Quick and accurate identification of water inrush types and sources of water inrush is of great significance for safe mining of coal mines. Laser-induced fluorescence (LIF) technology is rapid and sensitive in detection, which applies LIF to the detection of water inrush in coal mines and uses pattern recognition algorithm to quickly identify the source of water inrush. The current algorithms for identifying water samples are too dependent on pre-established water sample spectral databases When the water source is not in the library, it is easy to cause misidentification. The unsupervised learning algorithm DBSCAN does not require the label and category information of the sample set when clustering, which can reduce the misidentification of unknown categories. Therefore, the DBSCAN algorithm is used to identify the laser-induced fluorescence spectra in water inrush, and MVO is used for the parameter optimization of DBSCAN, which can eliminate the cumbersome manual parameter optimization process. In the experiment, four water samples were taken from the water intake point of Xieqiao Coal Mine, and 30 sets of spectral data were collected for each water sample. The fluorescence spectra of the water samples were collected using a USB2000+ spectrometer with a pixel of 2 048. First, the unsupervised learning algorithm automatic encoder (AE) reduces the dimension of the original spectral data to reduce the influence of redundant information in the spectral data on the clustering. The structure of the AE designed in this paper is a multi-layer network model between shallow and deep layers, which can reduce the original spectral data to 2 dimensions. In order to make the dimensionality reduction model sparse, the author adds a Dropout layer to the traditional AE algorithm. It can be seen from the experiment that the dimensionality reduction model after adding the Dropout layer has a faster convergence speed. Then, using the multivariate optimization (MVO) algorithm to optimize the DBSCAN parameters. In the parameter optimization process, the spectral recognition rate of the water sample after DBSCAN is up to 97.5%, and the corresponding range of the parameter Eps is ［0.023 66 0.040 65］. The normalized unscaled spectral data is used for DBSCAN cluster identification to verify the effectiveness of AE on the dimensionality reduction of water sample spectral data. The recognition rate of the original water sample spectrum by DBSCAN is up to 95%, which is 2.5% lower than that of the post-dimensional water sample. The results show that using AE dimensionality reduction data can improve the recognition rate of DBSCAN for different spectra. Finally, the supervised learning algorithm K nearest neighbor (KNN) is used to identify the water sample spectrum after dimension reduction, and the recognition result and the unsupervised learning algorithm DBSCAN are compared. The training set uses three water samples, and the test set uses four water samples. For the test set data, the supervised learning algorithm can only accurately identify the water sample categories contained in the training set, but all the categories that are not in the training set are identified incorrectly. On the contrary, DBSCAN can accurately identify the water sample spectrum not in the training set. The nonlinear dimensionality reduction algorithm AE can achieve dimensionality reduction on high-dimensional water spectral data. The use of MVO-DBSCAN for LIF spectral identification of coal mine water inrush can effectively reduce the misidentification caused by the incompleteness of the mine water source spectrum database.