School of Mines, China University of Mining & Technology
Jiangsu Engineering Laboratory of Mine Earthquake Monitoring and Prevention, China University of Mining & Technology
Xuzhou Wushuo Information Co., Ltd.
School of Computer Science & Technology, China University of Mining & Technology

煤矿智能化建设的大背景下，如何高效的从冲击地压海量监测数据提取有效信息与提高预测预警准确率是未来的研究重点与难点。为解决目前基于物理指标冲击地压预测方法泛化能力较差、对海量数据特征挖掘不充分的困境，结合深度学习技术，初步尝试建立了物理指标与数据特征融合驱动的冲击地压时序预测方法。以陕西彬长矿区某强冲击危险工作面为背景，分析了多次大能量事件发生前物理指标的变化特征，并统计剖析了仅使用物理指标驱动的冲击地压危险预测指标的短板与不足；提出采用物理指标与数据特征融合驱动的冲击地压时序预测方法，预测模型包括数据预处理、特征提取以及预警模型构建3个模块，数据预处理将原有微震监测数据处理为具有特定时间窗的前兆模式序列，特征提取主要包括基于物理指标的显式特征以及基于卷积神经网络的数据隐式特征提取，提出基于注意力机制的显式特征和隐式特征的深度融合方法，并通过全连接网络实现预测模型分类，实现对不同冲击危险等级的大能量事件进行预测。模型测试结果表明：预测时长为未来1 d、未来2 d以及未来3 d时预测F1分别可达0.956、0.950以及0.854，现场可根据需求选用预测时长；工程应用时模型可对大能量事件准确预测，误差分析结果表明模型预测准确率较高，可满足现场需求。

Under the background of coal mine intelligent construction, how to efficiently extract adequate information from the massive monitoring data of coal burst and improve the accuracy of prediction and warning is the focus and difficulty of future research. Aiming to solve the challenges of the poor generalization ability of current physical model-based coal burst prediction methods and insufficient feature mining of massive data, combined with deep learning technology, a time series coal burst prediction method driven by the fusion of physical model and data feature was initially established. In this paper, taking a strong coal burst risk longwall face in Binchang Mining area, Shaanxi Province as the background, the variation characteristics of physical indicators before several large energy events are analyzed, and the shortcomings of coal burst risk prediction indicators driven only by physical models are statistically analyzed. The coal burst time domain prediction method driven by the fusion of physical model and data feature is proposed. The prediction model includes three modules: data preprocessing, feature extraction, and early warning model construction−the data preprocessing processes the original microseismic monitoring data into precursor pattern sequence within a specific time window. Feature extraction based on a physical model mainly includes the characteristics of explicit and implicit data feature extraction based on convolution neural network. The explicit and implicit characteristics of the depth of the fusion method was put forward based on attention mechanism, and through the whole connection network forecast model classification, the impact of different risk levels of large energy events. The model test results show that the prediction F1 score can reach 0.956, 0.950, and 0.854, respectively, when the prediction time is one day, two days, and three days in the future. The site can choose the prediction time according to the demand. In engineering application, the model can accurately predict large energy events, and the error analysis results show that the prediction accuracy of the model is high, which can meet the needs of the site.

国家自然科学基金资助项目 (52274098，U21A20110)；江苏省自然科学基金资助项目 (BK20221109)

曹安业，刘耀琪，杨旭，等. 物理指标与数据特征融合驱动的冲击地压时序预测方法[J]. 煤炭学报，2023，48(10)：3659−3673.

CAO Anye，LIU Yaoqi，YANG Xu，et al. Physical index and data fusion-driven method for coal burst prediction in time sequence[J]. Journal of China Coal Society，2023，48(10)：3659−3673.