Microseismic (MS) monitoring technology is widely applied in early warning of rock-burst. Real-time arrivals picking of MS waveforms with low computational complexity is urgently needed, especially in terms of MS monitoring in smart mine. We constructed a real-time arrivals picking method named Real-Time Picking Network (RTP-Net), which combines the characteristics of recurrent and convolutional neural networks. This model includes three parts: a short-term feature extraction module based on convolution, a long-term feature extraction module based on gated recurrent unit (GRU) and a feature judgment module based on convolution. Considering the characteristics of MS arrival time picking, an improved receiver operating characteristic (ROC) curve is designed to evaluate the performance of RTP-net, which is proved to have good stability. 5-fold cross validation based on simulation data and measured data shows that, compared with Short Term Averaging/Long Term Averaging (STA/LTA), Akaike Information Criterion (AIC) and U-Net methods, RTP-net has higher real-time picking accuracy, which is 91.73% for P-wave and 86.67% for S-wave. Compared with U-Net, the computational complexity of RTP-Net is much lower and the processing time of each sampling point is 0.0621 millisecond, which means RTP-Net may have good application prospects in embedded MS sensors and real-time MS monitoring systems.

microseismic,arrival time picking,deep learning,real-time processing