Physics-Informed Neural Network for Real-Time Imaging and Evaluation of Defect Based on High-Definition ACFM Probe

Defects reduce the bearing capacity of underwater structures. The determination of defect size provides guidance for underwater structure safety assessment and maintenance. Alternating current field measurement technique is widely used in underwater structure detection for the advantages of non-contact detection and quantitative ability of defects. Generally, the calibration and physical mechanism inversion methods are proposed to evaluate dimensions of defects with a lot of assumptions and simplifications. The artificial intelligence quantitative method needs a mass of expensive sample data. This method is inapplicability in the alternating current field measurement of ocean engineering field. To solve the small samples and the critical quantization problems, a novel physics-informed neural network is constructed for real-time imaging and evaluation of defects. A high-definition alternating current field measurement probe and instrument are designed to achieve high-precision acquisition of distorted magnetic images. A universal current disturbance physical inversion mechanism is developed to obtain the real morphology of defects without considering defect type under uniform guidelines. The physical defect morphology inversion mechanism is embedded into a deep YOLO neural network as one layer to achieve imaging and evaluation of defects with small samples. The results show that the physics-informed neural network has higher evaluation accuracy for defects in real-time with small samples. The detection accuracy reached 99.6% and the average quantization error was 1.32.