The invention relates to an OTDR fault feature judgment method based on a differential evolution neural network. Firstly, sufficient OTDR fault feature diagnosis data are collected, the diagnosis dataare input into the neural network to be trained after being denoised and cut off, and the trained neural network can be applied to OTDR fault point judgment. According to the method, the fault pointjudgment problem of the OTDR is converted into the mode recognition of the neural network, and the position and type of the fault point are judged with high precision by using the mode recognition advantage of the neural network.