Abstract
As a key technology for the non-invasive human-machine interface that has received much attention in the industry and academia, surface EMG (sEMG) signals display great potential and advantages in the field of human-machine collaboration. Currently, gesture recognition based on sEMG signals suffers from inadequate feature extraction, difficulty in distinguishing similar gestures, and low accuracy of multi-gesture recognition. To solve these problems a new sEMG gesture recognition network called Multi-stream Convolutional Block Attention Module-Gate Recurrent Unit (MCBAM-GRU) is proposed, which is based on sEMG signals. The network is a multi-stream attention network formed by embedding a GRU module based on CBAM. Fusing sEMG and ACC signals further improves the accuracy of gesture action recognition. The experimental results show that the proposed method obtains excellent performance on dataset collected in this paper with the recognition accuracies of 94.1%, achieving advanced performance with accuracy of 89.7% on the Ninapro DB1 dataset. The system has high accuracy in classifying 52 kinds of different gestures, and the delay is less than 300 ms, showing excellent performance in terms of real-time human-computer interaction and flexibility of manipulator control.
Introduction
Surface electromyography (sEMG) captures the electrical activity of muscles through the skin, offering a non-invasive pathway for interpreting human motor intent. Despite its potential for advanced human-machine interfaces (HMIs), sEMG signals are inherently weak, noisy, and subject to physiological variability. Traditional gesture recognition systems often struggle with feature sparsity, the difficulty of distinguishing between subtle, similar movements, and the failure to account for the temporal dynamics of muscle activation. This research introduces the MCBAM-GRU architecture to address these challenges, aiming for high-accuracy, real-time control of prosthetic devices.
Proposed Methodology: The MCBAM-GRU Architecture
The core of the proposed system is the Multi-stream Convolutional Block Attention Module-Gate Recurrent Unit (MCBAM-GRU). This architecture is designed to process multi-modal sensor data through several specialized stages.
Multi-Stream Feature Extraction
Rather than processing all sensor data in a single stream, the model utilizes six independent network branches. This multi-stream approach extracts features from individual sEMG and acceleration (ACC) channels separately, preventing early-stage feature interference and ensuring that specific muscle signatures are preserved.
1D-CBAM and Temporal Modeling
The team adapted the Convolutional Block Attention Module (CBAM) for 1D time-series data. This module applies "Channel Attention" to prioritize informative muscle channels and "Temporal Attention" to focus on critical time frames within a gesture. To capture the sequential nature of muscle activation, a Gated Recurrent Unit (GRU) follows the convolutional layers. GRUs are computationally efficient compared to LSTMs while effectively managing the temporal dependencies required for dynamic gesture recognition.
Data Optimization and Sensor Fusion
A critical component of the methodology involves identifying and removing redundant information. By analyzing the redundancy rate across channels, the researchers were able to prune the input data, which improved both recognition accuracy and system latency.
Furthermore, the fusion of sEMG with accelerometer (ACC) data provides the model with a comprehensive view of the movement—combining the electrical "intent" from the muscles with the physical "result" of the limb's motion. This fusion is particularly effective for distinguishing dynamic gestures that might have similar sEMG profiles but different spatial trajectories.
Experimental Results and Discussion
The model was validated using both a custom dataset of 52 gestures from 10 subjects and the benchmark Ninapro DB1 dataset.
- Accuracy: The MCBAM-GRU achieved 94.1% accuracy on the custom dataset and 89.7% on Ninapro DB1. This significantly outperformed established methods such as Random Forests (75.2%) and standard CNNs (79.5%).
- Real-Time Performance: The system maintains a processing delay of less than 300 ms. This is a critical threshold for "seamless" interaction, as demonstrated by the successful real-time control of a Seed Robotics SR-RH8D humanoid hand.
- Ablation Studies: Removing either the GRU or the CBAM modules resulted in accuracy drops of 2.4% to 4%, confirming that both attention mechanisms and temporal modeling are essential for high-performance gesture classification.
Conclusion
The MCBAM-GRU framework represents a significant step forward in sEMG-based gesture recognition. By combining multi-stream attention with temporal modeling and sensor fusion, the system achieves high accuracy across a large vocabulary of 52 gestures while maintaining the low latency required for real-world prosthetic control. Future work will focus on improving the model's robustness against muscle fatigue and sensor shift, as well as optimizing the architecture for deployment on embedded hardware within prosthetic limbs.
Summary updated