Method for quantitative eddy current testing of rail cracks based on gated recurrent neural network

Quantitative nondestructive testing of steel rails is crucial for railway safety. Existing deep-learning eddy current testing (ECT) methods for rail cracks are limited to single-parameter inversion and require separate models for vertical and oblique cracks. This paper proposes a multi-parameter quantification method based on crack contour reconstruction. A gated recurrent unit-recurrent neural network (GRU-RNN) reconstructs vertical-crack contours from scanning phase signals, from which crack depth and width are calculated. To extend the pre-trained network to oblique cracks, a phase-curve symmetrization transformation is developed based on finite element analysis, converting asymmetric oblique-crack phase curves into equivalent symmetric phase curves of vertical cracks with identical dimensions. The inclination angle is determined by the asymmetry rate of the original phase curve. The method was evaluated on two identical vertical cracks from different rail samples for cross-specimen consistency. For oblique cracks, the transformation was calibrated on three cracks and tested on three others, all with identical depth and width but different inclination angles. Average relative errors are 4.50% and 7.50% for vertical-crack depth and width, and 3.67%, 6.67% and 9.48% for oblique-crack depth, width, and inclination angle, respectively. The method enables multi-parameter inversion of both crack types using a single neural network.