Embedded Machine Learning for EMG-Based Elbow Motion Recognition

This research introduces a novel setup for capturing, processing, and analyzing electromyographic (EMG) signals for elbow motion recognition. The setup, comprising a three-lead EMG cable assembly, an amplifier PCB, and an Arduino Nano BLE board, implements embedded machine learning and power conservation. The study applies Discrete Wavelet Transform (DWT) for spectral analysis and feature extraction, aiding in EMG signal denoising. It utilizes the XGBoost Random Forest system for model training and K-Means Clustering for anomaly detection. The research evaluates the performance of the Edge Impulse framework, focusing on the model's performance, live classification results on the Arduino Nano BLE board, and power consumption during continuous operation. The model's offline performance was assessed using thirty additional samples, achieving an overall accuracy of 74.15%. The high performance in the “Idle” category suggests potential for a power-saving method by transitioning the system to a sleep state when idle. This evaluation provides insights for further model optimization and refinement. During the test runs conducted on the Arduino Nano BLE board, the inferencing time was measured to be approximately 50 ms. The program's power consumption was approximately 2 mAh, suggesting that it could be feasibly powered by a small rechargeable battery (approximately 200 mAh) for daily usage. This research affirms the potential of Edge Impulse and similar platforms in the field of embedded machine learning.