Online detection of weld surface defects based on improved incremental learning approach

Complex welding processes and harsh industrial environment inevitably affect the tailored welding quality of large workpieces, and remain the major cause of surface defects. In this paper, an online defect detection method combining the improved incremental learning algorithm and the network-based recognition algorithm is proposed to improve the performance of real-time processing. The incremental learning algorithm, called generalized incremental two-dimensional principal component analysis (GI2DPCA), rapidly extracts the relevant pattern features from weld surface images. The principal components of weld images are incrementally estimated in a recursive form, rather than by directly solving the image covariance matrices. The GI2DPCA reduces the influence of irrelevant feature changes on the convergence rates of principal components. To improve the classification speed, the artificial feedforward neural network (FNN), instead of various complex deep learning networks, is applied to identify defects online by assigning the features extracted from images to weld classes. The proposed algorithm is evaluated on 3600 classified weld images, collected from different sources, and tested on the actual running platform. The experimental results indicate that the proposed algorithm is capable of enhancing the performances of the convergence rate and the classification rate, and the average processing speed also meets the real-time requirements for online defect detection.