【作者】 陈广洲；

【导师】 徐晓春；

【作者基本信息】 合肥工业大学 ， 矿物学、岩石学、矿床学， 2010， 博士

【摘要】 现代矿产勘查过程中会产生地质、地球物理、地球化学、遥感等大量专题信息,因此,多源信息成矿预测方法在成矿预测中占据重要地位。如何将地质理论和数学处理方法有机结合,充分挖掘多源信息中隐含的成矿信息,建立成矿预测综合模型,一直是地学界的研究热点和难点。皖南石台地区位于江南过渡带西段和长江中下游铜-铁-硫-金成矿带南侧,具有沿江铜-铁-硫-金成矿带和皖南多金属成矿带的过渡性特征,具有较好的成矿前景。但一直以来,该区地质勘查和研究程度较低,已发现的矿点和矿化点较少,给成矿预测带来困难。基于此,从地质、物探、化探和遥感等多源信息角度对该区的铜多金属矿进行预测研究,对指导该区的地质找矿具有重要的实际意义。论文结合安徽省省级地质勘查专项(AHGTT2006-4),选取皖南石台地区,以铜多金属矿预测为例,进行多源信息耦合的成矿预测研究。针对模型各自的特点,采用多种传统预测模型(综合信息找矿模型、专家证据权重法、信息量法、BP人工神经网络法)进行成矿预测研究,并进行预测结果比较;随后,以前述模型为基础,率先应用两种能处理复杂变量关系的新型模型(投影寻踪模型、支持向量回归机模型)进行成矿预测,并进行比较以检验两种新模型的适用性和先进性。论文主要研究内容和取得的成果如下:1、综合研究区域地质、地球物理、地球化学等多源信息资料,分析已有矿点(矿化点)的成矿规律特征,构建该区铜多金属矿的综合信息找矿模型。选取7个预测变量(地层、断裂、岩浆岩、围岩、航磁、化探、遥感),制作上述专题图层,根据专家知识设定变量图层权重,利用GIS得到综合信息成矿潜力。根据数据特征,划分四级成矿有利区,圈定找矿靶区,得到成矿预测专题图。经过对圈定靶区的成矿有利程度分析,结果表明该方法是一种简便、有效的预测方法,其预测结果是适用的和有效的,可以为进一步的勘探工作提供依据。2、采用上述综合信息找矿模型所用数据,使用DPIS软件,应用专家证据权重法进行成矿预测。该法可以克服数据驱动型证据权重法对数据量要求多的不足。制作上述各变量的shp格式文件,生成单元格网,根据专家知识设定权重(同综合信息模型权重),得到后验概率,依累积后验概率分布特征划分四级成矿有利区,圈定靶区,得到专题图。分析结果表明:该法所得结果与综合信息找矿模型所得结果基本是一致的。3、采用客观成矿预测模型-信息量法进行预测以避免专家的主观因素对成矿预测的影响。信息量法的计算表明:对成矿贡献从大到小的顺序为:化探、地层、岩浆岩、断裂、遥感、围岩、航磁。根据单元格信息量的分布特征,划分四级成矿有利区,圈定靶区,得到专题图。结果显示信息量法的预测结果与综合信息找矿模型、专家证据权重法模型的预测结果基本是一致的。4、本文提出了一种新的成矿预测方法:信息量-专家证据权重耦合的预测方法。为充分吸收来自主、客观方面的知识,该法利用信息量法获得的客观权重和来自专家的主观权重,采用D-S证据理论进行权重的融合,获得新的综合权重,以此权重为基础,使用专家证据权重法进行成矿预测。预测结果表明该法可以获得更为满意的预测效果。5、采用处理非线性能力强的人工神经网络方法进行预测。依据上述数据,根据信息量法的计算结果,并参考综合信息找矿模型中权重的赋值情况,构建了训练BP人工神经网络所需的128个样本。训练样本采用1、0两种类型数据,分别表示单元内某个地质条件的存在和不存在。采用MATLAB软件,编制程序实现BP人工神经网络的训练和预测。根据预测数据特征,划分四级成矿有利区,圈定靶区,得到专题图。分析表明预测结果与先前模型的预测结果也是基本一致的。此外,针对人工神经网络的训练容易陷入局部极值的不足,本文采用群智能优化算法-粒子群算法进行神经网络的权值、阈值的优化以便提升神经网络的泛化能力,并进行了预测效果的比较。6、率先采用一种能处理各变量间复杂非线性关系的降维处理新方法-投影寻踪插值模型,将其应用于成矿预测领域。采用上述的128个样本,本文提出一种改进的人工蚁群算法用于优化计算求得模型的最佳投影方向,以此方向为基础,计算未知单元的投影值。根据投影值的分布特征,划分四级成矿有利区,圈定靶区,得到专题图。分析表明该模型所得预测结果与前述方法的结果是基本一致的,该方法是有效的。7、总结成矿预测的过程,可以将其看作是一模式识别问题。本文率先采用一种模式识别的新方法-支持向量回归机模型,将其应用于成矿定位预测。模型的训练样本同上,分别采用网格搜索法和基于群体智能优化的方法-粒子群算法来搜索最佳的模型参数。根据最佳模型参数,计算未知单元的预测值。根据其分布特征,划分四级成矿有利区,圈定找矿靶区,获得专题图。预测结果表明在大多数的成矿位置上同前面模型的结果一致,同时有自己的特点。8、为检验能处理非线性复杂关系的三种模型(人工神经网络、投影寻踪、支持向量机)的有效性,从128个样本中等间隔选取共16个样本组成小样本,并进行建模实验。结果表明:针对小样本情况,投影法的预测结果较好,支持向量机一般,神经网络的预测结果较差。更多还原

【Abstract】 In the process of modern mineral exploration, the massive special information is produced from geology, geophysics, geochemistry, remote sensing and so on, therefore, metallogenic diagnosis based on multi-source information holds important status. It has been ever since one of geo-science research focuses and difficulties to effectively combine specialized geological theory knowledge with mathematical methods so as to fully extract the useful metallogenic information from multi-source data to establish the integrated finding mineral model. Shitai region of South Anhui province, which is located at the western section of South of Yangtze River transitional zone and the south side of Cu-Fe-S-Au mineralization belt of the middle and lower reaches of Yangtze River, owns transitional features of Cu-Fe-S-Au mineralization belt along the Yangtze River and polymetallic mineralization belt in South Anhui province with promising metallogenic prospects. However, up to now, it has been difficult for the mineralization prediction because of poor geological exploration and few mineralization points. Therefore, it is of great significance to make the mineral prediction for the copper-polymetallic deposits from the views of multi-source information on the basis of mentioned above.Taking the example of copper-polymetallic deposits prediction, the dissertation deals with metallogenic