The crack breathing model and crack identification method for rotors using nonlinearity induced by cracks are studied in this work. Firstly, the finite element method is utilized to model a rotor–bearing system with a response-dependent breathing crack to obtain the numerical data for crack identification. During the modelling, an improved breathing crack model is proposed, focused on the unreasonable assumption about crack closure line in the original crack closure line position (CCLP) model. Compared with the original model, the improved breathing model can reflect the nonlinear behavior of cracks better. Secondly, based on the established model, super-harmonic features at 1/3 and 1/2 of the critical rotating speeds under different crack locations and crack depths are extracted for crack identification. Additionally, the super-harmonic features from two measurement points are used as inputs into an artificial neural network with a Levenberg–Marquardt back-propagation algorithm, corresponding crack positions and depths as outputs. The robustness of the method is tested by examining the identification results under different levels of noise. The results show that the proposed crack identification method is efficient for simultaneous identification of crack depth and position in operating rotors.