Research On Sorting Index Prediction Of Compound Dry Sorter Based On Machine Learning

China is the largest coal producer and consumer in the world,and the coal-based energy structure has supported China’s rapid economic development.The efficient and clean utilization of coal resources has become a long-term strategic requirement.Coal preparation,as the source of clean coal utilization technology,is one of the important means to realize the "twenty report" to further promote the energy revolution and strengthen the clean and efficient utilization of coal.Our country’s coal resources and water resources are inversely proportional to each other,2/3 of the coal resources are distributed in the northwest area of the drought and lack of water,and the reserves of the low quality coal are larger,urgent need for an effective method suitable for coal sorting in the lack of water area.As an innovation technology in the field of dry coal preparation in our country,compound dry method sorting machine has gained popularity in coal mine in the arid and water-scarce areas in northwest with its features of no water use,simple process and low production cost.Ash content,calorific value and yield are important indicators to evaluate coal quality,which determine the value of coal.The existing methods of obtaining indicators generally have the problem of lag,which cannot reflect the current coal quality in real time.The machine learning method is used to predict the ash content,calorific value and yield of the product.Therefore,this thesis aims to realize the real-time detection of sorting index.The specific research work of this thesis is as follows:(1)In view of the phenomenon of unbalanced data distribution and data scarcity in the collection process of sorting index data set of the compound dry sorter,Cubic splines interpolation data enhancement method is adopted in this thesis to average expand the test data set,which solves the problem that the model is prone to overfitting when predicting the unevenly distributed data set.In view of the noise effect of the data set and the irregularities of the data set,this thesis proposes the Z-score data standardization of the data set and Savitzky-Golay filter denoising processing.Simulation experiments show that the Savitzky-Golay filter denoising method can effectively reduce the MSE and MAE of the model.Strong correlation between feature variables means similar contributions to the model,thus increasing the computational burden of the model and even affecting the accuracy of the model.In this thesis,Pearson correlation coefficient is used to generate correlation matrix heat map to explore the correlation between feature variables,so as to eliminate strongly correlated feature variables and achieve feature dimension reduction.The results show that there is no strong correlation between the characteristic variables(|r|=0.8～1.0),so the characteristic variables in the test data set are retained.(2)For the modeling work of machine learning,based on the Kfold crossvalidation partitioning data set method,the above preprocessed complete and accurate data set is divided into the training set and the test set.The folding times K=5,and the proportion of the test set is 20%.Secondly,random forest RF,support vector machine regression SVR,BP neural network and Adaboost integrated learning machine were selected for modeling experiments.The simulation results show that both RF and BP neural network models show good performance in the process of sorting index prediction modeling,but the training time of RF is significantly shorter than that of BP neural network.Therefore,based on the analysis of actual sorting process,this thesis finally chooses RF model of RF as the base model of sorting index prediction of compound dry sorter.(3)In view of the subjectivity in the selection of model superparameters in the modeling process,which affects the model accuracy and the traditional optimization algorithm is easy to fall into local optimal,this thesis proposes improved optimization algorithms based on particle swarm optimization algorithm PSO and simulated annealing optimization algorithm SA: improved particle swarm optimization algorithm IPSO and improved simulated annealing optimization algorithm ISA.Its purpose is to strengthen the search ability of the algorithm in the early stage of iteration and the convergence ability in the late stage of search to prevent the algorithm from falling into the local optimal solution.Then,the feasibility of IPSO and ISA is verified by four Baseline test functions.The results show that IPSO and ISA have certain improvement over PSO in SA on Baseline test function,which indicates that IPSO and ISA proposed in this chapter are effective.Then,IPSO and ISA were combined with RF model and data set preprocessing technology to generate a fusion prediction model and applied to the yield data set that needed to be optimized.The results showed that IPso-RF had better optimization effect than ISA-RF,and its fitness function value(MSE)was reduced by 4.02%.In addition,the training time of IPSO-RF is also better than ISA-RF,which is shortened by 14.02%.Therefore,IPSO-RF was selected as the fusion prediction model in this thesis.(4)Finally,this thesis takes the constructed IPSO-RF fusion prediction model as the prediction model of the sorting index prediction system of the composite dry sorter.The whole system adopts the front and back end architecture based on Vue.js front-end framework,Flask back-end framework and SQLite database to design and implement the system construction process.The system is mainly composed of Web client,server and database.Users send request instructions on Web client to generate corresponding URL connection and send request to server through HTTP protocol.The server accesses the SQLite database through API call according to the corresponding request to obtain specific data,and returns it to the server for corresponding calculation according to specific tasks to realize specific business logic;Finally,the server packages the calculated data through the Vue.js framework visual operation to generate HTML text form and sends it back to the Web client for the user to display the results;The system mainly includes user login,sorting data management and sorting index prediction functions.This thesis has 57 pictures,34 tables and 106 references.