朱德福,王德玉,于彪彪. 基于离散元−机器学习的铝土矿矿柱强度预测方法[J]. 煤炭学报,2024,49(7):1−13. doi: 10.13225/j.cnki.jccs.2023.0888
引用本文: 朱德福,王德玉,于彪彪. 基于离散元−机器学习的铝土矿矿柱强度预测方法[J]. 煤炭学报,2024,49(7):1−13. doi: 10.13225/j.cnki.jccs.2023.0888
ZHU Defu,WANG Deyu,YU Biaobiao. DEM-ML investigation of bauxite pillar strength prediction method[J]. Journal of China Coal Society,2024,49(7):1−13. doi: 10.13225/j.cnki.jccs.2023.0888
Citation: ZHU Defu,WANG Deyu,YU Biaobiao. DEM-ML investigation of bauxite pillar strength prediction method[J]. Journal of China Coal Society,2024,49(7):1−13. doi: 10.13225/j.cnki.jccs.2023.0888

基于离散元−机器学习的铝土矿矿柱强度预测方法

DEM-ML investigation of bauxite pillar strength prediction method

  • 摘要: 矿柱极限承载能力与矿柱尺寸参数密切相关,科学地预测矿柱强度是空场法安全高效开采铝土矿的关键。为了准确高效地预测矿柱强度,融合运用离散元方法(DEM)的模型参数化、样本数据强扩展性与机器学习(ML)方法的数据驱动优势,选取矿柱尺寸参数(长、宽、高)作为影响因子,开发了Grasshopper参数化建模电池组,实现了等块体密度的矿柱黏合块体模型(BBM)的参数化构建,结合矿体节理分布特征实测结果,利用3DEC程序构建了300组黏合块体−离散裂隙网络(BBM-DFN)矿柱离散元数值模型,开展了矿柱承载特性试验,监测并建立了机器学习数据集,且验证了此数据集的可靠性;分别以支持向量机(SVM)、BP神经网络、随机森林(RF)、高斯过程回归(GPR) 4种算法构建了矿柱强度预测模型,根据回归类模型评价指标(判定系数R2、可解释方差EEVS、平均绝对误差EMAE、均方误差EMSE)开展了最优模型的评选,结合改进的量子粒子群智能优化算法(IQPSO)进一步优化模型,利用该模型建立了矿柱影响因子与强度之间的非线性映射关系。研究表明:由矿柱强度参数化模拟结果可知,随着矿柱宽高比增加强度显著提升,长宽比对强度影响幅度相对较小,当矿柱高度和横截面积相同时,不同截面矿柱承载能力依次为:正方形 > 长方形;当矿柱宽高比大于1时,方形截面矿柱强度影响因子敏感性主次顺序为:矿柱宽(长)度 > 矿柱高度;根据机器学习算法指标综合评价,SVM模型是矿柱强度预测的最佳模型(R2=0.953、EEVS=0.953、EMAE=0.608、EMSE=0.551),结合IQPSO算法优化后模型预测性能得到了进一步提升(R2=0.985、EEVS=0.986、EMAE=0.373、EMSE=0.239);将IQPSO-SVM矿柱强度预测值与3种经典硬岩矿柱强度公式计算结果进行了讨论分析,得出了Hedley公式针对铝土矿强度计算不适用,Krauland公式适用于宽高比小于4时,Esterhuizen公式可通过调整不连续因子(F)进行较为准确的强度计算。研究成果为硬岩矿柱强度的预测提供了一种解决方案,拓宽了矿柱(群)稳定性评价的思路。

     

    Abstract: The ultimate loading capacity of the pillar is closely related to the bauxite pillar size parameters, and scientific prediction of pillar strength is the key to safe and efficient mining by the airfield method. In order to accurately and efficiently predict the pillar strength, combining the model parameterisation and high scalability of sample data using Discrete Element Methods (DEM) with the data-driven benefits of Machine Learning (ML) methods, selection of pillar size parameters (length, width and height) as influencing factors, grasshopper parametric modelling battery pack developed, parametric construction of a Bond Block Model (BBM) for pillars with equal block densities achieved, combined with the measured results of the distribution characteristics of the pillar joints, used the 3DEC program to construct 300-group Bond Block Model-Discrete Fracture Network (BBM-DFN) discrete element numerical model of the pillars, carried out tests on the loading characteristics of the pillar, monitor and build a machine learning dataset and verify its reliability; Four algorithms, namely Support Vector Machine (SVM), BP neural network, Random Forest (RF) and Gaussian Process Regression (GPR), were used to construct the pillar strength prediction model. The selection of the best model was carried out based on the regression class model evaluation indicators (R-Square R2, Explained Variance Score EEVS, Mean Absolute Error EMAE, Mean Squared Error EMSE), further optimisation of the model in combination with the Improved Quantum Particle Swarm Intelligent Optimisation Algorithm (IQPSO), and use the model to establish a non-linear mapping relationship between the pillar influencing factors and strength. The study shows that: from the parametric simulation results of the pillar strength, it can be seen that the strength increases significantly with the increase of the width to height ratio of the pillar; when the height and cross-sectional area of the pillar are the same, the loading capacity of different cross-sectional pillars is in the following order: square more than rectangular; When the pillar width-to-height ratio is greater than 1, the order of sensitivity of the strength factors affecting the strength of a square pillar is: pillar width (length) more than pillar height; Comprehensive evaluation based on Machine Learning algorithm metrics, the SVM model is the best model for pillar strength prediction (R2=0.953、EEVS=0.953、EMAE=0.608、EMSE=0.551), and the model prediction performance is further improved after combining with the IQPSO algorithm optimization (R2=0.985、EEVS=0.986、EMAE=0.373、EMSE=0.239); The IQPSO-SVM pillar strength prediction values are discussed and analysed with three classic hard rock pillar strength formulae calculation results, and it was concluded that the Hedley’s formula is not applicable for bauxite strength calculations, Krauland’s formula is applicable for aspect ratios less than 4, and Esterhuizen’s formula can be adapted for more accurate strength calculations by adjusting the large discontinuity factor (F) value. The results of the research have provided a solution for predicting hard rock pillar strength and have extended the idea of assessing pillar (group) stability.

     

/

返回文章
返回