Abstract: Mine microseismic signals contain a large amount of rock fracture information. How to accurately identify the rock fracture signals in complex signals has always been the key to microseismic monitoring and early warning. Four main types of signals to be identified (rock fracture, blasting, Bolting rigs and electrical noise) in mines in a high noise environment were comprehensively sorted out. The Mel−frequency cepstral coefficient method was used to convert the four types of signals into non−linear spectra on the Mel scale and then into the cepstral domain respectively. Combined with the results obtained by taking the difference in the time domain, they were finally transformed into heat maps in a one−dimensional visual way. Using the data set composed of Mel−frequency cepstral coefficient heat maps, a pre−trained network model was called through the transfer learning method and the training parameters were readjusted to form a new model. Finally, the automatic identification and classification of mine microseismic signals were realized through this model.The real−time microseismic monitoring data of a copper mine in Anhui Province was taken as the data basis for training. For the microseismic signals within the selected time period, they were input into the automatic identification and classification model. The results show that the Mel cepstrupt coefficient heat map is used as the input of the deep learning model to test the data monitored in the actual production process of the mine, and through the comparison of the accuracy and fitting time of the four network models, it is determined that they have high recognition accuracy in the training process of labeled and unlabeled data. As the main identification method for microseismic monitoring data, the accuracy and timeliness of the microseismic signal collected in the subsequent remining activities of the mine can be improved.