The microseismic signals collected in mines are often mixed with environmental noise (such as power interference, blasting signals, etc.). Traditional manual screening methods are inefficient and affect the accuracy and timeliness of disaster warnings. Therefore, a new signal processing and recognition framework was designed, and a signal recognition method combining EWT (empirical wavelet transform) and CNN (convolutional neural network) is proposed. Firstly, the signal decomposition effect of EWT and EMD (empirical modal decomposition) was compared through simulated signals, and then the EWT decomposition was carried out on the microseismic signal, environmental random noise (non-power factor interference and influence), transient tip pulse noise, power frequency harmonic interference noise, and knocking noise. It was decomposed into a series of different modes, and the CNN architecture with multi-channel input and multi-scale convolutional kernel was designed, so that different signal modes can realize feature extraction and classification under a unified framework. Make full use of CNN's advantages in feature extraction and classification, and effectively distinguish the five mixed signal categories in the data set. Finally, the model training classification accuracy reaches 96. 42%, and the recognition accuracy of microseismic signals reaches 99.3%, Compared with EMD-CNN and traditional neural network methods, the proposed method shows higher accuracy and stronger robustness in recognition. The multimodal EWT-CNN signal recognition method has effectively improved the recognition accuracy of mine microseismic signals, and provides efficient and reliable technical support for the early warning and prevention of mine geological disasters.