Аннотация
Dexterous prosthetic control demands intent decoders that calibrate quickly and remain robust to arm-pose variation. We present a transfer-learning framework that advances the calibration-based family of algorithms via a dual-branch Transformer fusing surface electromyography (sEMG) and inertial measurement unit (IMU) signals. Training proceeds in three stages:(i) SourceNet learns pose-robust, subject-agnostic features from a multi-subject cohort; (ii) TargetNet adapts these features to a new user with only two repetitions per class; and (iii) PFCNet (Parallel Feature Concatenation) combines SourceNet and TargetNet embeddings for final classification. To our knowledge, prior calibration-based methods have focused predominantly on sEMG-only pipelines and do not explicitly optimize the models using additional sensing modalities during calibration; our modality-aware design addresses this gap. On a held-out subject (9 subjects; held-out subject evaluation; two-repetition calibration), the Transformer decoder achieves 99.35% accuracy, surpassing a strong CNN baseline while using approximately five times fewer learnable parameters. This supports embedded, real-time deployment. This model comprises 3.71M parameters (≈ 14.84 MB in 32-bit float; ≈ 3.71 MB when 8-bit quantized). Accordingly, its embedded inference footprint and latency are small: we expect single-digit to low double-digit millisecond inference times depending on hardware (e.g., sub-10 ms with 8-bit quantization on an NPU, or tens of ms on a CPU-only platform), making the system compatible with real-time prosthetic control. Ablations confirm that (a) sEMG–IMU fusion is essential for accuracy under pose changes and (b) minimal, two-repetition calibration suffices for reliable personalization. To enable reproducibility, we release a systematically curated dataset of nine subjects with synchronized sEMG and IMU recordings. Collectively, the proposed framework delivers a more accurate, parameter-efficient, and rapidly calibratable solution for prosthetic hand control, extending the calibration-based algorithm family with a principled, multimodal optimization strategy.