Foundation Models for EMG Human-Machine Interfaces

The development of generalizable models for Electromyography (EMG) signal analysis is a significant challenge, limited by high variability across subjects, conditions, and acquisition devices and platforms, alongside a reliance on large, task-specific labeled datasets. This thesis introduces a new paradigm to address these limitations: a compact, pre-trained Foundation Model specifically for the EMG domain. We propose an encoder-only Transformer architecture trained using a self-supervised, masked-signal modeling objective on large-scale unlabeled data. By adapting vision-style tokenization for multi-channel EMG and incorporating Rotary Positional Embedding to allow for extrapolation, the model learns robust and transferable representations. The resulting 3.6 million parameter model demonstrates a remarkable combination of efficiency and high performance. It sets a new state-of-the-art on the EPN-612 (96.60% accuracy) and UCI EMG (97.86% accuracy) gesture recognition benchmarks, significantly outperforming prior models with over ten times the parameters. The model's versatility is further proven by achieving a competitive 8.53° Mean Absolute Error in cross-subject kinematic regression, surpassing LSTM baselines in discrete gesture decoding, and showing remarkable performance in silent speech recognition despite its unimodal, EMG-only pre-training regime. This work validates that a single, self-supervised encoder can serve as a powerful foundation for diverse EMG tasks. Its high accuracy, coupled with a modest parameter count, paves the way for a new generation of robust, data-efficient human-machine interfaces and opens the door to their deployment on resource-constrained embedded environments.