【作者】 王晓路；

【导师】 刘健；

【作者基本信息】 西安科技大学 ， 安全技术及工程， 2011， 博士

【摘要】 为了明确煤矿井下瓦斯浓度、瓦斯涌出量、煤与瓦斯突出等预测器的输入主因素、降低其样本的复杂度并简化样本空间，为了确定预测器输入变量的不确定性对预测结果的影响，以及在其相关影响因素作用程度发生改变的情形下还能够进行准确的预测，本文开展了下列工作：为了识别瓦斯浓度、瓦斯涌出量等预测器的输入主因素，提出一种与预测精度相关的预测器输入主因素的遴选算法。在一定显著水平下对增添或删除若干因素前后预测器的预测残差进行F检验，用以确定具有改进作用的增添或删除操作。遍历增添和删除的所有情形后，即可确定能获得预测器在该地质、环境等条件下的最佳输入主因素，即预测性能最大改进的主因素输入组合。针对煤矿井下瓦斯浓度、瓦斯涌出量等时间序列往往呈现出较强的随机性和复杂性的情形，提出一种基于小波变换的瓦斯浓度、瓦斯涌出量等时间序列的预测方法。通过小波分解降低时间序列的复杂度，将小波分解后各个分量转化为具有历史特征的新样本分别进行预测，所得到结果进行综合作为最终预测结果。用时间序列分解到小波函数空间（或尺度函数空间）上的能量作为尺度能量，依据尺度能量与满足预测精度的最大误差能量的比值关系提出了小波最佳分解级数的计算方法。基于粒子群优化算法计算得到了预测模型的优化参数。针对煤矿井下瓦斯分布具有模糊性和不确定性等特点，提出基于模糊C-均值聚类（FCM）的瓦斯浓度、瓦斯涌出量等预测方法，对样本数据进行FCM聚类分析，依据所得到的不同类别分别建立预测模型，并对样本空间的参数、预测器的参数和FCM的参数采用基于蚁群算法并以预测差残的F检验值作为适应值的方法进行了优化。为了确定瓦斯浓度、瓦斯涌出量等预测器输入变量的不确定性对于预测结果的影响，提出基于抽样盲数的估计方法，解决了相关因素任意概率分布特性的处理问题。提出相关因素的不确定性在预测过程中的传递算法、高阶抽样盲数降阶的计算方法以及输入因素不确定性下预测器输出结果的概率分布和置信区间的计算方法。为了在瓦斯浓度、瓦斯涌出量等的相关影响因素作用程度发生改变的情形下，还能够进行准确预测，提出了基于虚拟状态变量的、系统网络结构可以自适应的预测方法。用能够反映待预测对象的输入变量在模式识别网络输出空间上的特征响应作为虚拟状态变量。基于虚拟状态变量建立预测模型，对待预测对象进行预测。通过预测误差建立反馈机制，以此对模式识别网络的结构朝着符合待预测对象(如瓦斯涌出量)变化的方向进行调整，从而得到适合的虚拟状态变量，以适应相关影响因素作用程度的改变。基于本文的研究成果，将上述方法应用于一些实例，对煤矿井下瓦斯浓度、瓦斯涌出量和煤与瓦斯突出等进行预测，结果表明：预测性能改善显著，表明所建议的方法是可行的和有效的。更多还原

【Abstract】 The aim is to make clear the forecaster main input factors of coal mine underground gasconcentration, gas emission quantity, coal and gas outburst, etc, reduce the complexity of theirsamples and simplify the samples attributes. It is also for evaluation the influence of the inputfactors uncertainties on the forecasting results as well as accurate prediction the gas hazardeven when the effect of the related factors changes. The following observations have beenmade:For recognition the main input factors of gas concentration and gas emission quantity, ect,an approach related to prediction precision to select the main input factors of gas hazardforecasting is proposed. The necessary elimination or adding of input factors are determinedby F test for the forecasting variances before and after elimination or adding. Afterinvestigating on all cases, the most suitable input factor under the current geologicalenvironment and other conditions can be identified, that is the main input factor combinationof the most improved prediction performance.With regard to the situation that time series of gas concentration and gas emissionquantity often show strong randomness and complexity, an approach based on wavelettransform is put forward to predict gas concentration, gas emission quantity and other timeseries. Wavelet decomposition is used to reduce the complexity of time series, and the waveletcomponents are transformed into new samples with historical characteristics, which are usedin the prediction, respectively. The final forecasting results are obtained by combining all ofthe predicted results. The energy of time series decomposed into wavelet or scaling functionalspace is used as the scaling energy, based on which, the method calculation best waveletdecomposition level is presented according to the ratio of the scaling energy and the largestforecasting biases energy under the allowed prediction precision. The optimized parameters of forecasting model is obtained by a Particle Swarm Optimization based program.Concerning the distribution of coal mine underground gas having fuzziness, uncertaintyand other characteristics, an approach is suggested to forecast gas concentration and gasemission quantity based on Fuzzy C-means Clustering (FCM). FCM is introduced to classifythe samples into different categories. The corresponding forecaster is constructed for eachcategory, respectively. The parameters of samples space, predictor and FCM are optimizedbased on the ant colony optimization algorithm by using F test of the prediction residual asthe fitness value.In order to evaluate the influence of the input variables uncertainties on the forecastingresults of gas concentration and gas emission quantity, ect,, an estimation approach based onSampled-Blind-Number (SBN) is proposed, based on which, the problem of processing theprobability distribution with various types is solved. The transfer algorithm of the relatedfactors uncertainties in forecasting, the way of high order SBN converting lower order oneand the method of the probability distribution and the confidence interval of the forecastingresults with the input factors having uncertainty are put forward.To accurately predict the gas concentration and gas emission quantity even when theeffects of the related factors change, an approach based on the virtual state variables andself-adaptive network structure is proposed. The input variables with the capability ofreflecting the awaiting forecasting objec are mapped by a pattern recognition network, and theobtained characteristics responses of the output space are used as the virtual state variables.The awaiting forecasting objec is predicted by the virtual state variables based forecaster. Thestructure of the pattern recognition network is adjusted according with the change of theawaiting forecasting object such as gas concentration by the feedback mechanism basedforecasting biases. Consequently, the suitable virtual state variables which can adapt to thechanges of the related factors effect are obtained.Based on the research achievements in this paper, the above methods are applied in someexamples, and gas concentration, gas emission quantity, coal and gas outburst are predicted.The results show that the forecasting performance is remarkably improved indicating that theproposed approaches are feasible and effective.更多还原