Prediction Of Component Content In CePr/Nd Extraction Process Based On Virtual Sample Generation

Rare earth,known as the "universal soil",is an industrial vitamin.All high-tech products in the world cannot do without it.China’s rare earth extraction technology is leading the world under the guidance of cascade extraction theory,the research achievement of the Chinese Academy of Sciences academy of Sciences scientist Xu Guangxian.However,the online detection of component content in the extraction process still remains in the stage of "timing sampling and off-line analysis".Some scientists in the field of rare earth have applied soft sensor technology to the rapid detection of component content,and achieved a series of achievements.But traditional soft-sensing methods need to rely on a large amount of data support,while the complex rare earth extraction process has high data acquisition costs,data duplication and other reasons,resulting in not much effective process data,which easily leads to small sample problems such as low model prediction accuracy and poor generalization performance.Therefore,it is a new way to expand the sample size through reasonable methods and use the soft-sensing method under the background of big data technology to improve the prediction accuracy of the rare earth extraction process model.Aiming at the complex rare earth extraction industrial site,due to the high cost of data acquisition,data duplication and other reasons for small sample problem,This paper takes the praseodymium/neodymium(CePr/Nd)extraction and separation process as the research object,and proposes a component content prediction method based on virtual sample generation.The main research contents are as follows:1.From the perspective of the output attributes of the model,a CePr/Nd component content prediction method based on SCN hidden layer interpolation is proposed.First,the SCN algorithm is used to establish the non-linear function relationship between the first moment of the H,S,and I color characteristic components of the CePr/Nd mixed solution image and the Nd component content.Then,According to the stochastic configuration network the mapping relationship between the network input layer,the hidden layer and the output layer,virtual input and output data are obtained respectively.The generated virtual samples will be added to the real small sample training samples to reconstruct the component content SCN model,so as to achieve more accurate prediction of CePr/Nd component content.2.For the specific problem of virtual samples generated based on SCN hidden layer interpolation,data amplification is performed from the perspective of input attributes.a method for predicting the content of CePr/Nd components based on GA optimized MD-MTD to generate virtual samples is proposed.MD-MTD is used to generate virtual input data,and the corresponding virtual output data is obtained through the SCN model;in order to make the generated virtual sample conform to the expected space,the virtual sample generation process is transformed into a virtual sample optimization process more likely;GA algorithm is used to search for "Optimal" virtual sample set;adding the best virtual sample to the real small-sample training sample,reconstructing the component content SCN model,so as to achieve more accurate prediction of CePr/Nd component content.3.In view of the fact that the above two methods are one-sided in terms of output and input attributes respectively,a method of CePr/Nd component content based on mixed virtual samples is proposed.Considering the complementarity of data amplification methods based on input and output attribute perspectives,SCN hidden layer interpolation method and GA optimized MD-MTD method were used to generate two different types of virtual samples for repeated data cleaning and mixing.The rationality,validity and applicability of two different types of virtual samples were analyzed through the process data collected from the rare earth extraction production site.The results show that the CePr/Nd component content prediction model based on mixed virtual samples has better generalization ability than the real small sample prediction model.In conclusion,the hybrid virtual sample generation method can more accurately predict the content of rare earth elements,Which is an effective tool to solve the problem of small samples in rare earth extraction and separation industry.