Time-domain Mixup Source Data Augmentation of sEMGs for Motion Recognition towards Efficient Style Transfer Mapping

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Technology

Background:

• Surface electromyogram (sEMG) offers promising applications for motion recognition due to its integration with wearable devices and high signal-to-noise ratio.
• Inter-subject variability and limited data availability pose significant challenges for developing accurate sEMG-based motion recognition classifiers.
• Transfer learning (TL) can mitigate inter-subject variability, but small source datasets risk overfitting in TL-combined classifiers.

Purpose of the Study:

• To evaluate the impact of the mixup data augmentation technique on the accuracy of motion recognition using TL.
• To assess the performance of TL-combined classifiers with augmented versus non-augmented datasets.
• To investigate the effectiveness of mixup in mitigating overfitting in TL models with limited source data.

Main Methods:

• Utilized an 8-class sEMG dataset from 25 subjects, recorded using wearable sensors.
• Employed a time-domain data augmentation method, mixup, to increase the source dataset size up to 10 times.
• Compared the performance of motion recognition classifiers (support vector machine and a 4-layered fully connected feedforward neural network) with and without TL, and with augmented data.

Main Results:

• The mixup data augmentation method significantly improved the performance of TL-combined classifiers.
• Augmenting the source dataset with mixup enhanced the accuracy of motion recognition models.
• The study demonstrated the effectiveness of mixup in addressing overfitting issues associated with TL in sEMG classification.

Conclusions:

• Mixup is an effective data augmentation strategy for improving TL-based sEMG motion recognition, particularly when source data is limited.
• The findings suggest that data augmentation plays a crucial role in enhancing the robustness and accuracy of wearable-based motion recognition systems.
• Future research will explore diverse datasets and augmentation methods to further understand data quantity's influence on sEMG recognition.