TL;DR
This paper introduces SAFE-Net, a novel hierarchical Spiking Attentional Feature Decomposition Network for estimating lower limb joint angles from surface electromyography (sEMG) signals. SAFE-Net significantly reduces inference costs and improves accuracy in multi-subject scenarios by decomposing sEMG into kinematic and biological features using a spiking neural network architecture.
What Problem Does This Paper Address?
Rehabilitation exoskeleton robots hold immense promise for individuals with mobility impairments. A critical component for seamless human-robot collaboration is accurate and real-time motion intention detection. Surface electromyography (sEMG) is a promising non-invasive signal source for this purpose. However, existing sEMG-based joint angle estimation methods often struggle with high computational costs, limiting their real-time applicability, and exhibit poor generalization across different subjects. This heterogeneity in sEMG signals, arising from individual differences, poses a significant challenge for developing robust, single models that work for everyone.
Methods
The core innovation of this paper is the SAFE-Net architecture, which comprises two main modules:
- Spike-driven Sparse Attention Encoder (SSAE): This module compresses raw sEMG signals into neural spiking forms. It leverages a "spike-driven sparse attention" mechanism, which is designed to be more energy-efficient than traditional attention mechanisms by focusing computations on relevant signal parts. This is achieved by filtering attention scores based on their proximity to a uniform distribution, effectively reducing unnecessary computations.
- Spiking Attentional Feature Decomposition (SAFD) module: This hierarchical module takes the compressed features from SSAE and decomposes them into two distinct streams:
- Kinematic Features: These capture the general motion patterns, independent of individual characteristics. These features are then used to predict the joint angles.
- Biological Features: These capture individual-specific characteristics. These features are used for subject identity recognition.
The network uses a Leaky Integrate-and-Fire (LIF) neuron model for spiking, which is known for its energy efficiency. The overall training objective is a weighted sum of regression loss (for joint angles), classification loss (for subject identity), and an orthogonal loss to ensure the kinematic and biological features are independent.
Datasets
The paper evaluates SAFE-Net on two publicly available datasets:
- SIAT-DB1 (DB1): Collected from 8 healthy volunteers walking on a treadmill at different inclines (0°, 5°, 10°) and speeds (3 km/h). sEMG was recorded from five muscles, and hip, knee, and ankle joint angles were captured using IMUs.
- SIAT-DB2 (DB2): Collected from 12 healthy volunteers walking on a treadmill at different speeds (2.5, 3.0, 3.6 km/h). sEMG was recorded from eight muscles, and joint angles were captured using a motion capture system.
Key Insights and Ideas
- Decomposition is Key: The paper strongly suggests that decomposing sEMG signals into distinct kinematic and biological components is crucial for achieving robust multi-subject joint angle estimation. This mirrors how humans might process sensory information, separating general motion from individual nuances.
- Spiking Neural Networks for Efficiency: The adoption of SNNs, particularly with sparse attention, offers a significant advantage in reducing computational load and power consumption, making real-time applications more feasible.
- Attention for Relevance: The "spike-driven sparse attention" mechanism effectively prunes irrelevant computations, leading to substantial efficiency gains without sacrificing accuracy.
- Biological Interpretability: By separating biological features for subject identification, the network gains a degree of interpretability, allowing researchers to understand what aspects of the sEMG signal are indicative of individual differences.
Key Results
- Inference Cost Reduction: SAFE-Net (specifically the SSAE component) achieved significant power consumption savings of 39.1% and 37.5% compared to Informer and Spikformer, respectively, in terms of inference costs.
- Improved Accuracy: SAFE-Net consistently outperformed Informer and Spikformer in joint angle estimation accuracy (measured by RMSE) across both datasets and various walking conditions (slopes and speeds). For instance, on DB1, SAFE-Net reduced average RMSE by 27.3% compared to Informer and 59.8% compared to Spikformer for slope=0.
- Robustness: The model demonstrated good robustness, with recognition errors increasing the least for SAFE-Net as slope or speed increased, compared to other methods.
- Generalization: The SAFD module was shown to significantly improve generalization capabilities in multi-subject scenarios, as evidenced by statistical significance tests and visual T-SNE embeddings of biological features.
Limitations and Future Work
- Cyclic Movement Bias: The current experiments primarily used cyclic movement patterns. The authors acknowledge that human movement is more complex and variable, and future work should explore SAFE-Net's performance in mixed movement scenarios.
- Healthy Subjects Only: The study was conducted on healthy subjects. Future research should validate SAFE-Net's generalizability on individuals with walking disorders and incorporate data from patients.
- Further Elucidation: The authors suggest further research into the physiological and kinematic mechanisms underlying the decomposition of sEMG signals.
- Broader Applications: The framework for decomposing sEMG into distinct features could be applied to other myoelectric pattern recognition tasks.
- SNN Potential: The full potential of spiking neural networks for enhancing computational efficiency is still being explored and warrants further investigation.
Summary updated