diagnosis coupled by multi-source metallogenic information by selecting Dingxiang district of Shitai county in South Anhui province on the basis of the geological exploration special project of Anhui province (AHGTT 2006-4).Firstly, the metallogenic diagnosis is made by a variety of conventional prediction models (integrated finding mineral model, expert weight of evidence model, prospecting-information contents, BP artificial neural networks). Secondly, two new models that can cope with complicated relationships among variables are used to do that.The main contents and achievements obtained in the dissertation are as follows:1. An integrated finding mineral model of copper-polymetallic deposits is constructed by synthetically studying regional multi-source information from geology, geophysics, geochemistry and so on, and analyzing the metallogenic laws of mineralization points found. Seven forecast variables (stratum, fault, magmatic rock, adjoining rock, aeromagnetic, geochemical exploration, remote sensing) are selected to make the thematic layers of variables. According to expert knowledge, thematic layers of variables are given the corresponding weights by means of their importance in metallogenic diagnosis, and then comprehensive mineralization potential information of the research area can be obtained by using GIS. On the basis of prediction data, topic chart of mineralization forecast is obtained by dividing metallogenic favorability area into four grades and enclosing perspective zones. By the analysis of enclosed perspective zones, the prediction result shows the model is suitable and effective. In addition, the method is a convenient, effective forecast technique and can provide the basis for further exploration work.2. On the basis of data used by the integrated finding mineral model and DPIS software, expert weight of evidence model is used to carry out metallogenic prediction, which can overcome the shortage of data-driven weight of evidence model that needs more data. Firstly, documents of variables for the shp form are made and unit grid meshes are produced. Secondly, weights of variables according to the expert knowledge are set. Finally, topic chart of mineralization forecast is obtained on the basis of the posterior probability of every unit grid mesh and the curve of accumulation posterior probability with metallogenic favorability area divided into four grades and perspective zones enclosed. The prediction result shows that the result of model is basically consistent with that of integrated finding mineral model.3. The objective model of metallogenic diagnosis-prospecting-information contents is used in mineralization forecast to avoid the subjective influence from experts. Its computation result indicates that the mineralization contribution descending order is geochemical exploration, stratum, magmatic rock, fault, remote sensing, adjoining rock, aeromagnetic. In the light of distribution characteristic of prospecting-information contents, the corresponding topic chart can be obtained with metallogenic favorability area divided into four grades and perspective zones enclosed. The prediction result shows that the result of model is basically consistent with that of integrated finding mineral model and expert weight of evidence model.4. A new method of metallogenic diagnosis is proposed in the dissertation; that is, forecast technique based on the coupling of prospecting-information contents-expert weight of evidence. To make full use of knowledge from the subjective and objective aspects, it can obtain new comprehensive weights by fusing subjective weights which come from experts and objective weights which come from prospecting-information contents on the basis of D-S evidence theory. Subsequently, expert weight of evidence is used according to comprehensive weights. The forecasting result indicates that this method can obtain more satisfactory effect. 