Abstract: Aiming at the problems of low efficiency of manual discrimination and insufficient classification accuracy of
traditional machine learning methods for complex non-stationary signals in microseismic monitoring of deep mines,an automatic
classification model of microseismic signals based on Transformer architecture is proposed. Based on the real mine data,an annotated
data set including microseismic events,blasting vibration and noise was constructed. By extracting the time-frequency
features of the signal and combining the self-attention mechanism of Transformer,the model effectively captures the long-range
dependence and global features in the waveform,significantly improves the classification accuracy,and enhances the identification
ability of the waveform confusion phenomenon,showing good generalization and engineering applicability. The experimental
results show that the model achieves an overall classification accuracy of 96. 3% on the test set,and the recognition rate of microseismic
events and blasting signals exceeds 97%. It is significantly better than SVM,KNN,CNN-BiLSTM and VGG16 in
many performance indicators. This model provides technical support for the intelligent upgrading of mine microseismic monitoring
system and real-time accurate early warning of ground pressure disaster by effectively solving the problem of waveform confusion.

Key words: microseismic monitoring,signal classification,Transformer model,deep learning,deep mine