Federated Learning-Enhanced Edge Deep Learning Model for EMG-Based Gesture Recognition in Real-Time Human-Robot Interaction

• Published in: IEEE sensors journal Vol. 25; no. 5; pp. 9139 - 9151
• Main Authors: Zafar, Muhammad Hamza, Moosavi, Syed Kumayl Raza, Sanfilippo, Filippo
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.03.2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Bandpass filters; Datasets; Deep learning; Deep learning (DL); edge devices; Electromyography; Feature extraction; Federated learning; federated learning (FL); Fingers; Gesture recognition; Hands; Human engineering; Human-robot interaction; human-robot interaction (HRI); Machine learning; Muscles; Neural networks; Noise; Real time; Rescue operations; Robot learning; Robots; Search and rescue missions; Sensors; Signal quality
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2025.3529841

Electromyography (EMG)-based gesture detection plays a crucial role in human-robot interaction (HRI), providing a seamless interface for controlling robotic systems through muscle activity. Despite its potential, EMG systems face significant challenges related to the security and privacy of sensitive biometric data, as well as the computational limitations of deploying deep learning (DL) models on edge devices. To address these issues, we propose a federated learning (FL)-based DL model for EMG gesture recognition, specifically designed for edge devices. Our model utilizes a custom dataset collected using a Mindrove eight-channel EMG armband, capturing eight distinct hand gestures-rest, move left, move right, move down, move up, open fingers, close fist, and twist hand-from ten subjects with seven repetitions each, ensuring diverse and robust data for training. During preprocessing, a bandpass filter (50-450 Hz) was applied to remove noise and enhance signal quality, followed by a short-term frequency transform (STFT) with a 200-ms sample time and 50% overlap to extract relevant features from the EMG signals. The dataset was segmented into training and testing sets with a 70/30 split for evaluation. We evaluate several FL techniques, including FedAvg, FedProx, and FedSGD, demonstrating that FedAvg achieves the highest accuracy of 96.92% without quantization with Scenario 9 (15 epochs, 20 rounds) with minimal communication overhead. Additionally, our model is quantized, resulting in an 89% reduction in size and a high accuracy of 95.99%, representing a minimal loss of 0.93%, making it ideal for edge deployment without compromising performance. A comparative analysis with other DL models, such as multiconvolutional residual networks (MCRNs), multiconvolutional neural networks (MCNNs), temporal convolutional networks (TCNs), and InceptionNet, shows that our approach outperforms these models in both accuracy and efficiency. Experimental results validate the high accuracy of our model in both training/testing and real-time disaster scenario simulations using the Spot robot. The proposed solution provides a secure, efficient, and highly accurate framework for EMG-based gesture recognition on edge devices, ideal for HRI and assistive technologies such as in search and rescue operations.