【作者】 许满贵；

【导师】 徐精彩； 葛岭梅；

【作者基本信息】 西安科技大学 ， 安全技术及工程， 2006， 博士

【摘要】 论文在国家杰出青年基金项目《采矿环境与安全(50125414)》等纵向科研项目和大量煤矿安全现状综合评价项目研究与实践基础上，综述了国内外煤矿安全评价理论与技术的研究现状，建立了煤矿安全动态评价和预测模型，并结合现场数据实现了应用研究。论文还综合模糊数学和神经网络技术的优点建立了煤矿安全评价的模糊神经网络模型，增强了对不确性指标的表达能力，使评价模型具有更强的自学习、并行计算、全局寻优和复杂数据的处理能力，得出一套符合煤矿生产系统特点的动态(变权)安全评价和预测技术、方法。首先，论文在分析我国煤矿安全现状的基础上探讨了开展煤矿安全评价工作的重要意义，综述了国内外安全评价理论和煤矿安全评价技术的发展现状、神经网络技术的特点，提出课题研究的意义、研究思路和主要内容。然后，根据现代事故致因理论，结合人-机-环分析法与层次分析法等分析了煤矿安全生产的主要影响因素，并总结为11大类。根据安全评价指标体系建立的原则，提出几种煤矿安全评价指标的预处理方法，建立了煤矿安全评价指标体系。接着，根据神经网络结构特点和已经建立的煤矿安全评价指标体系，确定以误差反向传播的前向BP网络作为煤矿安全评价算法模型，并探讨了煤矿安全评价的网络结构设计、训练学习流程、性能改进方法。讨论了MATLAB神经网络工具箱及其图形用户界面GUI在神经网络模型的设计和训练过程中的强大功能，为煤矿安全评价网络模型的应用奠定基础。利用神经网络工具箱GUI实现了煤矿动态(变权)安全评价的神经网络模型设计，并结合大量的现场实际数据实现了煤矿安全动态评价的应用研究，评价结果与“煤矿安全现状综合评价报告”结果基本一致。论文还对煤矿安全评价神经网络模型作了重要改进，综合模糊数学和神经网络技术的优点建立了煤矿安全评价模糊神经网络模型，并完成了应用研究，使评价过程对不确定性指标的表达能力更强，具有更强的自学习、并行计算、全局寻优和复杂数据处理能力。煤矿安全预测也是煤矿安全动态评价的一部分，论文建立了煤矿安全指标(百万吨死亡率)预测的神经网络模型，并借助新疆焦煤集团某矿井的实际数据完成了应用研究，比较了三种网络结构、三种训练函数的预测过程及结果的准确度，得出三种网络结构中6×18×1三层网络模型的预测结果更稳定和准确，TRAINBR和TRAINLM函数在矿井安全动态预测中有明显的优势。在现场实践的基础上，适应煤矿安全评价神经网络模型需要，论文中编制了煤矿安全评价原始数据采集表，并制定了指标量化标准，有助于动态采集并量化数据，便于该安全评价技术的推广应用。更多还原

【Abstract】 Based on the distinguished young scholar project mining environment and safety, which wasfunded by the National Natural Science committee, this paper has been studied for the sake ofscientific, convenient and adaptable mine safety evaluation methods.Firstly, in the paper, in order to put forward the importance and necessity of safety evaluation,Chinese coalmine safety status is analyzed. The domestic and foreign safety evaluations aresummarized to analyze the problem of traditional coalmine safety evaluation methods, andbring forward Artificial Neural Networks (ANN) to deal with many problems, such asnonlinear, changeable weights, etc.Secondly, based on the Accidental Incidence Theory and other safety theories, the AnalyticHierarchy Process are combined with other methods to analyze the primary factors affectingcoal mine safety; furthermore, the processes are classified into 11 types including human,mechanical, environmental factors. And the evaluation index pretreatment methods are givento establish the coal mine safety evaluation index system.Thirdly, because ANN are adaptive models that can be learned from the data and generalizethings. Especially in contrast to traditional models, which are theory-rich and data-poor, theneural networks are data-rich and theory-poor in a way that a little or no prior knowledge ofthe problem is present. Neural networks (NN) can be used for coalmine safety evaluationfrom inputs to outputs of these kinds of black boxes. In this paper, the multilayer model isestablished, which learn using an algorithm called back propagation.Fourthly, the coal mine safety evaluation BP model is designed by the use of the NN tool boxof the MATLAB software, trained by the means of safety sample data of 33 mines to provethat the ANN model is applicable for coal mine safety evaluation. Moreover, the process and characteristics of coal mine safety evaluation BP model are analyzed to put forth the new method(FNN) integrating the fuzzy mathematics and ANN, which has the same nonlinear and changeable weights processing function as ANN, and the same functions of using expertise and little requiring sample data as the fuzzy mathematics. The coal mine safety evaluation FNN model has been set up and applied to 10 mine sample data.At last, safety forecasting model are also founded by the MATLAB software. By comparing three types of NN models and training functions, it is proved that the TRAINDM learning rules continue for too long; but the TRAINBR can minimize the error than TRAINLM and not continue such length as TRAINDM. Therefore, TRAINBR maybe is the most adaptive learning rule for coalmine safety forecasting. Besides, if the forecasting model is so simple, it is hard to meet with the precise requirements.The paper has two creative points: one is that the FNN are firstly used to evaluate coal mine overall safety conditions; the other is that three types of NN structures and training functions are comparably studied to prove that the TRAINBR function maybe is the more adaptive learning rule of coalmine safety forecasting than the TRAINLM function, and if the forecasting model is so simple that it can not meet with the precise requirements.更多还原