5. The prediction is carried out by artificial neural network that has strong ability of treating the non-linear relationship among variables in the paper. According to the above data, 128 samples to train BP artificial neural network are produced, considering the analysis result of prospecting-information and the variables’weight information of integrated finding mineral model. The training samples are used according to 1, 0 two types which express geological condition’s existence and no existence in the units respectively. MATLAB software is employed to realize the training and forecasting of BP artificial neural networks by programming. According to the distributed characteristics of forecasting values, topic chart of mineralization forecast is obtained with metallogenic favorability area divided into four grades and perspective zones enclosed. Analysis shows forecasting result is basically consistent with that of former models. In addition, aimed at the fault of easily falling into the local extremum in the training, the dissertation uses swarm intelligence optimization algorithm-particle swarm algorithm to carry out optimization of the weight and the threshold values of the neural network so as to promote its generalization ability, and makes the comparison on the prediction effect.6. The dissertation for the first time applies a new method-projection pursuit interpolation model which deals with the complex relationship among variables and reduces the dimensions into mineralization forecast. On the basis of 128 samples, an improved ant colony algorithm is presented to compute the best projection direction, by which the unknown projection values of every unit can be computed. According to the distributed characteristics of projection values, topic chart of mineralization forecast is obtained with metallogenic favorability area divided into four grades and perspective zones enclosed. Analysis shows forecasting result is effective and basically consistent with that of former methods.7. Studying the course of metallogenic prediction, it can be regarded as a pattern recognition problem. Therefore, a kind of new pattern recognition method-support vector regression model is used for the first time to mineralization forecast. Training samples are the same as the above models, the best model parameters are investigated through the grid method and swarm intelligent optimization algorithm-particle swarm algorithm, by which predicted values of units can be computed. According to their distributed characteristics, topic chart of mineralization forecast is obtained with metallogenic favorability area divided into four grades and perspective zones enclosed. The result indicates that this model has its own typical characteristics, besides, is consistence with former forecasting results for most locations.8. To test the validity of three models that can treat complex nonlinear relationship among variables, 16 samples out of 128 ones are selected and modeled in equidistance as sub-samples. The result indicates projection pursuit model is better, support vector regression model moderate, and neural network worse.更多还原