基于深度学习的露天铀矿可爆性智能分级模型研究

Intelligent Classification of Blastability for Open-pit Uranium Mine based on Deep Learning

湖山铀矿属于特大型露天铀矿山,目前矿山爆破生产为“一次设计,长期使用”,故存在爆破参数缺乏动态调整、炸药单耗高、爆破效果不理想的问题,对此,可通过对爆破区块进行动态可爆性分级管理并反馈调控爆破设计来解决。本研究利用该矿爆破区块的生产历史大数据,提出了采用钻孔率(α)、炸药单耗(β)和块度指标(γ)计算区块爆破性指数K的方法,并根据爆破性指数K的值对历史爆破区块的可爆性进行分级;再以爆破区块的单轴抗压强度(UCS)、矿石的质量指标(RQD)和矿体的地质强度指标(GSI)作为可爆性指标,建立了可爆性指标与可爆性等级相对应的数据集;然后构建了深度学习神经网络模型,并以可爆性指标作为输入,以可爆性等级作为输出对构建的深度学习神经网络模型进行了训练;最后通过现场试验验证了训练后的模型对可爆性等级预测的可靠性和准确性,同时优化了爆破设计和爆破效果。研究结果表明:建立的深度学习神经网络模型可用于爆破区块的可爆性分级与爆破效果优化。

Husab Uranium Mine is a super-large-scale open-pit uranium mine.Currently,the mine adopts a“one-time design,long-term use”approach to blasting production,leading to issues such as a lack of dynamic adjustment of blasting parameters,high explosive consumption,and unsatisfactory blasting results.To address these issues,a solution can be achieved through dynamic blastability classification management of blasting blocks and feedback-controlled blast design.This study utilizes the production history big data of the mine′s blasting blocks.It proposes a method to calculate the blasting index K using drilling rate(α),explosive consumption per unit volume(β),and fragmentation index(γ).Here,αrepresents the drill hole cross-sectional area per unit area,where a higher value indicates more drilling required and higher drilling costs.βrepresents the amount of explosives required per unit volume of crushed rock,where a higher value implies a more significant amount of explosives required and higher blasting costs.γrepresents the distribution of fragment size after ore blasting,where a higher value indicates worse blasting effects,higher transportation costs,and greater difficulty in blasting.Based on the value of the blasting index K,the blastability of historical blasting blocks is classified into different levels.Uniaxial compressive strength(UCS)of the blasting blocks,rock quality designation(RQD)of the ore,and geological strength index(GSI)of the ore deposit are used as blastability indicators,establishing a dataset correlating blastability indicators with blastability levels.The dataset consists of 69 sets of historical data,with 20 sets classified as level one(easily blastable),24 sets as level two(relatively difficult to blast),and 25 sets as level three(difficult to blast).Subsequently,a deep learning neural network model is constructed,comprising an input layer,five hidden layers,a dropout layer,and an output layer.The model is trained using blastability indicators as inputs and blastability levels as outputs.The traditional SVM model is used for comparison,revealing that the trained deep learning neural network model achieves higher prediction accuracy on the test set than the traditional SVM model.Finally,the reliability and accuracy of the trained deep learning neural network model in predicting the blastability level of blasting blocks are verified through on-site experiments,optimizing the blast design and blasting effects.The research findings indicate that the trained deep learning neural network model,based on a large amount of historical production data from Husab Uranium Mine,can be used for blastability classification of blasting blocks and optimization of blasting effects.