feature point extraction performance of the deep network and common feature point extraction networks, DeepLabCut and HR-net, proposed in this study were compared to analyze the theoretical accuracy of the surface parameter measurement. The results indicated that the theoretical accuracy of the parameter measurements was within 0.065 mm.