Abstract: Estimation of hand kinematics from surface electromyographic (sEMG) signals provides a non-invasive human-machine interface. This approach is usually subject-specific, so that the training on one individual does not generalise to different subjects. In this paper, we propose a method based on Bidirectional Encoder Representation from Transformers (BERT) structure to predict the movement of hands from the root mean square (RMS) feature of the sEMG signal following $\mu$-law normalization. The method was tested for within-subject and cross-subject conditions. We trained the model with two hard sample mining methods, Gradient Harmonizing Mechanism (GHM) and Online Hard Sample Mining (OHEM). The proposed method was compared with classic approaches, including long short-term memory (LSTM) and Temporal Convolutional Network (TCN) as well as a recent method called Long Exposure Convolutional Memory Network (LE-ConvMN). Correlation coefficient (CC), normalized root mean square error (NRMSE) and time costs were used as performance metrics. Our method (sBERT-OHEM) achieved state-of-the-art performance in cross-subject evaluation as well as high performance in subject-specific tests on the Ninapro dataset.
1. Introduction
Surface electromyography (sEMG) records muscle electrical activity non-invasively, offering a promising interface for controlling prosthetic limbs and exoskeletons. Despite its potential, two primary obstacles hinder widespread clinical adoption: subject specificity and the complexity of continuous regression. Subject specificity refers to the inability of models trained on one individual to generalize to others due to physiological differences. Continuous regression, unlike discrete classification, requires the precise estimation of joint angles over time, which is highly sensitive to signal noise and temporal shifts.
2. Proposed Methodology
The authors introduce sBERT-OHEM, a framework leveraging the Bidirectional Encoder Representation from Transformers (BERT). The pipeline incorporates several key innovations:
- $\mu$-law Normalization: This non-linear transformation magnifies low-magnitude signals while compressing spikes, ensuring the model captures subtle muscle activations often lost in standard linear scaling.
- BERT Architecture: By utilizing self-attention mechanisms, the model processes temporal windows bidirectionally. This allows the system to utilize both past and future context within a frame to estimate the current joint angle.
- Hard Sample Mining: The use of Online Hard Example Mining (OHEM) and Gradient Harmonizing Mechanism (GHM) forces the model to prioritize high-error data points during training, improving overall robustness.
3. Experimental Design
The model was validated using the Ninapro DB2 dataset. The input consisted of 12 sEMG channels, while the target outputs were 10 finger joint angles recorded via a CyberGlove II data glove. The experimental protocol involved six distinct grasping movements.
4. Results and Discussion
The sBERT-OHEM model achieved a Correlation Coefficient (CC) of 0.84 in cross-subject tests, establishing a new state-of-the-art for the Ninapro dataset. It outperformed classic architectures such as Long Short-Term Memory (LSTM) and Temporal Convolutional Networks (TCN).
- Efficiency: BERT's parallel processing capabilities allowed for a training time of 0.5 hours, significantly faster than the 2.5 hours required for LSTMs.
- Scalability: Unlike traditional methods that degrade as more subjects are added, the BERT-based approach maintained high performance when scaled from 10 to 38 subjects.
5. Conclusion
This study demonstrates that the bidirectional attention mechanisms of BERT are highly effective for decoding complex muscle-to-motion relationships. Future work will focus on reducing inference latency for real-time applications and testing the model's performance on unstructured, real-world movements.
Summary updated