基于分布共识的联邦增量迁移学习

Federated Incremental Transfer Learning Based on Distributed Consensus

相同生产工艺的工业过程协同建模是解决工业难测参数在线软测量的有效方法,但因生产原料、设备等因素差异,所形成的分布式数据往往呈现非独立同分布特性(Nonindependent Identically Distribution,Non-IID).同时,受生产环境变化影响,数据分布特性会随时间发生变化.因此,工业建模场景对模型的个性化配置和自主调整能力提出了更高的要求.为此,本文提出一种结构与参数并行优化的联邦增量迁移学习方法(Federated Incremental Transfer Learning,FITL).所提方法在增量式联邦学习框架下,建立了基于模型输出信息的联邦共识组织,并利用横向联邦进行组内增强;进而,面向联邦共识组织,通过最小化组间共识差异增量迁移不同共识组织信息;最后,结合组内横向增强和跨组织迁移学习,构造增量迁移下的联邦学习模型.在工业数据集和基准数据集上的实验结果表明,与现有方法相比,所提模型能更好地实现不同工况Non-IID情况下的协同建模.在过程工业回归任务和公开数据集的分类任务中,FITL能在多工况环境下相较基线方法提升9%和16%的模型预测精度.

Industrial process collaborative modeling with the same production process is an effec-tive method to solve the difficult industrial parameters online soft measurement.Due to the differences in production materials,equipment and other factors,the distributed data often prensent nonindependent identically distribution(Non-IID).Simultaneously,influenced by chan-ges in the production environment,the distribution characteristics of data change over time.Con-sequently,industrial modeling scenarios demand heightened requirements for personalized config-uration of models and autonomous adjustment capabilities.To address these concerns,this paper proposes a federated incremental transfer learning(FITL)strategy that achieves parallel optimi-zation of both structure and parameters.Under the framework of incremental federated learning,a federated consensus organization based on model output information is established,and horizon-tal federated is used for intra-group enhancement.Furthermore,the information of different consensus organizations is incrementally migrated for federal consensus groups by minimizing consensus differences between groups.Finally,a federation learning model under incremental transfer is constructed by combining intra-group horizontal reinforcement and cross-organization transfer learning.Experimental results on industrial data sets and benchmark data sets show that,compared with the existing methods,the proposed model can better realize collaborative modeling under different working conditions of Non-IID.In the regression task of process indus-try and the classification task using public datasets,FITL exhibits a notable enhancement of 9%and 16%in model predictive precision over baseline methods in multiple working conditions.