Application Research Of Machine Learning Method Based On Nonlinear Mineralization Dynamic System

With the continuous advancement of the digitalization process,geological science needs to introduce new technologies and methods that can realize automatic data processing and analysis.The development of Machine Learning provides a new solution to different problems in geological big data.The dissertation studies the dynamic process of the nonlinear metallogenic system,verifies the dynamic behavior of the system and the nonlinear function relationship among the system variables,and applies the Machine Learning model to different target problems.The main research results are as follows:(1)The dissertation takes mineralization elements as research objects.The dynamic process of ore-forming elements is studied by phase space reconstruction technique;correlation dimension and reaction diffusion equation.The results show that the ore-forming elements in ore-forming fluids have chaotic dynamic behavior through aggregation and diffusion,and the attractors in phase space have stable fractal dimension.The stability of the system determines the specific form of the attractor in the phase space.(2)Combined with the fluctuation dissipation theorem and the distribution characteristics of elemental content,the dissertation studies the relationship between the dissipation degree of ore-forming system energy and the distribution of ore-forming element content.The results show that the degree of energy dissipation in the ore-forming system determines the distribution of ore-forming element content in space..(3)The dissertation uses the DLA model to simulate the infiltration process of ore-forming fluids under different constraints.The results show that the ore-forming fluid migration path has spatial self-similarity and can be regarded as a fractal growth process.(4)Based on structured data,Machine Learning model is applied to nonlinear regression,lithology identification and feature extraction.The result shows: RandomForest and SVM models have a good effect on nonlinear relationships.GBDT algorithm has high classification accuracy for lithology identification problems.The features extracted by the Random Forest model have higher classification ability through the calculation of the entropy function.Compared with the PCA model,the KPCA model based on RBF kernel function and Sigmoid kernel function has better feature extraction effect,and the experimental sample is linearly separable in the newly feature space.The two manifold learning models,ISOMAP and LLE,can effectively learn the changes in high-dimensional space,making the classification boundary of the sample in the new feature space obvious after mapping.The Feedforward Neural Network model can realize the automatic extraction of features while solving the target problem.The weight distribution of the hidden layer affects the validity of feature learning.