Seismic phase picking is the first key step of mine microseisms detection, and its accuracy often directly affects the quality of subsequent event processing, so we proposed a method for P-arrival picking of mine microseisms which is based on deep learning method. Firstly the CNNDet model is constructed for events detection and P-arrival pre-picking, and then the CGANet model was constructed to accurately pick up the P-arrival time for the detected events by introducing the self-attention mechanism and the gated recurrent unit. Comparison with STA/LTA, DPick and PpkNet shows that the precision and the recall ratio of seismic event detection by our method are more than 98% for the test sets, and the mean error and the standard deviation of P-arrival are 0.014 s and 0.051 s, respectively. Our method is superior to the above three methods in terms of precision, the recall ratio and the standard deviation. In addition, the experimental tests on samples with different SNRs prove that our method can still maintain high precision on the condition of low SNR. In the source location, our method also shows more excellent performance. The P-arrival picking method proposed in this paper which is based on gated recurrent unit and self-attention mechanism provides a new idea for microseisms monitoring and accurate identification of rock burst and other disasters.