文献类型：期刊文献

中文题名：基于双路Transformer多分支网络的稀土萃取流程仿真方法

英文题名：A simulation method for rare earth extraction process based on dual-path Transformer multi-branch network

作者：朱建勇[1];杨飞[1];李中美[2];杨辉[1]

机构：[1]华东交通大学电气与自动化工程学院,江西南昌330013;[2]华东理工大学信息科学与工程学院,上海200237

年份：2026

卷号：77

期号：4

起止页码：1988

中文期刊名：化工学报

外文期刊名：CIESC Journal

收录：;EI(收录号：20262921108274);北大核心:【北大核心2023】；

基金：国家自然科学基金项目(62363010);江西省自然科学基金重点项目(20252BAC250019);工业控制技术国家重点实验室开放课题(ICT2024B50)。

语种：中文

中文关键词：稀土萃取;多分支网络;双路Transformer;多尺度特征提取;流程仿真

外文关键词：rare earth extraction;multi-branch network;dual-path Transformer;multi-scale feature extraction;process simulation

摘要：稀土萃取过程具有多变量、非线性、强耦合等特点,传统机理分析方法难以准确仿真稀土萃取流程。因此,提出一种双路Transformer多分支网络的稀土萃取流程仿真方法。首先,针对多级萃取槽串接而成的流程特点,选取基础的多分支深层网络,通过不同的分支输出每级萃取槽的组分含量;其次,为进一步捕捉稀土萃取流程的复杂特性,在多分支网络的基础上引入多尺度特征提取模块以及双路Transformer模块,特征提取模块通过对输入数据应用不同核大小的一维卷积,提取不同层次的特征,双路Transformer模块通过并行多头注意力和时间卷积网络,同时处理长距离依赖与短期动态变化,之后利用跨分支注意力以及门控融合机制对并行特征进行整合;最后,设计匹配网络深度的差异化分支层结构,采用贝叶斯优化算法自动迭代优化模型超参数,并针对模型结构特点提出一种分阶段动态加权策略,逐步解冻分支训练,后续分支继承前序阶段学习到的网络参数。仿真结果表明了所提方法的有效性。
The rare earth extraction process is characterized by multi-variable,nonlinear,and strong coupling properties,making it difficult for traditional mechanistic analysis methods to accurately simulate the process.Therefore,a dual-path Transformer multi-branch network is proposed for simulating the rare earth extraction process.First,based on the series-connected structure of multi-stage extraction tanks,a basic multi-branch deep network is selected to output the component content of each extraction tank through different branches.Second,to further capture the complex characteristics of the rare earth extraction process,a multi-scale feature extraction module and a dual-path Transformer module are introduced into the multi-branch network.The feature extraction module applies one-dimensional convolutions with different kernel sizes to the input data to extract features at different levels.The dual-path Transformer module simultaneously processes long-range dependencies and shortterm dynamic changes through parallel multi-head attention and temporal convolutional networks.Subsequently,cross-branch attention and a gated fusion mechanism are used to integrate the parallel features.Finally,a differentiated branch layer structure matching the network depth is designed,and a Bayesian optimization algorithm is used to automatically iteratively optimize the model hyperparameters.A staged dynamic weighting strategy is proposed to address the characteristics of the model structure,gradually unfreezing branch training,with subsequent branches inheriting the network parameters learned in previous stages.Simulation results demonstrate the effectiveness of the proposed method.