Abstract
This study presents a hybrid deep learning framework, PSO-LSTM, designed to continuously estimate upper limb joint angles from surface electromyography (sEMG) signals. By integrating Particle Swarm Optimization (PSO) with Long Short-Term Memory (LSTM) networks, the model automatically tunes hyperparameters to capture the non-linear temporal dependencies of muscle activity. Experimental results from eight healthy subjects demonstrate that the PSO-LSTM model significantly outperforms traditional back-propagation networks, achieving high goodness-of-fit ($R^2$ up to 0.94) and low error rates for elbow and wrist motion estimation, offering a promising control interface for rehabilitation exoskeletons.
Introduction
For patients suffering from spinal cord injuries (SCI) or stroke-induced paralysis, powered exoskeletons represent a vital technology for restoring mobility. However, a persistent challenge in exoskeleton design is achieving smooth, natural control. Many current systems rely on discrete action classification, which often results in jerky and unnatural movements.
To facilitate fluid, "human-like" motion, robotic systems must interpret a user's motion intention in real-time. Surface electromyography (sEMG)—the electrical signal generated during muscle contraction—provides a rich data source for this intent. However, sEMG signals are notoriously non-linear, non-stationary, and noisy. Mapping these complex electrical bursts to precise, continuous joint angles requires advanced modeling capable of handling high-dimensional temporal data.
Methods
The researchers proposed a PSO-LSTM framework to bridge the gap between raw muscle activity and joint kinematics. The system follows a structured pipeline from signal acquisition to regression modeling.
- Data Acquisition and Pre-processing: sEMG signals were recorded from eight target muscles while joint angles were simultaneously captured using an optical motion capture system (NDI Optotrak) to provide ground truth data.
- Feature Engineering: Seven time-domain and frequency-domain features were initially extracted. To optimize performance, the team employed Spearman correlation analysis to reduce the feature set to the four most informative metrics per muscle, minimizing overfitting and computational overhead.
- LSTM Regression: Given that joint movement is a time-series process, an LSTM network was selected for its ability to retain temporal dependencies and "remember" previous states.
- PSO Optimization: The core innovation is the use of Particle Swarm Optimization to automatically tune the LSTM’s hyperparameters, such as the number of hidden neurons and training iterations. This eliminates the guesswork of manual tuning and ensures the model is optimized for the specific non-linearities of the sEMG data.
Experimental Setup and Datasets
The study utilized a dataset collected from eight healthy male subjects (ages 22–27). The experimental protocol involved monitoring eight specific muscle channels: ED, ECU, FCU, FCR, BM, BB, LTB, and MTB.
Subjects performed five distinct movement types, including Elbow Flexion/Extension (EFE), Elbow Pronation/Supination (EPS), Wrist Flexion/Extension (WFE), Wrist Adduction/Abduction (WAA), and a combined continuous motion (CEW). Joint angles were recorded at a frequency of 100 Hz.
Key Results and Discussion
The PSO-LSTM model demonstrated a significant performance advantage over standard Back Propagation (BP) neural networks across all metrics:
- Estimation Accuracy: For elbow motion, the PSO-LSTM achieved a Root Mean Square Error (RMSE) of 0.0599, vastly superior to the 0.1986 recorded by the BP model. Wrist estimation followed a similar trend, with the PSO-LSTM achieving an RMSE of 0.1025 versus 0.2348 for the BP model.
- Correlation: The model reached an average Goodness of Fit ($R^2$) of 0.9443 for the elbow and 0.8924 for the wrist, indicating a high degree of precision in tracking actual human movement.
- Complexity Insights: The results highlighted that wrist estimation is inherently more complex than elbow estimation. This is attributed to the intricate muscle synergies required for wrist stabilization, whereas the elbow functions as a simpler hinge-like joint.
Conclusion
By combining the temporal processing power of LSTM with the global search capabilities of PSO, this research provides a robust method for continuous motion estimation. While future work is needed to validate the model on clinical populations (such as stroke survivors) and in more complex multi-planar movements, the PSO-LSTM framework marks a significant step toward more responsive and intuitive rehabilitation robotics.
Summary updated