ABSTRACT The evaluation of fatigue cracks is essential for the detection of early damage to railway tracks. Laser nonlinear wave modulation spectroscopy using broadband excitation is an effective method for fatigue crack detection that can solve the problem of frequency combination optimisation. However, due to the complex components of laser nonlinear ultrasonic signals, the evaluation of fatigue cracks using laser nonlinear ultrasonic signals remains a challenge. In this study, wavelet packet decomposition and principal component analysis were used to extract various features of nonlinear ultrasonic signals in the frequency domain, and four new combinations of features were obtained. These new features were used as input for the evaluation model. Finally, a fatigue crack evaluation model based on adaptive particle swarm optimisation-support vector machines was proposed to accurately evaluate fatigue cracks of different lengths. The experiments conducted showed that the fourth new combination of features and evaluation model proposed in this paper can improve the accuracy of evaluating rail fatigue cracks of different lengths, reaching a classification accuracy of 97%.