Weld surface defect detection based on improved two-dimensional principal component analysis

Robot welding inevitably brings various weld defects due to the irregularity of part shape and the complexity of welding process. When applying two-dimensional principal component analysis to the detection of weld surface defects, it often faces problems such as high computational complexity, low classification accuracy and inability to incrementally learn. To resolve these problems, a mean updated incremental two-dimensional principal component analysis (MUI2DPCA) algorithm is proposed. MUI2DPCA and feedforward neural network (FNN) are combined to detect weld surface defects online. Firstly, the local block images are obtained by preprocessing the video frame images captured by the camera. Then, the pattern features of weld block images are extracted online by MUI2DPCA. The algorithm performs incremental iterative estimation on principal components of block images, significantly reduces the computational complexity, and incrementally updates the current sample mean to reduce the influence of irrelevant feature changes on the convergences of principal components. Finally, FNN is used to establish the relationship between pattern features and weld categories, and the weld defect detection information is obtained in real time. The test results show that the average classification accuracy of the detection method is 95.40%, and the average processing speed reaches 29 frames per second, and it can meet the real-time requirements of weld detection.