摘要
煤矿采空区覆岩破裂信号作为开采沉陷的前兆特征,对其开展监测有助于预警采空区塌陷事件.但现有手段难以实现大范围、全方位、分布式的监测.以我国宁东矿区羊场湾煤矿为研究区域,引入分布式声波传感技术(distributed acoustic sensing,简称DAS)对采空区覆岩破裂信号开展连续监测.针对DAS数据信噪较低的问题,对比试验了 5种去噪方法.对预处理后的信号开展时频分析,提取覆岩破裂信号;进一步将DAS信号转换为递归图以构建数据集,训练基于卷积神经网络的覆岩破裂信号智能识别模型.结果表明,同步压缩小波变换能够很好地压制DAS数据的噪声.覆岩破裂信号与非覆岩破裂信号的递归图之间具有明显区别,训练得到的VGG-16模型在分类二者的任务上实现了 85％的准确率.因此,利用DAS技术监测覆岩破裂具有可行性,本研究所提出的基于递归图和卷积神经网络VGG-16的深度学习方法可实现对覆岩破裂信号的智能识别.研究成果为后续开发基于DAS系统的开采沉陷智能预警系统提供了一定技术支撑.
Abstract
[Objective]As socioeconomic development advances,challenges associated with coal mining beneath structures such as buildings,water bodies,and railways have intensified markedly.It is increasingly imperative to extract underground minerals within acceptable boundaries while diligently monitoring the environmental impacts of such activities.Previous research showed that mining-induced subsidence was a primary contributor to environmen-tal geological disasters in mining areas,particularly when the integrity of the overlying strata is breached.There-fore,it is crucial to develop methodologies for the early detection of surface subsidence,which requires in-depth re-search into monitoring the fracture signals from overburden rock.Existing methods,including acoustic emission and microseismic monitoring systems,face significant challenges in achieving widespread,comprehensive,and distribu-ted monitoring.In response to these limitations,distributed acoustic sensing(DAS),a state-of-the-art optoelec-tronic sensing technique,has recently gained prominence and been extensively employed across geophysical explo-ration fields such as oil and gas exploration and seismic monitoring.We explore applying DAS technology to en-hance the monitoring and identification of fracture signals in the overburden of mined-out areas,aiming to improve both safety and sustainability in mining operations.[Methods]This research selects a coal mine in Ningdong town,Lingwu city,Ningxia Hui Autonomous Region,China.DAS technology is used to continuously monitor the fracture signals of the overburden in underground voids.A fibre optic cable was installed at the bottom of a trench stretching parallel to the coal mining face with dimensions of approximately 1 km in length,15 cm in width,and 30 cm in depth.Additionally,several triaxial node seismometers were deployed along the route for comparison validation.Given the low signal-to-noise ratio of DAS data,comparative experiments were conducted using five denoising tech-niques:high-pass filtering,empirical mode decomposition,Fourier transform,F-X deconvolution,and synchronous compression wavelet transform.The DAS signals were preprocessed through detrending,mean removal,and denois-ing,followed by time-frequency analysis to extract overburden fracture signals.The event signals collected by the DAS system which formed a dataset were converted into recurrence plots.This dataset was used to train an intelli-gent recognition model for overburden fracture signals based on the convolutional neural network VGG-16.[Re-sults]The results demonstrate that synchronous compression wavelet transform effectively eliminates noise from DAS data.The overburden fracture signals detected by DAS were consistent with those by seismometers.Recur-rence plots of DAS-collected fracture signals differed from nonfracture signals,which can be distinguished by the trained VGG-16 model with an accuracy of 85％.[Conclusions]Monitoring overburden fractures via DAS technolo-gy is feasible.The proposed deep learning approach,based on recurrence plots and the VGG-16 convolutional neu-ral network,can effectively recognize fracture signals.This research provides significant technical support for devel-oping an intelligent early warning system for mining subsidence based on DAS.
维普
万方数据