首页|期刊导航|内蒙古民族大学学报(自然科学版)|基于生成对抗网络的表面肌电信号数据增强方法研究

基于生成对抗网络的表面肌电信号数据增强方法研究OA

Surface Electromyography Signal Data Augmentation Based on Generative Adversarial Networks

中文摘要英文摘要

表面肌电信号(surface electromyography,sEMG)在人机交互、康复辅助和智能假肢控制等场景中具有重要应用价值,但其采集过程易受受试者差异、电极位置、肌肉状态及实验环境影响,导致深度学习模型在小样本和跨受试场景下的泛化能力受限.针对上述问题,构建一种面向sEMG时序特征的编码器辅助生成对抗网络(EAGAN),用于表面肌电信号数据增强.该方法首先对原始 sEMG信号进行窗口化、Min-Max缩放和条件标签编码,随后利用LSTM编码器提取真实信号的时序语义特征,并将其与随机噪声和条件标签共同输入生成器,生成指定类别的合成sEMG信号.判别器采用条件Critic结构输出Wasserstein评分,并结合梯度惩罚完成WGAN-GP对抗训练.实验采用多项评价指标对生成信号质量进行评估,并通过编码器消融实验分析LSTM编码器对生成性能的影响.结果表明,该方法能够生成与真实sEMG较为接近的合成样本,可作为小样本肌电识别场景中的一种候选样本扩充方式.

Surface electromyography(sEMG)signals are of significant value in human-computer interaction,reha-bilitation assistance,and intelligent prosthetic control.However,sEMG acquisition is susceptible to inter-subject vari-ability,electrode placement,muscle condition,and experimental environments,which limits the generalization ability of deep learning models in small-sample and cross-subject scenarios.To address these issues,an encoder-assisted gen-erative adversarial network(EAGAN)is developed to exploit temporal information in sEMG for data augmentation.The method first processes raw sEMG signals through window segmentation,Min-Max scaling,and conditional label en-coding.Then,an LSTM encoder is employed to extract temporal semantic features from real sEMG signals,which are combined with random noise and conditional labels as inputs to the generator to synthesize category-specific sEMG sig-nals.The discriminator adopts a conditional Critic structure to output Wasserstein scores,and WGAN-GP adversarial training is performed with gradient penalty.The experiment employs multiple evaluation metrics to assess the quality of generated signals,and analyzes the impact of the LSTM encoder on generation performance through encoder ablation experiments.The results indicate that the method can generate synthetic samples that are relatively close to real sEMG signals,suggesting its potential as an alternative augmentation source for small-sample myoelectric recognition tasks.

王一博;路扬;白青海;马小蕾;张强;赵乾志

内蒙古民族大学 计算机科学与技术学院,内蒙古 通辽 028043内蒙古民族大学 计算机科学与技术学院,内蒙古 通辽 028043内蒙古民族大学 计算机科学与技术学院,内蒙古 通辽 028043内蒙古民族大学 计算机科学与技术学院,内蒙古 通辽 028043内蒙古民族大学 计算机科学与技术学院,内蒙古 通辽 028043内蒙古民族大学 计算机科学与技术学院,内蒙古 通辽 028043

信息技术与安全科学

表面肌电信号数据增强生成对抗网络LSTM编码器

surface electromyographydata augmentationgenerative adversarial networkLSTM encoder

《内蒙古民族大学学报(自然科学版)》 2026 (5)

33-41,9

内蒙古自治区自然科学基金项目(2023LHMS06021)内蒙古自治区直属高校基本科研业务费项目(GXKY22134,GXKY23Z015)内蒙古民族大学博士科研启动基金项目(BSZ002)

10.14045/j.cnki.15-1220.2026.05.005

评论