Crack Orientation From Eddy Current Signals: Deep Learning Coupled With Fracture Mechanics

Crack orientation governs how the crack‐tip field is partitioned between fracture modes, yet automated nondestructive testing (NDT) rarely resolves it. This work couples eddy current testing (ECT), a one‐dimensional convolutional neural network (1D‐CNN) and linear elastic fracture mechanics. A finite element model of a differential ECT probe scanning a surface crack in an S355 steel plate provides impedance signatures for orientations between 0° and 180°; the real part, imaginary part, and magnitude of ΔZ train the network, which resolves orientation over 10 classes with 87.72% accuracy, 2.85° mean absolute error, and 3.41° root mean square error. Predicted angles feed a Raju–Newman stress intensity factor and a Paris‐law integration of remaining life. Because the induced currents run transverse to the service stress, ECT amplitude and crack driving force are anticorrelated: The strongest signals come from the least critical cracks. Amplitude‐based triage is therefore unsafe, and orientation‐resolved assessment is required.