Presentation description

The long-term goal of this research is to improve the performance of neural engineering algorithms by developing methods to quantify fatigue in real-time.
Muscle activity, as measured via surface electromyography (sEMG), is widely used in neural engineering to monitor muscle activity and control assistive robots such as prostheses and exoskeletons. It is well established that fatigue progressively alters muscle activity, often impairing the effectiveness of control algorithms due to a problem known as domain shift. Therefore, the ability to reliably characterize fatigue-related features in sEMG signals to compensate for domain shift could significantly enhance the robustness and adaptability of these algorithms. However, there remains no standardized or reliable method for quantifying muscle fatigue during dynamic, concentric movements, such as those involved in walking or running.
Here, we explore a method for continuously quantifying muscle fatigue from sEMG data using a hybrid convolutional-recurrent artificial neural network (ANN). The ANN is trained to predict muscle fatigue trends based on extracted sEMG signal features, including Mean Absolute Value (MAV), signal envelope (ENV), and Discrete Wavelet Transform (DWT) coefficients. These features are computed over time and formatted as sequences of 3D image arrays to enable learning of both spatial and temporal patterns.
To validate the network's predictions, participants self-report their perceived exertion using a Visual Analog Scale (VAS) during three distinct calf muscle exercises: (1) dynamic concentric ankle extensions, (2) isometric bodyweight ankle extensions, and (3) a combined dynamic-isometric sequence. Interpolated VAS scores serve as ground truth for fatigue levels.
The resulting model is expected to learn individual-specific markers of muscle fatigue and generalize well across sessions. This model could enhance adaptive control for assistive robotics and fatigue monitoring.