基于域间Mixup微调策略的跨被试运动想象脑电信号分类算法

Cross-subject motor imagery EEG classification based on inter-domain Mixup fine-tuning strategy

为了缓解传统微调算法的灾难性遗忘问题,本文提出了一种基于域间Mixup微调策略的跨被试运动想象脑电信号分类算法Mix-Tuning。Mix-Tuning通过预训练、微调的二阶段训练方式,实现跨领域知识迁移。预训练阶段,Mix-Tuning使用源域数据初始化模型参数,挖掘源域数据潜在信息。微调阶段,Mix-Tuning通过域间Mixup,生成域间插值数据微调模型参数。域间Mixup数据增强策略引入源域数据潜在信息,缓解传统微调算法在样本稀疏场景下的灾难性遗忘问题,提高模型的泛化性能。Mix-Tuning被进一步应用于运动想象脑电信号分类任务,实现了跨被试正向知识迁移。Mix-Tuning在BMI数据集的运动想象任务达到了85.50%的平均分类准确率,相较于被试–依赖和被试–独立训练方式的预测准确率58.72%和84.01%,分别提高26.78%和1.49%。本文分析结果可为跨被试运动想象脑电信号分类算法提供参考。

In order to alleviate the catastrophic forgetting problem of vanilla fine-tuning algorithms,we propose a crosssubject motor imagery EEG classification method based on inter-domain Mixup fine-tuning strategy,i.e.,Mix-Tuning.Mix-Tuning realizes cross-domain knowledge transfer through a two-stage training manner consisting of pre-training and fine-tuning.In the pre-training stage,Mix-Tuning uses the source domain data to initialize the model parameters and mine potential information of the source domain data.In the fine-tuning stage,Mix-Tuning generates inter-domain interpolation data to fine-tune the model parameters through inter-domain Mixup.Inter-domain Mixup data enhancement strategy introduces latent information of the source domain data,which alleviates the catastrophic forgetting problem of Vanilla Fine-tuning in sparse sample scenarios and improves the generalization performance of the model.Mix-Tuning is further applied to the motor imagery EEG classification task and achieves cross-subject positive knowledge transfer.Mix-Tuning achieved an average classification accuracy of 85.50%on motor imagery task BMIdataset.Compared with 58.72%and 84.01%for Subject-specific and Subject-independent training manner,Mix-Tuning increased by 26.78%and 1.49%,respectively.The analysis results in this paper can provide a reference for cross-subject motor imagery EEG classification algorithm.