The Application Research Of Dimensionality Reduction Algorithms On Heap Leaching Process Of Uranium

Uranium biological heap leaching process is a technology that combined uranium heap leaching and bio-leaching. It not only retains the characteristics of bio-leaching technology, but also possesses the superiority of uranium ore heap leaching technology. Thus, biological heap leaching process now has developed into a uranium production supporting technology in many countries of the world by virtue of its remarkable economic and environmental benefits.At present, uranium ore in our country has many characteristics, such as complex types, low and medium grade in the majority. It can not meet the needs of China’s long-term development of nuclear power. Therefore, in order to improve the theoretical research and practical application about uranium biological heap leaching process, using some data mining methods to establish the mathematical model, the simulation experiment with the help of statistical analysis software will have a great significance for the study of biological leaching of uranium. So, in this paper, we take a set of sample data which produced in the uranium ore biological heap leaching process by the ZQ7 heap leaching column as samples, and establish a few mathematical models to research the biological heap leaching process:(1) Construct BP neural network model to do simulation experiments about uranium accumulative leaching rate. To rule out the contingency in dividing samples, we divide the samples in different ways. Experiments show that the model has good simulation effects, and it can meet the general needs of production in actual. Among the divisions, the one which chooses the first 53 samples as a training set and the after 15 samples as a test set has the best simulation effect, the root mean square error is 0.7012, the mean relative error is 0.491823;(2) Construct the support vector machine(SVM) model based on principal component analysis(PCA), called PCA-SVM for short, to do simulation experiments about uranium accumulative leaching rate. Because the factors in biological heap leaching process are interrelated, so we use the method of PCA to extract three principal components among them, whose contribution rates are more than 85%, as the input variables of SVM model. In the course of the experiment, we use grid search, genetic algorithm(GA), particle swarm optimization(PSO) to optimize the parameter c and g. Experimental results show that using PCA to extract features has a better simulation effect than before. Among the three optimization algorithms, PSO algorithm has the best search efficiency. The PCA-PSO-SVM has the best simulation results: the root mean square error is 0.0673 and the mean relative error is 0.000793;(3) Selecting the support vector machine as classifier, construct the support vector machine model based on mean impact value(MIV-SVM), to choose character subsets. The MIV algorithm can sort the factors and select the first 7 factors of the highest accuracy, which are Eh出, p H出, Fe2+出, Fe2+进, Fe3+进, Fe3+出, leachate volume. Take the 7 factors as the input variables of SVM model, which simulate the effect is better than the original 10 features, results are that the root mean square error is 0.3519 and the mean relative error is 0.004173;(4) Selecting the support vector machine as classifier, construct the support vector machine(SVM) model based on discrete binary particle swarm optimization(BPSO) to select character subset. The BPSO can screen the optimal character subset: spray intensity, Eh出, Fe2+进, Fe2+出, U出 five factors. The simulation results of the model are: the root mean square error is 0.3332 and the mean relative error is 0.003985;(5) Since the MIV-SVM algorithm and BPSO have their different advantages and limitations, therefore, we can establish combined algorithm to simulate and analyze the samples.MIV-SVM algorithm can sort the characters and remove the irrelevant characters.Then use the better subset after sorted to initialize part of the populations in the following BPSO algorithm, it can get a better starting point for the search. The simulation results of MIV-SVM-BPSO are: the root mean square error is 0.3071, the mean relative error is 0.003528. Its simulation effects are better than single MIV-SVM model or BPSO model.