In order to quickly and accurately identify the source types of coal mine water inrush, we propose a method of constructing a multilayer regularization extreme learning machine (M-RELM) model, which combines the functions of nonlinear feature extraction and classification learning. The fluorescence spectra of water samples are obtained by laser induced fluorescence (LIF) technique as the input of model. The features of fluorescence spectra are extracted by the improved auto encoder (AE) to form the feature space of the model hidden layer. In order to reduce the effect of noise and anomaly of spectra on classification results, the algorithm of extreme learning machine(ELM) is optimized regularly. According to whether the unknown samples are used to construct the training set, the model is optimized regularly by the L2 norm regularization (L2-RELM) or the graph manifold regularization (GM-RELM), which realizes the supervised or semi-supervised classification learning. By propagating between the hidden layers of different functions, M-RELM is constructed, and the integration of pre-training and training is completed. The water inrush samples in Huainan area coal mine as the experimental object, the performance compares with the support vector machine (SVM) and ELM with a single hidden layer. On the samples set containing mixed water, the average testing accuracy of the model can reach more than 94% and the training time is about 0.2 s. On all water samples containing the unknown samples, the testing accuracy of GM-RELM is increased by 2% than L2-RELM. The experimental results show that the M-RELM model is more suitable for the identification requirements of coal mine water inrush.