Process Simulation Of Rare Earth Extraction Based On Improved Multi-Branch Deep Network

Rare earth resources have the reputation of "industrial vitamin",as an indispensable raw material for the development of high-tech industries such as new energy and intelligent equipment manufacturing,its importance is self-evident.For the intelligent transformation requirements of the industrial production process,enterprises urgently need a stable and efficient,intelligent and flexible extraction process simulation method for production decisions to reduce the waste of raw materials and improve product purity.Since the extraction reaction is a complex physicochemical change process,the traditional mechanistic model cannot accurately describe the actual rare earth extraction process.For this,based on the characteristics of cascade separation in extraction process and industrial big data,this paper designed a new neural network with multi-branch output structure.The network outputs the component content of the extraction tank by step through different branches,and is further improved by a multifeature fusion technique and sparse auto encoder.The main research contents of the article are as follows.1.To analyze a multi-component two-export rare earth cascade extraction process.The structure of conventional neural networks as black-box models with one-way data transfer between hidden layers cannot accurately represent the rare earth extraction process with multiple outputs,nonlinearity,and strong coupling.Based on the actual reaction sequence of the extraction process,this paper firstly proposes a basic multi-branch deep neural network(MB-DNN),which initially solves the shortcomings of the traditional model by introducing branching outputs step by step.2.In order to effectively alleviate the gradient disappearance problem of deep networks,a multi-branch deep feature fusion network(MB-DFFN)is proposed based on the structure of MB-DNN and prior information of rare earth extraction.This paper designs a multi-feature fusion mechanism trough introducing residual structure and feature short-circuit operation in the branches,which lets MB-DFFN can efficiently learn original features,deep features,and inter-branch coupled features for component content prediction.3.The performance of the neural network depends largely on the scale of the label data,but it is difficult to collect component content data(label information)in the actual rare earth extraction factory.In this regard,a multi-branch deep feature fusion network combined with sparse auto encoder(SAE-MBDFFN)is proposed for rare earth extraction process simulation by introducing an unsupervised pre-training method.Based on the unsupervised pre-training and supervised fine-tuning,a large amount of unlabeled data information is extracted for the initialization of the model parameters,which significantly improves the output accuracy of SAE-MBDFFN as well as the convergence speed of the loss function.In summary,this paper proposes a process simulation method of rare earth extraction by combining extraction prior information and a multi-branch neural network.The simulation results show that the proposed method has a small error for component content prediction and meets the actual production needs,which can provide intelligent decision support for process reorganization or parameter optimization.