Development of a Physics-Informed Doubly Fed Cross-Residual Deep Neural Network for High-Precision Magnetic Flux Leakage Defect Size Estimation

Defect depth is an essential indicator in magnetic flux leakage (MFL) detection and estimation. The quantification errors for defect are closely related to length width errors, this feature has always been used support the operator's judgment identification. However, existing algorithms based on shallow deep neural networks only employed simple general network structures inspired by field of artificial intelligence; consequently, these lack physical concepts result large regarding size, especially depth. In article, describe integrate above theory into a network, we propose physics-informed doubly fed cross-residual (DfedResNet) suitable MFL learning. Physics-based studied integrated loss functions during training. DfedResNet quantifies defects data automatically extracts features defects. experimental results show that it effectively achieves high-precision length, width, simultaneously, Moreover, considers from all three dimensions training, use originally measured signal place recognized images avoid information further improve accuracy. model proposed article reduces within 0.3 mm 0.4% t . addition, compared with other traditional algorithms, improves accuracy 1–2 orders magnitude thus high performance.