Online Prediction Of Elemental Component Content In Rare Earth Extraction Process Based On Machine Learning Algorithm

The rapid detection of elemental component content in the extraction process of rare earths is the primary condition for the quality of export products,however,the way to obtain component content values in the automated production process of China’s rare earth separation industry is still basically at the low level of off-line analysis.Therefore,the search for a soft measurement method of rare earth element content with the advantages of low cost,simple maintenance and rapid continuous detection is the key to improve the automation level of rare earth separation industry.In view of the color characteristics of some rare earth elements,such as Pr and Nd ions,machine vision methods based on machine learning algorithms can be used for soft measurement of elemental component content.In order to improve the accuracy of soft measurement,this paper carries out the research of online prediction of elemental component content in rare earth extraction process based on machine learning algorithm,and the main research work and related conclusions are as follows.1)A prediction model of rare earth elemental component content based on deep learning method in machine learning algorithm is proposed,and the abstract representation of rare earth extraction solution image dataset is extracted by establishing a lightweight VGG(Lightweight Visual Geometry Group)model,and a deep neural network regression model is simultaneously constructed to optimize the network weights for the image of rare earth extraction mixed solution The soft measurement of the content of individual rare-earth element components in the sample is performed.2)In order to explore the commonality between the component content of multiple rare earth elements and between the component content and concentration,a multi-task learning model is constructed,and a multi-objective optimization algorithm,a multiple gradient descent algorithm based on the optimized upper bound,is proposed for the prediction of the component content of rare earth elements.By acquiring Pr and Nd extraction mixed solution image datasets and conducting simulation and simulation experiments on soft measurement of elemental component content during rare earth extraction,relevant conclusions were drawn through a large number of comparative experiments as follows:1)The operations of convolution and pooling in the VGG model can fully capture the internal differences in the images of rare earth extraction mixed solutions,and the measurement accuracy of the lightweight VGG model is higher than that of other VGG models and the traditional machine learning algorithm LSSVM(Least Squares Support Vector Machines),and the method in this paper,compared with the traditional machine learning methods,eliminates the complicated process of manual screening In addition,the ten-fold crossover method proves that this method has high stability and strong generalization power for the prediction of rare earth elemental component content,and the maximum relative error is 2.773 8%,which meets the minimum requirement of±5% for the maximum relative error of elemental component content detection in the process of rare earth extraction and separation,and the prediction time of a single sample is within 2.8s,and the time cost meets the requirement of practical extraction production.The time cost meets the requirement of component content detection in the actual extraction production.2)The multi-task learning model can improve the generalization ability and robustness of the model by joint training among multiple tasks compared with the single-task learning model;the proposed multiple gradient descent algorithm based on the optimized upper bound is close to the shortest single-task learning and homoscedasticity uncertainty in the training process,and the training time does not increase with the number of tasks,which is much lower than the multiple gradient descent method.The overall prediction error in the three-prediction task is only second to the single-task learning,and the overall prediction error in the four-prediction task is lower than that of other optimization methods.The prediction accuracy of component content can only be improved by combining the two tasks of component content and concentration,but not that of concentration;and the prediction accuracy of component content is only slightly higher than that of dual prediction tasks for four prediction tasks,but not significantly.The maximum relative error of component content prediction in each prediction task is within ±5% of the maximum relative error of component content detection in rare-earth extraction production,and the prediction time cost is within 3s,which can meet the accuracy and real-time of online soft measurement of multi-element component content in the process of rare-earth linkage extraction.The method proposed in this paper can meet the minimum requirement of ±5% relative error in the extraction and separation of rare earths,and the prediction time cost can meet the requirement of component content detection in the extraction production,which provides a new idea for the soft measurement method of rare earth elemental component content and has practical significance in the extraction production of rare earths.