摘要
针对脑电数据不足和数据异构导致运动想象解码性能下降,提出一种融合双服务器结构和多头注意力机制的联邦学习模型,实现异构脑电数据的运动想象多任务识别。通过欧几里得对齐,减少客户端数据分布差异。设计双服务器架构解决数据异构,服务器一选择并共享本地模型的最佳特征,将其作为全局共享资源改进客户端更新策略,解决客户端漂移问题;服务器二聚合本地模型参数,并进行全局参数微调训练,增强模型适应不同场景脑电数据异构的能力,提高全局模型的通用性。利用Transfomer多头自注意力,提高运动想象脑电的特征表示和模型学习能力。在BCI IV 2a脑机接口竞赛数据集上,该模型与联邦学习基准模型相比,脑电解码平均准确率提升了21.05百分点,Kappa值提升了0.283。在两个脑电数据集进行不同用户、不同环境和设备的跨数据集测试,脑电解码分别获得了71.13%和86.63%的平均准确率以及0.615和0.822的Kappa系数。结果表明:该模型在多用户、多设备且数据高度异构的场景下,运动想象脑电识别能够获得较好的性能,具有较强的泛化性。
To address the decline in decoding performance due to the lack of motor imagery electroen‐cephalography(EEG)data and data heterogeneity,this study proposes a federated learning model.This model integrates a dual-server structure and multi-head attention mechanism to facilitate the recognition of motor imagery tasks in heterogeneous EEG data.Reducing differences in EEG data distribution between clients by Euclidean alignment methods.The dual-server architecture is designed to tackle data heteroge‐neity.Server One identifies and shares optimal features from the local model,using them as a globally shared resource to improve client update strategies and resolve client drift issues.Server Two consolidates parameters from the local model,performing global parameter fine-tuning training to enhance the model′s adaptability to heterogeneous EEG data in different scenarios,thereby improving the universality of the global model.The incorporation of the Transformer’s multi-head self-attention enhances feature representation and learning capability of motor imagery EEG.On the BCI IV 2a Brain-Computer Interface competition dataset,this model shows a 21.05 percent point improvement in EEG decoding average accuracy and a 0.283 increase in the Kappa value compared to the federal learning benchmark model.In cross-dataset testing with different users,environments,and equipment on two EEG datasets,the EEG decoding achieved average accuracies of 71.13%and 86.63%and Kappa coefficients of 0.615 and 0.822,respectively.The results demonstrate that this model can deliver stable performance in scenarios with multiple users and devices,and highly heterogeneous data,exhibiting strong generalizability.
作者
吴健民
张圆
乔晓艳
WU Jianmin;ZHANG Yuan;QIAO Xiaoyan(College of Physics and Electronic Engineering,Shanxi University,Taiyuan 030006,China)
出处
《测试技术学报》
2025年第5期581-591,598,共12页
Journal of Test and Measurement Technology
基金
山西省研究生教育创新计划项目(2024SJ021)。
关键词
运动想象脑电
数据异构
联邦学习
双服务器架构
多头自注意力
motor imagery electroencephalogram
heterogeneous data
federated learning
dual-server architecture
multi-head self-attention
