Research On Ash Content Prediction And System Control In Coal Flotation Process Based On Deep Learning

The research on intelligent coal preparation is imperative driven by the intelligent coal mine policy of the government and the new artificial intelligence technology.As key separation processes of coal slime(-0.5 mm),the intelligent research progress of coal flotation is relatively slow,because it is difficult to make a breakthrough in the accurate prediction of ash content and real-time control of the dynamic system.At present,academic is mainly faced with four scientific problems.First,the accuracy and performance of the concentrate and tailings ash content prediction model need to be improved.Second,the contribution and mechanism of froth and tailings image features to the ash content model are not clear.Third,the modeling of flotation concentrate ash content under small-size datasets is more difficult.Fourth,the coal flotation intelligent control system based on ash content feedback regulation in the flotation process and its industrial examples have not yet been formed.For the above scientific problems,this thesis carries out relevant research based on deep learning and computer vision technology.The main contents are as follows.The feature engineering of coal flotation froth image was established from three perspectives:morphoscopic,statistical,and color-space.First,a flotation froth image instance segmentation method based on Mask R-CNN is proposed,and morphological features are only for flotation froth images.Secondly,statistical features were extracted based on the grayscale histogram,grayscale co-occurrence matrix(GLCM)and statistical model.Then,the color-space matrix(CCM)features were extracted in the HSI color space.Finally,the correlation between feature engineering for coal flotation froth/tailings image and ash content was explored.Based on the principal component analysis(PCA)dimensionality reduction feature engineering and the fitting ability of feature engineering after PCA dimensionality reduction is verified by support vector regression(SVR).Based on the convolutional neural network(CNN),the classification prediction model between coal flotation concentrate/tailings ash content was studied.The classification performance of coal flotation froth/tailings images in different ash content intervals was studied,and the results showed that the ash content span of concentrated/tailings was the best at±0.5%and±0.8%,respectively.The model training results show that the optimized Res Net 101 model has the best prediction performance on the test set of seven different ash content interval classes after data enhancement,and the model accuracy is 97.1%,the lower the flotation foam image classification accuracy.The training results of the corresponding flotation tailings image dataset show that the optimized Res Net 50 model has the highest prediction performance on the test set of 12 different ash content interval classes,and the model accuracy is 92.7%.In contrast,the CNN model has a better classification performance on the coal flotation froth image dataset between different ash content intervals.A method for predicting flotation concentrate ash content interval by transfer learning on small-size datasets is proposed.It is found that the classification performance of the coal flotation froth image dataset on the CNN model is more susceptible to the scale of the dataset,and the fifth chapter proposes two source domain datasets(Image Net and lab flotation dataset S1)and one target domain industrial flotation dataset S2.The weights of the best model in the source domain are transferred layer by layer to the target domain network through Fine-tuning and freezing to initialize and train the target domain model.The results show that the weight parameters of the S1 optimal model contribute more to the initialization of the S2 model,and the number of layers of the optimal initialization weight parameter is1-6 layers.The classification accuracy and F1-Score values of the model for coal flotation froth images in different ash content intervals are 0.8573 and 0.8551,respectively.Combined with feature engineering and CNN features,an explainable regression modeling method of coal flotation concentrate/tailings ash content based on feature fusion is proposed.In order to further improve the performance of the concentrate and tailings ash content prediction models,four typical regression model algorithms(polynomial regression(Poly),random forest regression(RFR),extreme gradient boosting(XGBoost)and deep neural network(DNN))are selected to train coal flotation concentrate and tailings feature fusion datasets and explain and visualize the feature contribution and mechanism inside the"black box"model to optimize the model based on SHAP.The results show that XGBoost has the smallest prediction error for flotation concentrate ash content under the same optimization conditions,and the MAE,RMSE,and R~2 on the test set are 0.7254,1.3688,and 0.6592,respectively.the flotation tailings ash content model performs best with the RFR model,the MAE,RMSE,and R~2 of the RFR model on the test set were 0.4916,0.7211,and 0.9947,respectively.The contribution of feature engineering to the model was much higher than that of CNN features.Finally,according to the optimal model,the flotation concentrate and tailings ash content prediction system software were developed,respectively.An intelligent coal flotation control system based on flotation tailings ash content feedback was developed and put into industrial application to verify the industrial application value of the results in this thesis.A coal flotation intelligent control system solution is proposed,and the system software is designed and developed to establish an intelligent control system for coal flotation.Compared with the production data before and after the use of the intelligent control system,it is found that after application,the tailing ash content of the flotation machine increased by 3.76%,the concentrate ash content decreased by 0.32%,the concentrate yield increased by 0.41%,and the combustible recovery rate increased by 0.68%,especially saving 6.79 L/h of collector and 15.61 L/h of the frother.The application results show that the coal flotation intelligent control system developed in this thesis can improve the quality and efficiency of coal preparation enterprises.The thesis contains 120 figures,38 tables,and 212 references.