数据降维算法在铀矿堆浸工艺中的应用研究

发布时间：2018-03-24 20:47

本文选题：铀的累计浸出率　切入点：支持向量机　出处：《东华理工大学》2016年硕士论文

【摘要】：铀矿生物堆浸工艺是将铀矿堆浸和生物浸出相结合的一项技术,它不仅保留了生物浸出技术的特点,还兼具铀矿堆浸工艺的优越性。因此,生物堆浸工艺凭借其良好的经济和环境效益,现如今已经发展成为世界很多国家生产铀矿的支撑性技术。目前,我国的铀矿石呈现出类型复杂、中低品位居多等特点,而且已不能满足我国核电长远发展的需要。因此,为了完善铀矿生物堆浸工艺的理论研究与实践应用,采用一些数据挖掘方法建立数学模型,并借助统计分析软件进行仿真实验对生物浸铀工艺进行研究具有重要的意义。为此,本文选取某矿区ZQ7堆浸柱在铀矿生物堆浸工艺中产生的一批实测样品数据作为样本,建立了以下几种数学模型对其进行研究分析:(1)构建BP神经网络模型对铀的累计浸出率进行仿真实验。为了排除划分样本的偶然性,本文对样本进行了不同方式的划分。实验表明,该模型具有较好的模拟效果,能满足一般实际生产的需求。其中,前53个样本作为训练集,后15个样本作为测试集的划分方式具有最好的模拟效果,其均方根误差RMSE=0.7012,平均相对误差MRE=0.491823;(2)建立基于主成分分析(PCA)的支持向量机(SVM)模型(PCA-SVM)对铀的累计浸出率进行仿真实验。由于生物堆浸工艺中各因素间存在着相互关联,所以采用PCA提取出累计贡献率大于85%的三个主成分作为SVM模型的输入变量。在实验过程中利用网格搜索、遗传(GA)算法、粒子群(PSO)算法对参数c和g进行优化。实验表明,经PCA特征提取比未经PCA特征提取的模型具有更好的模拟效果;在三种寻优算法中,PSO算法有最好的搜索效率;PCA-PSO-SVM模型的模拟效果最好,其结果为:均方根误差RMSE=0.0673,平均相对误差MRE=0.000793;(3)将支持向量机作为分类器,建立基于平均影响值(MIV)的SVM模型(MIV-SVM)选择特征子集。基于MIV算法可对特征因子进行排序,以此选取得到最高精度时的前7个特征,即Eh出,p H出,Fe2+出,Fe2+进,Fe3+进,Fe3+出,浸出液体积,将这7个特征作为SVM模型的输入变量,其模拟效果优于原始10个特征的效果,其结果为:均方根误差RMSE=0.3519,平均相对误差MRE=0.004173;(4)将支持向量机作为分类器,建立基于离散二进制粒子群(BPSO)算法的SVM模型选择特征子集。最终BPSO算法筛选出最优的特征子集包含:喷淋强度,Eh出,Fe2+进,Fe2+出,U出共5个特征,该模型的模拟结果为:均方根误差RMSE=0.3332,平均相对误差MRE=0.003985;(5)由于MIV-SVM和BPSO算法有其不同的优势及局限性,因此,可建立组合算法对样本进行模拟分析。采用MIV-SVM算法对特征进行排序,快速去掉无关特征,然后以排序后的优良子集初始化后续BPSO算法的部分种群,使其有一个好的搜索起点。最终MIV-SVM-BPSO模型的模拟结果为:均方根误差RMSE=0.3071,平均相对误差MRE=0.003528,其模拟效果均比单一的MIV-SVM和BPSO模型好。
[Abstract]:Bio-heap leaching of uranium ores is a combination of heap leaching and bioleaching of uranium ores. It not only retains the characteristics of bioleaching technology, but also has the advantages of heap leaching of uranium ores. Because of its good economic and environmental benefits, biological heap leaching technology has been developed into a supporting technology for uranium ore production in many countries in the world. At present, uranium ores in China are characterized by complex types, medium and low grade, and so on. Therefore, in order to perfect the theoretical research and practical application of bioreactor leaching process of uranium mine, some data mining methods are used to establish mathematical models. It is of great significance to study the bioleaching process with the help of statistical analysis software. In this paper, a batch of measured sample data produced by ZQ7 heap leaching column in uranium mine are selected as samples. The following mathematical models are established to study and analyze them. (1) A BP neural network model is constructed to simulate the cumulative leaching rate of uranium. In order to eliminate the contingency of dividing samples, In this paper, samples are divided in different ways. Experiments show that the model has good simulation effect and can meet the requirements of general production. Among them, the first 53 samples are used as training sets. The latter 15 samples have the best simulation effect as the partition of the test set. The root mean square error (RMSE) is 0.7012, and the mean relative error (MREE) is 0.491823 / 2). A support vector machine (SVM) model based on principal component analysis (PCA) is established to simulate the cumulative leaching rate of uranium. Therefore, three principal components whose cumulative contribution rate is more than 85% are extracted by PCA as input variables of SVM model. The parameters c and g are optimized by using grid search, genetic algorithm and particle swarm optimization algorithm. PCA feature extraction has better simulation effect than the model without PCA feature extraction, among the three optimization algorithms, it has the best search efficiency and the PCA-PSO-SVM model has the best simulation effect. The results are as follows: the root mean square error (RMSE) is 0.0673, and the mean relative error (MREE) is 0.000793K3). The support vector machine (SVM) is used as the classifier, and the SVM model based on the average influence value (MIV) is established to select the feature subset. Based on the MIV algorithm, the feature factors can be sorted. In this way, the first seven features of the SVM model are selected, that is, the first seven features, i.e., the volume of the leaching solution, the volume of the leachate, the simulation effect of which is better than the original 10 characteristics, and the simulation effect is better than that of the original 10 features. The results are as follows: root mean square error (RMSE) 0.3519, mean relative error (Mre) 0.004173 / 4) support vector machine (SVM) is used as classifier. The SVM model selection feature subset based on discrete binary particle swarm optimization (BPSO) algorithm is established. Finally, the optimal feature subset is selected by BPSO algorithm. The optimal feature subset consists of five features: spray intensity, Eh, Fe2, Fe2, Fe2, and Fe2, respectively. The simulation results of this model are as follows: root mean square error (RMSE) 0.3332, mean relative error (Mre) 0.003985 / 5) because MIV-SVM and BPSO have different advantages and limitations, a combined algorithm can be established to simulate and analyze samples. MIV-SVM algorithm is used to sort the features. Quickly remove irrelevant features, and then initialize some populations of the subsequent BPSO algorithm with a sorted fine subset. The simulation results of MIV-SVM-BPSO model are as follows: root mean square error (RMSE) 0.3071, average relative error MREE 0.003528. The simulation results are better than that of single MIV-SVM and BPSO models.
【学位授予单位】：东华理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TD958

【相似文献】