Crack parameter inversion using an ultrasonic-magnetic fusion neural network

Crack defects in ferromagnetic components may lead to structural degradation and safety risks during manufacturing and service. Ultrasonic testing and magnetic memory testing have complementary advantages, but single-modality methods are often limited in quantitative crack parameter inversion. To improve inversion accuracy, this paper proposes an ultrasonic-magnetic fusion neural network for the simultaneous prediction of crack length, horizontal coordinate, vertical coordinate and inclination angle. The model integrates ultrasonic full matrix capture (FMC) data and dual-component magnetic memory signals, introduces physical prior features, and uses a cross-modal attention fusion module (CAFM) with differentiated regression heads to enhance multi-parameter regression. The ultrasonic and magnetic memory datasets were generated using finite element simulations, and experimental data from 45 steel specimens were used for validation. On the simulation test set, the fusion model achieved mean absolute percentage error (MAPE) values of 4.40%, 1.36%, 3.34% and 9.62% for crack length, horizontal coordinate, vertical coordinate and angle, respectively. In experimental validation, the corresponding MAPE values were 5.41%, 1.81%, 7.54% and 11.84%. These results demonstrate that the fusion strategy can make better use of the complementary ultrasonic and magnetic information, leading to more stable performance across different crack parameters than the single-modality models.