Research On The Analysis Method Of Gold Mine Wall Rock Alteration With LIBS Technology Combined With Machine Learning

The phenomenon of layering of different alteration zones in the process of hydrothenmal gold mine is an important basis for gold mine exploration.Laser Induced Breakdown Spectroscopy(LIBS)is a kind of atom emitting spectrum technology that supports erosion and fast detection.Compared with the Inductive Coupled Plasma Emission Spectromcter(IPS)LIBS has the advantages of no need for sample preparation,support for remote detection and wide range of coverage elements,and has become a star analysis method in the field of geologi-cal material analysis.Based on the actual situation of gold mine exploration,this article has carried out research on the analysis method of machine learning and LIBS based on the charac-teristics of spectral discrimination analysis and composition regression prediction prediction.The main contents are as follows:(1)Build LIBS experiment platform.Optimize the delay time,laser energy and laser superimposed times in the LIBS system.Analysis and discussion were analyzed and discussed for the noise,spectral drift and abnormal spectrum in the spectral data,and proposed a set of standardization processes suitable for spectral pre-processing for mission and experimental platforms for this article.(2)Study on classification of typical gold mineral alteration based on full-spectrum data,a LIBS spectral dataset comprising eight typical gold-related mineral alteration samples was produced.First use the self-organized mapping neural network(SOM)to achieve two-dimensional visualization of LIBS spectrum.aiming at the problem that traditional SOM does not use the labels of training samples,it is proposed to reduce the phenomenon of similar sample confusion mapping into the same neuron according to the dynamic change weight vector update of the winning node clustering situation,and the classification accuracy of the improved SOM network reaches 92.56%.After that,after parameter tuning experiments,the accuracy rate of shallow neural networks in the same data set was 93.89%.Finally,migrate the MLP-MIXER network in the field of image classification to the LIBS spectral recognition task,determine the optimal structure through repeated experiments,and the accuracy of the category in the test concentration reaches 97.92%.On this basis,this thesis use the Boltzmann slope method and the Stark extension method to calculate the physical parameters implied in each spectrum.After the normalization,use them to enhance MLP-Mixer,and the accuracy of the final classi-fication reaches 99.44%.(3)The study of the quantitative analysis methods of Si,Ca,Al and Mg in the geological samples were carried out,and the experimental data was collected in the laboratory standard sample.For the characteristics of this experimental spectrum data,the Boruta algorithm that is based on the decision-making tree algorithm construction system will be applied to the spec-tral analysis through assumptions and causal inferences.Base on raw Boruta,an improved algorithm is proposed to reconstruct the Boruta algorithm kernel using LightGBM and SHAP algorithms and loose feature check based on cross-validation with support vector regression(SVR).Experimental results show that the proposed method can show better analysis effect in SVR and partial least squares regression regression(PLSR)analysis,and at the same time,show certain generalization ability in neural network regression modeling.