关键词: 焊缝缺陷, 主成分分析网络, 深度学习, 二维主成分分析, 鲁棒性, 范数

Abstract: Principal Component Analysis Network (PCANet) is a simplified deep subspace learning model based on Convolutional Neural Network (CNN).When PCANet is applied to the weld defect detection,it cannot reflect the complete structure information of data and is sensitive to noise.To solve these problems,a Robust Non-Greedy Bi-Directional two-dimensional PCANet algorithm (RNG-BDPCANet) weld defect online detection method was proposed,which used a bi-directional two-dimensional principal component analysis as the convolution kernel under norm distance metric to obtain the optimal global projection matrix of the objective function with a non-greedy strategy.It was robust to outliers.The experiments were carried out on the self-built weld artificial dataset,ORL and Yale B face dataset respectively.The results showed that the proposed algorithm had a significant improvement in the classification and robustness performances.

Key words: weld defects, principal component analysis network, deep learning, two-dimensional principal component analysis, robustness, norm