College of Energy Engineering
Key Laboratory of Western Mines and Hazard Preventionof China Ministry of Education,Xi’an University of Science and Technology
Key Laboratory of Coal Resources Exploration andComprehensive Utilization, Ministry of Natural Resources
Xinjiang Energy Co., Ltd., State Energy Group

冲击地压是制约煤炭安全高效开采的重大灾害之一,实现冲击地压的智能化预警是保障煤矿智能安全开采的关键路径。b 值作为监测冲击地压的有效指标,掌握矿井开采过程中 b 值演化趋势对冲击地压的及时预警具有重要意义。为此基于相空间重构 (PSR) 与深度学习提出了对矿井开采中时间序列 b 值的短期预测方法,运用相空间重构技术将卷积神经网络识别及降噪后的 b 值映射到高维空间,混合遗传算法 (GA) 优化的长短期记忆网络 (LSTM) 学习高维数据特征构成 b 值预测模型 (PSR–GA–LSTM)。实例结合冲击地压矿井宽沟煤矿 W1123 综采工作面,计算了降噪后 b值的重构参数且实现了数据的重构。评价了不同模型的预测性能并对最优预测模型进行了实例分析。研究结果表明:时间序列 b 值经过降噪技术处理后,能增强模型对于 b 值趋势特征的学习能力和降低噪点对于冲击前兆信息的干扰;时间序列 b 值经过相空间重构及长短期记忆网络的超参数得到优化后,模型的预测精度能得到明显提升;较其他模型相比 PSR–GA–LSTM 的残差波动范围最小稳定在 0.005 以内,其误差 (RMSE)、平均绝对误差 (MAE) 和平均绝对百分比误差 (MAPE)分别为 0.001 51、0.001 33、0.29% 都低于其他模型;PSR–GA–LSTM 模型经过时间序列 b 值训练后,所预测的 b 值趋势蕴含着冲击前兆信息,能预先对冲击事件的发生提供 b 值预警指标。该模型对于匀速推进的冲击地压矿井 b 值趋势发展有着较好的预测能力,所用方法可为在冲击地压时间上演化发展的预测预警研究提供借鉴与参考。

Rock burst is one of serious disasters that inhibit safe and high efficient coal mining. The realization of intelligent pre-warning of rock burst is the critical path to ensure coal mine intelligent and safe mining. As the b value is an ef-fective monitoring indicator of rock burst, it is of great significance for a timely pre-warning of rock burst to grasp theevolution trend of b value in the process of mining. Therefore, based on the phase space reconstruction (PSR) and deeplearning, a short-term forecast method for the b value of time sequence in mine exploitation is proposed. The b value oftime sequence identified by CNN and denoised is extended to a high-dimensional space through phase space reconstruc-tion technique, and then the long short-term memory (LSTM) network optimized by the genetic algorithm (GA) learns thehigh-dimensional data feature, which constructs the b value prediction Model (PSR−GA−LSTM). Combined with theW1123 fully mechanized mining face of the Kuangou coal mine identified rock burst mine, the b value of time sequencedenoised is reconstructed using the optimized parameters of PSR. The prediction performance of different models is evalu-ated and the case research of the optimal prediction model is carried out. The research results show that after the b value oftime sequence is processed by noise reduction technology, the learning ability of the model for the b value trend featurecan be enhanced and the interference of noise to the precursory information of rock burst can be reduced. After the b valueof time sequence is reconstructed in phase space and the hyperparameters of the LSTM are optimized, the prediction ac-curacy of the model can be significantly improved. Compared with other models, the residual fluctuation range of thePSR−GA−LSTM model is the smallest and stable within 0.005, and its root mean square error (RMSE), mean absolute er-ror (MAE) and the mean absolute percentage error (MAPE) is 0.001 51, 0.001 33, 0.29%, which are lower than other mod-els. After the PSR−GA−LSTM model is trained on the b value of time sequence, the predicted b value trend contains theprecursory information of rock burst, which can provide b value pre-warning indicators for the occurrence of rock burstevents in advance. The model has a better ability to predict the trend development of the b value of rock burst mine withuniform advance, and the method used in this paper can provide a reference for the prediction and pre-warning research onthe evolution of rock burst in time.

国家自然科学基金资助项目(51874231);陕西省创新能力支撑计划资助项目(2020KJXX-006);陕煤联合基金资助项目(2019JLZ-04)

崔峰,何仕凤,来兴平,等. 基于相空间重构与深度学习的冲击地压矿井时间序列 b 值趋势[J]. 煤炭学报,2023,48(5):2022−2034.

CUI Feng,HE Shifeng,LAI Xingping,et al. Trend of time sequence b value of rock burst mine based on phasespace reconstruction and deep learning[J]. Journal of China Coal Society,2023,48(5):2022−2034.