Study On Ash Prediction Experiment Of Flotation Tailings Based On Deep Learning

In order to achieve the goal of "double carbon",coal,as the main source of carbon emissions,must take the path of intelligent green low-carbon development and utilization innovation,coal preparation plant as the terminal of coal production process,its degree of intelligence directly affects the overall height and level of intelligent construction of coal mines.Coal slurry flotation is an important part of coal sorting,with the continuous development of instrumentation technology and computer technology,its automation control level is also growing,however,the closed-loop control of the whole flotation process is always difficult to achieve,and its core constraint is the serious lag of flotation product quality detection,which cannot meet the intelligent demand of coal processing plant.In recent years,machine vision and deep learning technologies have been widely used in industry to provide new solutions for flotation product ash detection.Therefore,this thesis investigates the prediction method of flotation tailing ash based on image processing and proposes a collaborative prediction method based on CNN prediction model with migration learning combined with BP residual estimation model based on color features to design an online prediction system for flotation tailing ash.The thesis first analyzes the feasibility of the method of predicting ash based on images,formulates samples with different ash contents,builds an image acquisition system for flotation tailings,and selects key equipment for the system.The feasibility of the method to predict ash based on the image of flotation tailings coal slurry is analyzed from the image and camera theory,and it is proved that the slurry image is only related to the properties of the sample itself.The samples were con Figureured using different products of 1/3 coking coal from Linliao coal processing plant of Huabei Mining Group,and samples with ash content interval of 20%-70% were obtained.The flotation tailing coal image acquisition system was designed and built in the laboratory,and the selection study was conducted for the main equipment of the system including industrial cameras,lenses,light sources and containers.Optimization tests were conducted for key parameters of the system to analyze the effect of concentration conditions on image acquisition,pre-processing for the acquired tailings pulp images,and extracting image feature data with actual production experience.In order to acquire excellent quality images,the key parameters of the system were optimized and the results showed that the best conditions were achieved when the light source intensity was 13500 Lux,the aperture size was F=2.8 and the exposure time was 100 ms.The effect of concentration on the image of tailing coal slurry was studied.By con Figureuring the slurry with different ash samples of 5%-50%concentration,the variation of the mean ash value with concentration was investigated,and the results showed that the concentration had a mild effect on the high ash slurry,but the relative error was less than 5%,and the concentration had little effect on the image of flotation tailing coal from the overall trend as well as the practical application point of view.The image acquisition was performed under optimal conditions for different ash slurries,and finally 300 sets of image data and their corresponding ash values were obtained.By cropping and removing the irrelevant interference parts of the tailing coal images,sufficient information was retained while reducing the size and the amount of operations.The relationship between gray features and gray score of tailing coal images was analyzed,the gray feature values of the images were extracted,the color space of the images was analyzed,the color feature values of the images were extracted,and the extracted 12 feature values were used as the data set of the model.The CNN model based on migration learning combined with the BP residual estimation model based on color features is proposed for collaborative prediction,and the CNN-BP network model is constructed to further improve the prediction accuracy while reducing the sample requirement.Five representative convolutional neural network models are selected,and the CNN models are trained with the pre-processed images as input variables and the actual gray values as output variables,and the prediction effects of the five models are compared using the validation set data,and the results show that the VGG network has the best effect,the Alex Net network has the second best effect,and the Res Net network has the worst effect.Comparing the prediction effect of CNN-BP model with CNN model alone and BP model alone,the results show that the algorithm proposed in this thesis can effectively correct the CNN model,and the effect is better than BP model,and the accuracy of the prediction result for the test set data reaches 90%,the mean square error is 0.4155,the coefficient of determination is 0.9983,and the average absolute error is only 0.47.Design an online ash prediction system for flotation tailings,combine image acquisition hardware with ash prediction model,and analyze the system operation effect by flotation test.Implement the communication design between MATLAB and industrial camera,combine the image acquisition and model prediction unit by writing code,and complete the software interface based on App Designer toolbox.By changing the amount of flotation chemicals to get different ash tailing pulp,the operation effect of the ash online prediction system is analyzed,and the results show that the system is stable,with good real-time performance and high accuracy of ash detection,which can meet the demand of flotation production and provide a method to solve the key problem of flotation intelligence.The thesis has 38 figures,14 tables,and 106 references.