EMG-based gesture recognition: CNN feature contribution and ablation study

• Main Author: Pinheiro, Miguel Marques Alves Maya
• Publication Date: 2024
• Format: Master thesis
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: http://hdl.handle.net/10400.26/53179
• topic: Electromyography (EMG); Surface EMG (sEMG); EMG Signal Processing; Feature extraction; Gesture recognition; Prosthetic control; Machine learning; Deep learning; Convolutional neural networks; Ablation study; Model recovery

Electromyography (EMG)has emerged as a pivotal technology in biomedical engineering, offering valuable insights into muscle activity with significant applications in adaptive prosthetics. This thesis delves into the application of EMG signal classification to enhance gesture recognition for prosthetic control.Utilizing the GRABMyo dataset, which includes comprehensive EMG data from 43 participants performing various hand and wrist gestures, this study provides a robust platform for testing and validating classification algorithms. The data set was meticulously organized, ensuring clear separation between training and testing sets to facilitate rigorous evaluation.There search focuses on the implementation of Convolutional Neural Networks (CNNs) withan AleNet architecture for their superior feature extraction capabilities. An ablation study was conducted to systematically assess the influence of different neural network components on the model’s performance. By selectively disabling specific filters within the convolutional layers, the study aimed to identify the most critical features and components that significantly contribute to accurate gesturer ecognition.The objectives of this ablation study include optimizing the network by focusing on impactful elements and improving gesture classification accuracy. Following the ablation study, a recovery phase was implemented to restore and potentially enhance the model’s performance. This phase tested the network’sr esilience and adaptability, providing insights into its built-in redundancy and ability to compensate for lost functionalities through targeted training strategies. The recovery effort not only aimed to regain lost performance but also to deepen the understanding of which features and layers are most critical to the network’s operation. The fina phase involved validating the optimized model on a test set comprising participants not included in the training phase. To visually demonstrate the model’s effectiveness, the Unity engine was used to simulate the gestures based on the model’s classifications, offering a nintuitive and interactive representation of its capabilities.This visualization underscores the practical applications of the technology in adaptive prosthetics, bridging the gap between theoretical research and real-world implementation.