基于改进的能量生成对抗网络表面肌电数据增强方法

A surface EMG data enhancement method based on improved energy generation adversarial networks

摘要:

肌电控制是智能假肢研究的重点，其识别算法需要大量的肌电数据支持，而表面肌电信号存在采集困难、数据多样性匮乏及质量不稳定的问题。因此，提出一种基于改进的能量生成对抗网络(EBGAN)的肌电数据增强方法。将卷积神经与EBGAN模型相结合，完成模型改进优化从而模拟原始数据生成过程，并采用动态时间翘曲和信号快速傅里叶变换幅度的均方误差作为评价指标，从时、频两域评估生成数据的真实性；使用支持向量机等模型对合成以及原始数据进行分类识别，验证其有效性。实验证明，改进的EBGAN模型生成的肌电信号与原始信号具有高度相似性，合成数据集显著提升了分类准确率，提升幅度在1%至9%之间。证实了数据增强方法的有效性，为肌电信号的智能化分析和应用提供了新的途径。

Abstract:

Electromyography (EMG) control is a focal point in intelligent prosthetics research, where the recognition algorithms rely on extensive EMG data. However, the collection of surface EMG signals faces challenges due to difficulties in acquisition, lack of data diversity, and instability in quality. Hence, an EMG data augmentation method based on an improved Energy-based Generative Adversarial Network (EBGAN) is proposed. This method combines convolutional neural networks with the EBGAN model to enhance the model"s optimization and simulate the generation process of original data. Dynamic Time Warping and the Mean Squared Error of the Fast Fourier Transform amplitude are employed as metrics to evaluate the authenticity of the generated data across time and frequency domains. Support Vector Machines and other models are used to classify and validate the effectiveness of both the synthesized and original data. Experiments demonstrate that the EMG signals generated by the improved EBGAN model highly resemble the original signals, with the synthesized data set significantly improving classification accuracy by 1% to 9%. This confirms the effectiveness of the data augmentation method and provides a new approach for the intelligent analysis and application of EMG signals.