A parallel and efficient transformer deep learning network for continuous estimation of hand kinematics from
Chuang Lin1, Xifeng Zhang2, Chunxiao Zhao2
1Dalian Maritime University, Dalian, 116026, China. linchuang_78@126.com.
Scientific Reports
|October 16, 2025
Summary
A new lightweight model, PET, efficiently decomposes surface electromyography (EMG) signals into hand joint angles for human-machine interaction. PET achieves state-of-the-art accuracy with reduced latency, memory, and power consumption, making it ideal for wearable devices.
Area of Science:
- Biomedical Engineering
- Human-Computer Interaction
- Machine Learning
Background:
- Surface electromyography (EMG) offers non-invasive human-machine interaction but faces challenges like high latency, memory, and power consumption in current models.
- Transformer-based architectures, while effective, suffer from high time complexity in their attention mechanisms, increasing inference time and power demands for continuous motion estimation.
Purpose of the Study:
- To introduce PET (Parallel Efficient Transformer), a lightweight, parallel transformer model designed to overcome the limitations of existing surface EMG analysis methods.
- To enable real-time decomposition of surface EMG signals into hand joint angles with significantly reduced latency, memory footprint, and power consumption.
Main Methods:
- Developed a novel, bottom-up architecture for PET, focusing on parallel processing and efficient power mechanisms.
- Evaluated PET's performance against state-of-the-art methods (SVR, TCN, LSTM, GRU, LE-LSTM, LE-ConvMN, Transformer, Bert, MAFN, Conformer) on challenging datasets (Ninapro DB2, DB7, FMHD, SEEDS).
Main Results:
- PET demonstrated superior performance across various metrics including Correlation Coefficient, RMSE, NRMSE, and AME, outperforming all compared architectures.
- Achieved a PET correlation coefficient of 0.85 ± 0.01 on the Ninapro dataset and 0.81-0.82 on FMHD and SEEDS datasets.
- Significantly reduced end-to-end latency and power consumption compared to existing methods without compromising accuracy.
Conclusions:
- PET offers a highly accurate and efficient solution for real-time surface EMG signal decomposition into hand joint angles.
- The lightweight and parallel architecture of PET makes it suitable for deployment on clinical edge and public wearable devices.
- PET represents a significant advancement in non-invasive human-machine interaction, particularly for applications in prosthetics and rehabilitation engineering.
