Introduction
Reliable myoelectric prosthetic control is limited by the scarcity of large, diverse, and well-annotated surface electromyography (sEMG) datasets. Existing generative approaches, primarily based on generative adversarial networks (GANs), often suffer from training instability and limited sample diversity, reducing their effectiveness for data augmentation. We present DiT-EMG, a class-conditioned diffusion transformer framework that synthesizes realistic multichannel sEMG signals while preserving gesture-specific temporal and spatial characteristics for prosthetic machine learning.
Methods
DiT-EMG integrates a Denoising Diffusion Probabilistic Model (DDPM) with a Diffusion Transformer (DiT) backbone, Adaptive Layer Normalization for class conditioning, Classifier-Free Guidance (CFG), and Denoising Diffusion Implicit Models (DDIM) sampling. The framework was evaluated on the NinaPro DB2 benchmark using 18 controlled ablation experiments investigating DDIM sampling resolution, guidance scale, temporal patch size, and model depth. Performance was assessed using the Fréchet Inception Distance (FID), Train-on-Synthetic Test-on-Real (TSTR) ratio, Power Spectral Density (PSD) error, Dynamic Time Warping (DTW) distance, and time-domain descriptors including Root Mean Square (RMS), Waveform Length (WL), and Zero Crossings (ZC).
Results and Discussion
The optimal configuration (patch size = 8, CFG = 3.0, DDIM = 1000 sampling steps) reduced the Fréchet Inception Distance (FID) from 3219 to 256, representing a 92.1% improvement over the baseline. Synthetic signals achieved a Train-on-Synthetic Test-on-Real (TSTR) ratio of 0.93, with a Power Spectral Density (PSD) error of 3.1%, Dynamic Time Warping (DTW) distance of 0.07, Root Mean Square (RMS) error of 2.4%, Waveform Length (WL) error of 4.1%, and Zero Crossings (ZC) error of 3.7%, demonstrating close agreement with real sEMG signals across spectral, temporal, and morphological characteristics.
Conclusions
DiT-EMG demonstrates that class-conditioned diffusion transformers can generate high-fidelity multichannel sEMG while maintaining downstream classification performance. The proposed framework provides a scalable approach for synthetic data augmentation, reducing dependence on costly annotated datasets and supporting the development of more robust intelligent prosthetic control systems.