Research On Prediction Method Of Dust Concentration In Fully-Mechanized Excavation Face Based On Ensemble Learning

Dust is one of the main disasters and occupational hazards in underground coal mines.The pneumoconiosis induced by it seriously endangers the occupational health of coal mine workers,and the coal dust explosion caused by it seriously threatens the life safety of people.With Chinese coal mining from mechanization to intelligent and unmanned,dust prevention and control means will also develop to the direction of intelligence.The introduction of dust prediction method is conducive to the development of dust perception technology,and provides basic data for intelligent dust prevention and dust control.However,there is a lack of research on mine dust concentration prediction,which restricts the research and development of mine dust intelligent perception technology.To this end,the author introduces the relevant algorithms and ideas of deep learning and machine learning,and carries out the research on the dust concentration prediction method of mine comprehensive excavation working face.The main work and results are as follows:The optimization methods of parameters and structures of various basic learner models are studied.In order to achieve the best effect of the ensemble prediction model,the base learner model should also achieve the best prediction effect.Therefore,the parameters and structure of the base learner model need to be optimized.Through analysis,the maximum depth and number of optimal decision trees for random forest,XGBoost and Light GBM are 12 and 120,6 and 20,11 and 60 respectively.Besides the model super parameter,the deep learning variants need to consider how to set the model structure.It is found that the optimal structure of the model is a neural network with three hidden layers,in which the first two layers are the RNN layer,the neurons in the first layer are larger than the second layer,and the third layer is the fully connected layer.Next,the number of neurons in the hidden layer of the model is used as a variable to explore the prediction effect of the model,and finally the optimal number of neurons in the hidden layer of RNN,LSTM,GRU and Bi-RNN are 40,35,40,35,30,35,50,45 and 50 respectively.Bi-RNN structure and parameter setting: The hidden layer consists of two RNN layers with equal neurons(70 neurons)and opposite directions,and one fully connected layer with 70 neurons.The reasons for the different prediction characteristics of machine learning and deep learning models are proved.Through a large number of simulations,the mean variance(MSE)and the mean absolute percentage error(MAPE)of the various base learners’ models are obtained.From the analysis of the prediction effect of the basic learner model,it can be seen that the MSE of the machine learning model is generally smaller than that of the RNN-model,but its MAPE is larger than that of the RNN-model.By analyzing the formulas of MSE and MAPE,we can see that MSE is the set of the average of the square of the prediction error,and MAPE is the set of the prediction error divided by the average of the actual value,so MSE will amplify the expression effect of the error.From the overall point of view,the prediction effect of deep neural network model is better than that of machine learning model,so its MAPE is smaller;However,the prediction error of RNN model is larger than that of machine learning model in some time periods,so the MSE of RNN model is larger.The optimal integration mode of the dust prediction model is proposed.The method to improve the accuracy of prediction model is determined: based on the Stacking method,all kinds of base learners are integrated to enhance the prediction ability of the model.According to the types of basic learners,this thesis can be divided into three integration methods: machine learning integration,deep learning integration and hybrid model integration.After integration,it is found that the MSE and MAPE of the machine learning integration model,the deep learning integration model and the hybrid integration model are 52.54,6.65,56.48,4.90,45.90 and 4.72,respectively,which are 31.11%,-7.27%,17.29% and 14.44 lower than the average MSE and MAPE of each basic learner.The integration of the model also continues the characteristics of its basic learner model.For the machine learning integration model,the MSE has decreased,but the MAPE has increased.Both MSE and MAPE of the deep learning integration model have decreased.The hybrid integration model has the best effect,not only the advantage that machine learning algorithm can effectively eliminate large error points in the prediction model,but also has the characteristics of high overall accuracy of deep learning model.Its MSE and MAPE are also the smallest among the three integration models.Therefore,the hybrid integration method is finally chosen to build the dust prediction model.The dust concentration prediction model and system based on ensemble learning were developed.There will be some differences in the dust production law in different workplaces.If the dust production law changes,the parameters and structure of the prediction model will also change.It is complicated and not intuitive to modify the model directly from the source code.Therefore,a man-machine interaction interface is designed for the built dust prediction model,so as to adjust the model quickly.The model and interactive interface mentioned above are all compiled by code,and the source code needs to run in a certain virtual environment.If you want to run the model on other computers,you need to prepare the virtual environment in advance.Moreover,if the reconstructed virtual environment is different from the version of the virtual environment in which the prediction model was compiled,it is very likely that the prediction model cannot be run.In order to improve the general applicability of the prediction model,the dust prediction system is packaged into an executable program to make it more convenient to use.There are 61 figures,15 tables and 59 references in this thesis.