The rolling contact fatigue cracks on the rail surface and upper surface usually exist in the form of inclined cracks or multiangle complex cracks, which are difficult to detect and evaluate. Based on this, the wireless power transfer-eddy current testing (WPTECT) is adopted. A new probe structure is designed, and neural network algorithms are combined to detect and evaluate cracks. Firstly, different from the existing wireless power transfer-eddy current testing methods, the resonant circuit is constructed by increasing the excitation frequency instead of the series-parallel capacitance. Secondly, according to the characteristics of complex cracks, a directional probe structure consisting of two eight-figure excitation coils and two rectangular receiving coils is designed. Finally, the features of the detected signal are fully extracted, and the cracks are identified by the radial basis function neural network algorithm. Simulation and experimental results show that the proposed probe structure is sensitive to defects at any angle. Meanwhile, the recognition accuracy of the radial basis function algorithm for oblique crack, T crack, Y crack, and T crack with 1. 2 mm lift-off is 92. 00% , 95. 27% , 96. 64% , and 89. 50% , respectively.