Research On Data-Driven Concentrate Grade Prediction Of Hematite Ore Dressing Processes

The plant-wide hematite ore dressing production consists of a number of sequenced sub-processes and entire process control systems.The sub-processes include the raw ore crushing and sieving cell,the high intensity magnetic grinding-separation cell,the shaft furnace roasting cell,the low intensity magnetic grinding-separation-reverse flotation cell,the concentrate process cell and the tailings process cell.Raw ore is transformed into iron concentrate through physical and chemical interaction in the entire process.Concentrate grade is a production index reflecting the quality of the whole process product in mineral processing production.The management staffs determine the target value range of concentrate grade,and production line managers decompose it into the operation indices of the whole process of ore dressing production and the process parameters of the quality,efficiency and consumption of the processed products and the setting values of the process control system.The process control system is applied to make the output of the process track the setting values,so that the concentrate grade can be controlled within the target value range.Since the concentrate grade cannot be measured online,the fluctuations in production conditions and operating conditions make it difficult for the managers to make accurate decisions under dynamic environment.This is the main reason why it is difficult to guarantee quality stability.Fortunately,industrial data such as input and output variable values of each process,operating indices and concentrate grade have been recorded and stored continuously.The development concentrate grade prediction method with those industrial big data is of great practical significance for realizing intelligent optimization decision-making of ore dressing production processes.Also,it is of great academic significance for the research on a new machine learning method combining mechanism analysis with machine learning.Supported by the National Basic Research Program of China(973 project)"the total control strategy and operational control approach for complex manufacturing processes(No.2009CB320601)",the National Natural Science Foundation project"Advanced Control Technologies of the Overall Mineral Dressing Process(No.2012BAF19G01)",and the National Natural Science Foundation project "Data Driven Complex Industrial System Operation Optimization Control and its Application",the research on data-driven concentrate grade prediction of hematite ore dressing processes has been carried out.The major contributions are summarized as follows:1)Based on the mechanism analysis of the entire ore dressing production processes and the dynamic performance analysis of the whole process control system,the concentrate grade prediction problems under stable and dynamic operating conditions are presented.Also,the challenge of the concentrate grade prediction method of the entire ore dressing production processes is analyzed.2)Based on the analysis of the characteristics of the steady-state operation of the ore dressing production processes under the control system,a concentrate grade prediction model of the ore dressing production processes under the steady-state operation is developed.A concentrate grade prediction approach based on ? support vector regression is proposed.In view of the problem that the unified penalty factor is used in classical support vector regression,which cannot effectively reflect the influence of prediction errors of different sizes on the performance of the model,a new penalty factor coefficient-based prediction method is proposed.The penalty factor coefficient is selected according to the prediction error of training samples and the variance,standard deviation and distribution distance of training samples.The simulation results on an open source data set show that the improved method is more accurate and convergent than the classical algorithm.Based on the improved method and the concentrate grade prediction model,a concentrate grade prediction method is proposed.The simulation results using industrial data from a hematite ore dressing production process indicate that the prediction accuracy of the proposed method is higher than that of the existing methods.3)Through the analysis of the dynamic operation characteristics of the ore dressing production line under the action of the whole process control system,the prediction model of concentrate grade is established when the production conditions and operating conditions change.In view of the fact that concentrate grade is related to concentrate grade at previous time,operation indices of each process and input and output variables,and those three categories of information have different time scales,a concentrate grade prediction approach under dynamic working conditions is proposed,which combines mechanism analysis with multiscale information deep learning.The proposed approach is composed of a multiscale deep combined convolutional network prediction sub-model reflecting the influence of process variables on concentrate grade,a fully connected neural network prediction sub-model reflecting the influence of process operation indices on concentrate grade,a fully connected neural network prediction sub-model reflecting the impact of concentrate grade on concentrate grade in the past time and a multiscale information neural network ensemble model.Among them,each feature extraction layer of the multi scale deep combined convolutional network uses kernels with different sizes,and the optimal scale feature information outputted by each layer is spliced together across layers,and the output is transformed as the last fully connected layer of the network.The prediction accuracy of the influence of process variables on the concentrate grade is improved.The multiscale information neural network ensemble model trains the network parameters of the above described three prediction sub-models at the same time,according to the gradient information of the prediction error loss function,and achieves collaborative optimization of the weights of the influence of different time scale information on concentrate grade.Based on the actual data of a whole process of ore dressing production,the simulation experiment of the proposed concentrate grade prediction approach was carried out.The results indicate that the prediction accuracy of the proposed approach is significantly higher than that of the existing methods when the working condition changes.4)A simulation experiment system for concentrate grade prediction of plant-wide ore dressing process has been designed and developed.It consists of the hardware platform and the software system.The hardware platform is comprised of a prediction model training server in cloud,a local data acquisition and processing computer and a concentrate grade forecasting computer.The software system is comprised of a prediction model training software,a multiscale data acquisition and the processing software and a concentrate grade prediction software.The forecast model training software is comprised of several modules such as model training program management,model parameter visualization configuration,training process monitoring,prediction model evaluation,model and parameter downloading,production condition visualization analysis,data management and user management,etc.The Multiscale data acquisition and processing software is comprised of modules including process variable acquisition,operation index and concentrate grade acquisition,data pre-processing,batch uploading,and so on;The concentrate grade prediction software is comprised of proposed prediction algorithm module,working condition analysis module,parameter remote correction module,input pretreatment module,concentrate grade real-time prediction module,result display and uploading module,etc.The forecasting model training software runs on the server in the cloud,and selects a forecast algorithm for steady state operating conditions or dynamic operating conditions through visual analysis of the production condition.At the same time,it refines the weight parameters of the prediction algorithm and carries out a simulation experiment to decide whether to download the corrected weight parameters.The software of multiscale data acquisition and processing runs on the data acquisition and processing computer to collect and process the input and output of each process and operation indices and concentrate grade.The concentrate grade prediction software implements prediction approaches proposed.Effectiveness of the proposed approaches and the developed simulation system is demonstrated using industrial data from an ore dressing plant.Under steady-state operating conditions,the mean square error of concentrate grade prediction is 0.01,the maximum relative error is 0.56%,that is,the predicted errors are all within the allowable error range.Under dynamic operating conditions,the mean square error of concentrate grade prediction is 0.11,the maximum relative error is 1.7%,and more than 90%of the predicted errors are within the allowable error range.