【作者】 缪燕子；

【导师】 马小平；

【作者基本信息】 中国矿业大学 ， 控制理论与控制工程， 2009， 博士

【摘要】 煤矿瓦斯灾害是我国主要灾害之一,矿井瓦斯突出预测是一个十分复杂的理论和实验技术课题,原因在于瓦斯突出是一个包含地质学、物理、化学等多学科、多种效应交叉的复杂现象,并且对于这一灾害缺少有效的监测方法和手段。论文针对瓦斯突出这一不确定性和非线性灾害问题,建立了多传感器信息融合瓦斯突出预测系统模型,并分别对融合系统的数据层、特征层和决策层进行了分析和研究,构建了一个基于多规则决策的瓦斯突出智能预警系统,从而有效地提高对瓦斯突出预测的准确度。论文通过对多传感器信息融合一般框架结构分析,提出了基于多传感器信息融合的瓦斯突出预警系统结构,以及特征级和决策级分层融合模型,并通过多传感器管理子系统实现反馈建立预测系统闭环控制模型。在分析目前已有的突出指标及临界值确定依据的基础上,引入层次分析法得到几个典型突出指标的权重排序,对瓦斯突出多传感器融合预测系统的数据来源信息进行分析。并根据选取的重要影响指标选择瓦斯突出监测传感器,由于静态传感器定点布置存在一定的局限性,提出了利用主动嗅觉技术研究瓦斯监测动态传感器,以增加多传感器预测系统的实时、动态数据源。分析多传感器信息融合各层次算法的优缺点,选取人工神经网络作为特征层分层融合算法。通过瓦斯突出实例,对比分析了三种改进的BP神经网络模型,实验数据表明三种改进的BP神经网络能有效解决传统BP网络收敛速度慢和易陷入局部极优的缺陷,从而有效改善融合的效果和速度。基于神经网络固有的缺点,进一步提出用Dempster-Shafter (D-S)证据理论作为决策级融合方法,构成特征级和决策级的分层融合结构,增加决策的可靠性。D-S证据理论以其优越的不确定性推理成为很好的决策融合算法,论文对决策层D-S证据理论进行了深入研究,总结了证据理论存在的主要问题,并提出了对应的解决办法。特别针对D-S证据理论中证据冲突的问题,引入证据距离、证据源可信度等概念,提出了一种新的合成规则。该规则把冲突信息按证据源的可信度进行分配,对一致性证据采用反映聚焦程度的与运算。理论分析和数值实验表明本文提出的合成规则对高冲突性证据和一致性证据都非常有效,能够解决多传感器信息来源之间的证据冲突问题。为了定量表示实际应用中的模糊概念和数据,论文研究了D-S证据理论推广到模糊集理论的方法。总结了目前比较典型的证据理论向模糊集推广的方法和结论。针对前人的不足,给出了一种新的模糊集合之间相似度的定义,并进一步提出一种有效组合模糊证据理论的决策层融合算法,并给出其数学证明。通过实例验证了该算法能更有效地获取模糊焦元的变化信息,具有更好的融合效果。论文选取五个典型高瓦斯矿区突出数据,验证了本文提出的多传感器信息融合瓦斯突出预测系统模型的可行性和有效性;并根据改进的D-S证据理论融合规则,采用Windows Vista sp1 + Visual Studio 2008 sp1系统开发平台实现了矿井瓦斯突出智能预警系统,为现场工程技术人员提供实时决策辅助支持。经过矿区现场数据验证该瓦斯突出多规则决策融合系统能够提供准确、可靠的灾害预测,有效地提高了煤矿安全管理水平。更多还原

【Abstract】 Gas outbursts in mines are among the main disasters in China. Gas outburst prediction is a complex theoretical and technical project, as the phenomenon of gas outbursts results from multiple effects of a geological, physical and chemical nature, and there is a lack of effective monitoring methods for this disaster. For resolving the seeminlgy unpredictable and non-linear problem of gas outbursts, a new model of a gas outburst prediction system based on multi-sensor information fusion is put forward, and the data level, feature level and decision level is analysed and studied respectively. And an intelligent decision prediction system is presented, the example results of this system show that it could significantly increase the accuracy of gas outburst prediction.After summarizing the general architecture of multi-sensor information fusion, this work puts forward a new model of a gas outburst prediction system. It is based on multi-sensor information fusion, which is described as hierarchical fusion in the feature fusion level and the decision level. And it is a closed contrl model through the feedback of sensor management system.Based on analysing the current indexes and their critical values which were put forward by different researchers, this thesis presents a hierarchical analysis method for arranging the main typical influencing factors in the order of their significance and chooses the most suitable sensors according to the most important factors. After analysing the limitations of local static sensors, a real-time dynamic sensor is studied using the active olfaction technology.As the merits and demerits of each multi-sensor information fusion method, the Artificial Neural Network (ANN) is chosen as feature fusion algorithm. Three improved BP neural network could improve the learning velocity and capability of convergence of traditional BP neural network, which is validated by example data analysis of gas outbursts. As the inherent disadvantages of ANN, the Dempster-Shafter (D-S) evidence theory is presented for decision fusion, which could increase the credibility of the decision-making.As one of the more effective uncertain reasoning methods among the information fusion methods, the D-S evidence theory has proven to be a good decision fusion algorithm. This work sums up the main deficiencies of the D-S evidence theory, and gives the corresponding solutions. It puts forward new combination rules of the D-S Evidence Theory for resolving the problem of evidence conflicts based on introducing the distance between the evidence, the credibility of evidence sources, and other notions, and also provides the mathematical proofs for these rules. The improved rules allocate the conflicts to various focal elements according to the credibility of the coherence evidence. The AND-algorithm is adopted to combine the coherence evidence, which reflects the intersection of focus elements. The corresponding mathematical and theoretical analysis proves that the given rules are rational and effective for both highly conflicting and coherent evidence, and that they can deal with conflicting information from different sensors.For combing the D-S evidence theory and the fuzzy set theory, typical current methods and conclusions of extending the evidence theory to fuzzy sets are summarized. Pointing out the disadvantages of previous methods, this paper introduces a new definition of the similarity degree between fuzzy sets, and also a decision-making fusion algorithm which can combine fuzzy evidence effectively.Mathematical proofs and example analyses validate the new algorithm and demonstrate that it is more effective, acquires more information on the change of fuzzy focal elements, and produces better fusion results than existing methods. In order to validate the model of the gas outburst prediction system based on multi-sensor information presented in this thesis, the gas outburst data of five typical mines are chosen. According to the improved combination rules of the D-S evidence theory, an intelligent prediction system of gas outbursts is developed on a Windows Vista +Visual Studio 2008 software platform which provides real-time decision support for underground workers in coal mines. The results from the analysis of our gas outburst multi-regular decision fusion system show that the system supplies accurate and credible disaster prediction, which effectively improves the level of coal mine safety management.更多还原