Efficient Spiking Neural Network for Cross-Subject Lower Limb Joint Angle Estimation from Surface Electromyography

Concepts de base

A hierarchical spiking neural network architecture, SAFE-Net, is proposed to efficiently estimate lower limb joint angles from surface electromyography signals, achieving high accuracy in cross-subject scenarios.

Résumé

The content presents a novel neural network architecture, SAFE-Net, for continuous joint angle estimation from surface electromyography (sEMG) signals. SAFE-Net consists of two main components:

1. Spike-driven Sparse Attention Encoder (SSAE):

• Compresses sEMG signals into coarse-grained features using spiking neurons and sparse attention mechanism
• Significantly reduces inference costs in terms of power consumption, latency, and computational complexity compared to existing Transformer-based methods

2. Spiking Attentional Feature Decomposition (SAFD) module:

• Decomposes the compressed features into fine-grained kinematic and biological features
• The kinematic features are used for joint angle regression, while the biological features are used for subject identity recognition
• Enhances the model's generalization ability in cross-subject scenarios

Experiments on two datasets, SIAT-DB1 and SIAT-DB2, demonstrate that SAFE-Net outperforms existing methods in both recognition accuracy and inference efficiency. The proposed architecture shows great potential for applications in lower limb rehabilitation exoskeletons with synchronous and proportional control.

Stats

The RMSE of hip-knee-ankle joint angle prediction on DB1 dataset:
SSAE: 4.02, 4.28, 5.79
Informer: 5.52, 5.94, 7.23
Spikformer: 9.99, 9.95, 9.96
The RMSE of hip-knee-ankle joint angle prediction on DB2 dataset:
SSAE: 3.63, 3.39, 2.97
Informer: 4.14, 4.93, 4.45
Spikformer: 8.58, 8.90, 8.33

Citations

"SSAE reduced the average RMSE for hip-knee-ankle by 27.3%, 28.0%, and 19.8% compared to Informer, and by 59.8%, 57.0%, and 41.9% compared to Spikformer."
"Compared to Informer, SSAE reduced by 60.8%, 34.9%, 39.1%, 60.8%, and 51.09% across the five metrics. Compared to Spikformer, SSAE reduced by 61.8%, 38.3%, 37.5%, 61.3%, and 51.62%."

Idées clés tirées de

Efficient sEMG-based Cross-Subject Joint Angle Estimation via Hierarchical Spiking Attentional Feature Decomposition Network

by Xin Zhou,Chu... à arxiv.org 04-12-2024

https://arxiv.org/pdf/2404.07517.pdf

Questions plus approfondies

How can the proposed SAFE-Net architecture be extended to handle more complex and variable human movement patterns beyond the cyclic motions studied in this work?

To extend the SAFE-Net architecture to handle more complex and variable human movement patterns, several strategies can be implemented:

Data Collection: Gather a diverse range of data encompassing various movement patterns, including irregular and spontaneous movements, to train the model on a wider spectrum of activities.

Feature Engineering: Incorporate additional features or sensors to capture more nuanced information about the movements, such as inertial sensors, pressure sensors, or additional muscle groups for sEMG data collection.

Model Adaptation: Modify the architecture to include more layers or modules that can capture the intricacies of complex movements, such as hierarchical attention mechanisms or recurrent neural networks for temporal dependencies.

Transfer Learning: Utilize transfer learning techniques to adapt the model trained on cyclic motions to new movement patterns, leveraging the knowledge gained from the initial training on simpler activities.

Data Augmentation: Augment the existing dataset with variations in speed, direction, intensity, and other parameters to simulate a broader range of movement patterns and enhance the model's generalization capabilities.

By implementing these strategies, SAFE-Net can be enhanced to effectively handle a wider array of human movement patterns beyond the cyclic motions studied in the current work.