Study On Bubble Detection And Fluctuation Prediction In Gas-solid Fluidized Bed Based On Machine Learning

As the basic energy and important industrial material in our country,the coal provides reliable energy guarantee for the economic and social development.Coal preparation is a key link in the process of coal production and use,which is one of the most effective ways to ensure the low-carbon utilization of high carbon energy.The main methods of coal preparation in our country are jigging,dense medium,flotation and air classification.Gas-solid fluidized bed is a typical application of dry dense medium coal preparation technology.The fluidization quality and stability of the fluidized bed can be reflected to some extent by the characteristics of bubbles and described intuitively and effectively by the inside pressure signal.At present,methods of image analysis and intrusive measurement are mainly adopted by the researchers to study the bubble characteristics and bed pressure signal in gas-solid fluidized bed.However,in the process of data collection and processing,the extremely strict prior knowledge and professional skills are required and the problems,such as high labor cost,large measurement error and complicated calculation,obviously restrict the research on optimizing the quality of fluidized bed.Therefore,based on the machine learning method,this thesis studies the bubble and the fluctuation state in the vibrated fluidized bed with its bubble image,pressure signal and operating parameters,and designs a set of gas-solid fluidized bed intelligent detection system.The main research work is as follows:1.Aiming at the problems of miss and false detection,low efficiency,caused by the tiny bubbles,interference of impurities and traditional detection methods,an end-to-end detection model based on YOLOv5 is proposed to detect bubbles quickly and accurately in the fluidized bed.Firstly,in order to reduce the problem of false detection caused by the interference of the medium impurities on the real bubbles in the complex environment,a new self-attentional convolutional structure is designed and added to the backbone feature extraction network to enhance the ability to focus on the global features and long-distance information of fluidized bed images.Then,a cross-scale weight fusion structure is introduced into the feature fusion network to solve the loss of positioning feature information,improve the efficiency of feature fusion between deep and shallow networks and the utilization of low-dimensional positioning features,so as to reduce the probability of miss detection for tiny bubbles.Finally,in the network detection head part,a decoupled detection head structure fusing inverting convolution is proposed to improve the accuracy and speed of detection model.Experiments show that the proposed algorithm has the best detection performance compared with other mainstream algorithms on the self-built data set,which proves the effectiveness of the improvement.2.To improve the problems of large labor cost,high measurement error and complicated calculation process for the study of fluctuation in fluidized bed,a model based on random forest regression is proposed to predict the density fluctuation in fluidized bed.Firstly,the dataset is established by combining the bubble characteristic information extracted from the proposed bubble detection model,the operating parameters of the vibrated fluidized bed and the pressure signal synchronously obtained during bubble image data collection.Then,the key variables are selected by the importance analysis of variables,and the training parameters of the prediction model are optimized by the parameter tuning ways including learning curve and grid search.At last,the prediction task of the density fluctuation in fluidized bed is realized with low training cost and high accuracy.3.On the basis of the above works,an intelligent detection system of gas-solid fluidized bed is innovatively proposed by uniting the gas-solid fluidized bed separation system and models of bubble detection and density fluctuation prediction.While optimizing the control mode of data collection hardware equipment,the system integrates the two models into a whole process visual operation interface,which covers data loading,model weight selection and threshold adjustment,bubble detection and characteristics extraction,and bed fluctuation prediction.Experiments show that the system can help users in different fields to realize the tasks of bubble detection and density fluctuation prediction accurately and in real-time for the fluidized bed,and effectively improve the problems of model deployment and system incompatibility in practical application,which has great practical value for the research of intelligent coal preparation.