TL;DR
This paper introduces a novel wearable Functional Electrical Stimulation (FES) system that uses customized textile electrodes to simultaneously recognize lower limb motion and muscle fatigue. By leveraging a deep learning model (FES-sEMGNet) that processes sEMG time-frequency images, the system aims to improve FES usability and safety for rehabilitation and sports training.
What Problem Does This Paper Address?
Functional Electrical Stimulation (FES) is a valuable tool for rehabilitation and sports training, helping to induce muscle movement and improve strength. However, current FES devices suffer from several limitations:
- Poor Wearability: They are often bulky, uncomfortable, and not suitable for daily activities or prolonged use.
- Lack of User Intent Recognition: Existing systems are typically passive and cannot adapt stimulation based on the user's intention to move. This limits active engagement and can lead to suboptimal training.
- Inability to Detect Muscle Fatigue: Without knowing when a muscle is fatigued, FES can lead to overexertion, muscle damage, spasms, and even falls.
- Skin Irritation: Traditional gel electrodes can cause discomfort and allergic reactions with prolonged use.
These issues hinder the seamless integration of FES into users' daily lives for effective and safe rehabilitation and training.
Methods
The researchers developed a wearable FES system with two main components: a lower machine (the wearable FES device) and a host machine (a PC).
- Wearable FES Device (Lower Machine):
- Customized Textile Electrodes: Replaced traditional gel electrodes with flexible, breathable, and stretchable textile electrodes made from Ag/AgCl conductive yarns. These are designed to conform to the body and reduce skin irritation, making them suitable for long-term wear.
- sEMG-FES Module: A miniaturized module containing an MCU, boost circuit, and sEMG sensors to acquire sEMG signals and generate FES pulses. It offers independent four-channel FES output with adjustable parameters (frequency, pulse width, amplitude).
- Host Machine (PC):
- GUI Control Interface: For initial configuration of FES parameters.
- Data Preprocessing Module: Segments sEMG signals, filters out FES artifacts using Empirical Mode Decomposition (EMD) and a notch filter, and converts the clean sEMG signals into time-frequency images.
- Recognition Module: Employs a novel deep learning model called FES-sEMGNet to simultaneously recognize lower limb motion and muscle fatigue status.
FES-sEMGNet Model: This is a hybrid deep learning architecture designed for efficiency and simultaneous output.
- Input: sEMG time-frequency images (generated using STFT, CWT, or HHT).
- Feature Extraction: A CNN-based module with four parallel simplified CNNs extracts features from the time-frequency images.
- Simultaneous Recognition: The extracted features are fed into two separate recognition modules:
- Lower Limb Motion Recognition Module: Uses LSTM networks for feature fusion and classification.
- Muscle Fatigue Status Recognition Module: Also uses LSTM networks for feature fusion and classification.
- Output: The model simultaneously outputs the recognized lower limb motion and muscle fatigue status.
Datasets
- Newly Introduced Benchmark: The study collected a dataset of sEMG signals from five healthy adult males performing six lower limb motions (sitting, walking, climbing stairs, ankle dorsiflexion, ankle plantarflexion, and cycling) under three muscle fatigue conditions (no fatigue, medium fatigue, extreme fatigue). This dataset was collected using the developed wearable FES system.
- Data Split: The collected data was randomly divided into training, testing, and validation sets using an 8:1:1 ratio.
Key Insights and Ideas
- Textile Electrodes for Comfort and Durability: The shift from gel to textile electrodes is a significant step towards improving wearability and reducing skin irritation, making FES more practical for daily use.
- Simultaneous Motion and Fatigue Recognition: The FES-sEMGNet model's ability to perform both motion recognition and fatigue detection concurrently is a key innovation. This allows for adaptive FES control and enhanced safety in a single system.
- HHT for Superior Time-Frequency Representation: The Hilbert Huang Transform (HHT) proved to be the most effective method for generating time-frequency images of sEMG signals, leading to higher accuracy in both motion and fatigue recognition compared to STFT and CWT. This highlights the importance of choosing the right signal processing technique for complex biological signals.
- Deep Learning for Feature Extraction: The FES-sEMGNet model eliminates the need for manual feature engineering, which is a laborious and often suboptimal process in traditional machine learning approaches. This allows the model to learn optimal features directly from the data.
- Lightweight and Parallel Architecture: The FES-sEMGNet model is designed to be computationally efficient, making it suitable for potential embedded deployment on wearable devices. The parallel structure allows for simultaneous processing of different tasks.
Key Results
- Lower Limb Motion Recognition: The FES-sEMGNet model using HHT time-frequency images achieved an overall accuracy of 93.33%. This significantly outperformed traditional methods using LDA with time-domain features (average accuracy of 87.08%).
- Muscle Fatigue Status Recognition: The FES-sEMGNet model with HHT inputs achieved an accuracy of 89.02% for muscle fatigue status recognition, outperforming the 4TFFs-LDA method (83.87%).
- Performance Comparison: HHT consistently yielded better results than STFT and CWT for both motion and fatigue recognition tasks across various metrics (precision, recall, F1 score).
- User Feedback: Subjects reported the system was easy and comfortable to wear, did not interfere with aerobic exercise, and no skin allergies were observed.
Limitations and Future Work
- FES Interference: While artifact filtering was applied, FES interference still slightly impacts the temporality of sEMG signals, leading to marginally lower accuracy in fatigue recognition compared to motion recognition. Future work should focus on better sEMG signal extraction techniques under FES interference.
- Subjective Fatigue Judgment: The slight decrease in F1 scores for medium and extreme fatigue might be influenced by the subjective nature of the Borg CR10 scale used for self-assessment.
- Hardware Integration: Currently, the FES-FES module and power supply are connected to the lower leg via wires, which can be a point of failure or inconvenience. Future work should aim to integrate these components into leggings or shoes for improved flexibility and user acceptance.
- Clinical Validation: The researchers plan to conduct clinical trials to validate the practical rehabilitation effects of the system.
- Edge Deployment: The goal is to deploy the recognition model on edge computing devices (FES devices, smartphones) to enable real-time operation without reliance on laboratory environments.
Summary updated