Multi-Step Intelligent P-phase Picking Model for Risk Assessment in Deep Underground Mines

Abstract

Accurate P-phase first arrival time is a premise for improving accuracy of seismic source localizations and achieving hazard warning. Traditional algorithms failed to meet the requirements of high precision and accuracy for microseismic (MS) monitoring in deep geological engineering. In this study, a multi-step model: convolutional neural network combined with K-means and AIC (CNN-KA) for picking arrival of P-phases is proposed. Firstly, convolutional neural network (CNN) technique is used to recognize waveforms of MS fractures. Secondly, maximum overlapping discrete wavelet transform and multi-resolution analysis are combined to denoise signals. Subsequently, a new picker was developed by introducing k-mean clustering to AIC. Finally, performance of the hybrid model was evaluated with open-source and field data. The results show that mean absolute error of CNN-KA is 0.0915s at 200Hz frequency, which is 86.65% lower than STA/LTA. In addition, a strategy is proposed to evaluate real-time mining risk by improving MS source location. An application in Pan Er Mine, Anhui Province, China showed that automatic location error of MS events was reduced from 37.33 m to 10.89 m. CNN-KA successfully warned of two potential geological hazards, which was verified by real-time mining pressure data. The proposed model greatly improves accuracy of p-phase arrivals and MS parameters. This study is of great value for early warning of geological hazards in underground geotechnical engineering.