基于CNN-LSTM的岩爆危险等级时序预测方法

Time-series prediction method for risk level of rockburst disaster based on CNN-LSTM

岩爆灾害是一种典型的矿山动力灾害现象,具有高度不确定性和不可预知性,严重威胁矿井安全生产。为了防止岩爆灾害破坏事故,需要掌握准确可靠的预测预报信息。为了预测岩爆未来时刻危险等级,提出一种基于CNN-LSTM的集成方法。首先,选取微震监测角频率、能量和凹凸体半径等特征量,根据灰色关联方法求出特征量关联度作为权重以标记岩爆危险指数;第二,对具有混沌特性的特征量时序数据进行相空间重构以表达其空间特征;第三,根据卷积神经网络(CNN)提取重构后相空间的空间特征,利用长短期记忆网络(LSTM)学习时间序列特征,预测岩爆特征量的未来状态;最后,使用粒子群算法优化广义神经网络模型(PSO-GRNN),并根据岩爆特征量未来状态评估其危险等级。实验结合冬瓜山铜矿微震监测实际工程,使用CNN-LSTM预测角频率、能量和凹凸体半径等未来状态,并采用PSO-GRNN预测其未来状态的岩爆等级。研究结果表明:本文提出的集成方法能够有效利用微震监测数据提前预测岩爆特征的未来状态及岩爆危险等级;与现有支持向量机和BP神经网络等经典方法相比,本文方法得到的未来状态预测值准确性更高。本研究成果为正确识别矿山岩爆当前活动及未来状态时的危险性提供理论支撑,为及时掌握矿山岩爆活动未来状态提供重要依据,同时,也可利用本文方法为类似岩爆灾害的其他地质灾害预警提供参考。

Rock burst disaster is a kind of typical mine dynamic disaster phenomenon,which has high uncertainty and unpredictability,and seriously threatens mine safety production.In order to prevent the occurrence of rockburst disaster,it is necessary to master more accurate and reliable forecast information.In order to predict the danger level of rockburst in the future,an integrated method based on CNN-LSTM was proposed.Firstly,according to the grey correlation method,the correlation degree of the angular frequency,energy and concave convex radius were calculated as the weight to mark the rockburst risk index.Secondly,phase space reconstruction was carried out to express the spatial characteristics of these characteristic variables.Thirdly,the spatial features of reconstructed phase space were extracted according to the convolutional neural network(CNN),and the long and short term memory network(LSTM)was used to learn the time series features and predict the future state of rockburst characteristic variables.Finally,Particle swarm optimization(PSO-GRNN)was used to optimize the generalized neural network model,and the risk level of rock burst was evaluated according to the future state of rockburst characteristic variables.In the experiment,in combination with the actual microseismic monitoring project of Dongguashan Copper Mine,CNN-LSTM was used to predict the future state of angular frequency,energy,concave-convex radius,etc.,and PSO-GRNN was used to predict rock burst risk level.The results show that the integrated method proposed in this paper can effectively predict the future state of rockburst characteristic variables and the rockburst risk level in advance with the microseismic monitoring data.Moreover,compared with the existing classical methods such as support vector machine and BP neural network,the predicted accuracy obtained by the proposed method is higher.The results provide theoretical support for correctly identifying the risk of rock burst in the current and future state,and provide an important basis for timely mastering the future state of rock burst.Furthermore,the method can also be used to provide reference information for the early warning of other geological disasters similar to rock burst disasters.