Intelligent phase picking of microseismic signals based on ResUNet in underground engineering

Abstract

With the continuous expansion of deep underground engineering and the growing demand for safety monitoring, microseismic monitoring has become a core method for early warning of rock mass fracture and engineering stability assessment. To address problems in existing methods, such as low data processing efficiency and poor phase recognition accuracy under low signal-to-noise ratio (SNR) conditions in complex geological environments, this study proposes an intelligent phase picking model based on ResUNet. The model integrates the residual learning mechanism of ResNet with the multi-scale feature extraction capability of UNet, effectively mitigating the vanishing gradient problem in deep networks. It also achieves cross-layer fusion of shallow detail features and deep semantic features through skip connections in the encoder-decoder structure. Compared with traditional short-time average/long-time average (STA/LTA) algorithms and advanced neural network models such as PhaseNet and EQTransformer, ResUNet shows superior performance in picking P- and S-wave phases. The model was trained on 400000 labeled microseismic signals from the Stanford earthquake dataset (STEAD) and was successfully applied to the Shizhuyuan polymetallic mine in Hunan Province, China. The results demonstrate that ResUNet achieves high picking accuracy and robustness in complex geological conditions, offering reliable technical support for early warning of disasters such as rockburst in deep underground engineering.

1. Introduction

With the rapid development of China's economic construction and the gradual depletion of surface resources, underground engineering projects are increasing, and resource extraction is advancing into deeper regions. Due to the complexity of the underground environment, frequent underground hazard incidents occur during underground engineering and underground space development, leading to various losses such as casualties, project delays, and equipment damage. Therefore, preventing and mitigating underground hazards has become a crucial issue in the process of underground engineering construction and resource extraction [1−3].

In the field of underground mining, major mining countries such as South Africa, the United States, Australia, Canada, and China have developed and applied various microseismic monitoring systems in recent decades to prevent and monitor hazards during the underground mining process [4]. Notably, the Engineering Seismology Group (ESG) microseismic monitoring system developed by the Canadian Engineering Seismology Group and the Institute of Mine Seismology (IMS) microseismic monitoring system developed by the South African Institute of Mine Seismology are widely utilized [5].

During the rock mass fracturing process in underground engineering, acoustic signals are generated, which is a phenomenon known as microseismicity. By studying these signals, we can obtain the acoustic characteristics of rock fractures and infer damage, fracture mechanisms, and instability processes of the rock mass from an acoustic perspective [6, 7]. In the research and development of microseismic monitoring systems, it has been confirmed that these systems can effectively capture vibration signals related to rock fractures and underground hazards (such as rock bursts, ceiling collapses, and slab failures) in the monitoring area, enabling analysis and research of the rock mass. However, two main issues currently hinder signal processing: first, the vast amount of data make manual identification and picking of these signals time-consuming and labor-intensive; second, the signals often contain substantial noise, as the microseismic monitoring system collects not only the desired microseismic signals but also irrelevant noise from blasting, equipment, and personnel, necessitating signal filtering and processing [8].