Defect Detection for Laser Welding Seam by Transfer Learning of Multiple Features of PEC Signal

ABSTRACT Pulsed eddy current (PEC) testing is an efficient method for online laser welding defect detection. However, obtaining sufficient feature target data of PEC signals to construct an accurate, reliable, and generalized defect detection model is challenging. An intelligent detection method is proposed for laser welding defects based on singular value decomposition (SVD) and multiple features transfer learning. SVD is used to decompose and reconstruct the PEC signal of a laser welding defect for achieving noise reduction. Time domain and frequency domain feature parameters of laser welding defect PEC signals are calculated for constructing the feature vector. The feature transfer is achieved by using a neural network-based transfer learning model, and the transferred feature data samples are used to train a Naive Bayes classifier. The proposed method is used to identify laser welding defects, achieving consistent results with destructive testing, which indicate the effectiveness of the detection method.