基于多分支残差深层网络的稀土萃取流程模拟

Process simulation of rare earth extraction based on multi-branch residual deep network

稀土萃取过程机理复杂,存在非线性、强耦合以及大滞后等特点,依据传统机理分析方法不能实现精确的萃取流程模拟.对此,本文提出一种多分支残差深层网络(MB-RDN)用于稀土萃取工艺流程模拟.首先,针对多级萃取槽串接而成的稀土萃取流程提出具有多分支结构的深层神经网络,该网络可以通过不同的分支出口计算每级萃取槽的组分含量;其次,为了有效缓解深层网络的梯度消失问题,在分支中引入残差结构和特征短接操作,并设计出一种多特征融合机制.所设计网络可以有效地学习原始特征、深层特征和分支间耦合特征以提升预测精度.最后的仿真结果表明了所提方法的有效性.

Process simulations based on mechanism methods hardly meet the requirements of industrial rare earth extraction since rare earth extraction is a complex nonlinear process with strong coupling and large time lags. Therefore, a multi-branch residual deep network(MB-RDN) is proposed for process simulation of the rare earth extraction process. First,a deep neural network with multi-branch structure is proposed for a rare earth extraction process with multiple extraction tanks, which can calculate the component content of each extraction tank by different branch outlets. Secondly, in order to effectively alleviate the gradient disappearance problem of deep networks, we design a multi-feature fusion mechanism by introducing residual structure and feature short-circuit operation in the branches. The designed network can efficiently learn original features, deep features, and inter-branch coupled features to improve the prediction accuracy. Finally, the simulation results show the effectiveness of the proposed method.