Abstract: Convolutional neural network technology is applied to the identification of small faults in coalfield earthquakes. In view of the characteristics of geological structure gradually revealed in the mining process of coal mines，the accuracy of the model is improved by using dynamically updated training sam‐ ples. The end-to-end three-dimensional fully convolutional network is studied and constructed. The origi‐ nal seismic data，the imaginary part of Hilbert transform and the instantaneous phase sine or cosine four�channel input are used. The learning step convolution layer is used to replace the traditional pooling opera‐ tion to retain the detailed features，and the exponential linear unit activation function is introduced to en‐ hance the nonlinear mapping ability. In the verification of the 110905 working face，the three fault areas ex‐ plained by this method were verified by mining，and 29 faults were exposed，including 11 faults with a drop greater than 1 m. The main faults of DF6 and DF12 and their associated structures were clearly identified， which were highly consistent with the location of the exposed faults. The results show that this technology sig‐ nificantly improves the accuracy of small fault identification，but in complex geological structure areas，it is still necessary to further optimize the network structure and multi-source data fusion strategy.

Key words: convolutional neural network, comprehensive working face, three-dimensional seismic, minor fault identification