基于交会图和多元统计法的神经网络技术在火山岩识别中的应用

针对火山岩储层的特殊性(复杂性、离散性和随机性),应用BP神经网络技术对火山岩测井解释中岩性识别问题进行了研究。该方法的技术关键是样本集和初始权重的建立,以及模型的优选。本文提出了一种基于交会图和多元统计法的学习样本生成方法,即根据取心岩样的地球化学和岩石学研究,用交会图技术建立样本集,采用聚类分析和距离判别法确定初始权重。将研究方法应用在松辽盆地杏山地区火山岩岩性识别问题中,取得了很好的效果,岩性解释符合率高于90%。文中通过四种岩性识别处理模式的对比研究,表明赋权重处理模式为最优处理模型。在神经网络模型预测过程中,需充分利用已有的地质经验和测井曲线信息建立典型可靠的样本文件,同时考虑神经网络方法中各种因素的影响,优选模型和计算参数才能使预测结果符合实际情况。

Application of neural networks technique based on crosspiot and multielement statistics to recognition of volcanic rocks

The study has been carried out for application of BP neural networks technique to lithologic recognition in logging interpretation of volcanic rocks in view of the speciality of volcanic reservoir (complexity. discreteness and stochastic property). The technical key of the method is to create the sample set and initial weight as well as to optimize the model. The paper presented a generation method of study samples based on crosspiot and multielement statistics, that is, building up sample set based on the study of geochemistry and lithology of cored samples and using cluster analysis and distance criterion to determine the initial weight. Application of studied method to lithologic recognition of volcanic rocks in Xin mountainous area of Songliao basin achieved good effects. The coincident rate of lithologic interpretation reaches to 90% and above. It is shown by correlation of four lithologic recognition patterns in the paper that weighted processing pattern is optimized processing model. In a predicting process of neural networks model,it needs to fully use available geologic experiments and logging information to create typical and reliable sample file and consider the influence of various factors in neural networks at the same time,so that the optimized model and calculated parameters can make predicted results be coincident with the objective reality.