Surface electromyography (sEMG) is key for human-robot interaction and assistive exoskeletons, but cross-subject variability often degrades recognition accuracy and increases computational load. Researchers propose OmniXceptionDBN, a multi-scale integrated sequence deep belief network that combines time-domain and frequency-domain processing for robust sEMG classification.
Key takeaways
- Hybrid architecture: Singular spectrum analysis (SSA) and fast Fourier transform (FFT) preprocess raw signals; XceptionTime and OmniScaleCNN extract time and frequency features, which are fused via a deep belief network (DBN) for classification.
- Performance: The method achieves 97.2% accuracy for single-subject classification and 85.9% for multi-subject scenarios without additional adaptation, outperforming traditional approaches in cross-subject generalization.
- Application: Designed for upper-limb movement recognition, the method supports intuitive control of exoskeletons and prosthetics, reducing user-dependent calibration.
By effectively merging temporal and spectral information, OmniXceptionDBN offers an efficient and robust solution for sEMG-based intent recognition, potentially accelerating deployment in industrial and rehabilitation robotics.
Source: 《机器人》期刊 (robot.sia.cn) · Published 2026-03-15 · “基于OmniXceptionDBN的表面肌电信号智能识别方法”