Deep Convolutional Generative Adversarial Network-Based EMG Data Enhancement for Hand Motion Classification.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Acquiring human bio-signal data, such as electromyography (EMG), requires stringent experimental conditions and ethical approvals, resulting in limited datasets for classifier training.
• The era of big data necessitates methods to expand existing datasets, particularly for complex biological signals.

Purpose of the Study:

• To propose and evaluate a method for enhancing multiple-channel EMG data using a deep convolutional generative adversarial network (DCGAN).
• To assess the similarity and diversity of synthetically generated EMG data compared to real data for improving hand motion recognition.

Main Methods:

• EMG signals were transformed into grayscale images via matrix transformation, normalization, and histogram equalization.
• A DCGAN was trained on these images to generate synthetic EMG data from random noise inputs.
• Classification accuracy using Support Vector Machine (SVM) and Random Forest (RF) was employed to evaluate data similarity and diversity.

Main Results:

• Adding synthetic data to the training set had minimal impact on classification performance, indicating similarity between real and synthetic data.
• A slight average accuracy increase of 1%-2% was observed for SVM and RF classifiers with additional synthetic data.
• Cross-validation demonstrated that synthetic samples possess large inter-class distances, suggesting enhanced diversity.

Conclusions:

• The DCGAN-based method can generate synthetic EMG data that is similar to real data, with potential to enrich training datasets.
• Histogram equalization was found to significantly improve the performance of EMG-based hand motion recognition.
• The generated synthetic data exhibits new characteristics that can potentially improve classifier robustness and generalization.