Introduction. This thesis aims to investigate the performance of free-model techniques in continuous wrist motion estimation, using forearm muscle EMG signals. More in details, we compare the performance of feature-based models and signal-driven models in predicting two different types of motion: wrist flexion/extension and wrist adduction/abduction. Differently from the common literature set-up, in which single-unite surface electrodes are used, we acquire forearm EMG signal using a wearable device. Comparison between our results and similar literature study results is successively done to assess the validity of this set-up in wrist angle estimation experiment. Material and Methods. 6 healthy subjects are asked to complete two different wrist motion tasks: firstly, a flexion/extension task and in a second time an adduction/abduction task. Forearm muscles EMGs are acquired using OYMotion GForcePro+, while wrist cinematic ones, used as regression targets, are acquired using a next generation inertial measurement unit (NGIMU). Once the EMG signals are processed and segmented, 4 time-domain features (Hudgins’ set) and one frequency-domain features are extracted. These latter are used to train and validate the model outcoming from Random Forest (RF) algorithm, while a Convolutional Neural Network (CNN) is trained and validate using the processed EMG signals. Performance metrics are calculated to evaluate the estimation performances both in training and testing phase. Results and Discussion. The outcomes of this study show the CNN has better performances with respect to the RF algorithm, both in training and testing phase and in prediction of both wrist motion tasks. The lower performance in feature-based model (RF) can be attributed to the extraction of low discriminative features, despite Hudgins’ set is considered a gold standard approach in this type of studies. Moreover, comparing the obtained performance metrics with those found in similar literature studies, the implementation of wearable device is demonstrated to be a valid option to surface electrodes in wrist motion estimation experiment.
Introduction. This thesis aims to investigate the performance of free-model techniques in continuous wrist motion estimation, using forearm muscle EMG signals. More in details, we compare the performance of feature-based models and signal-driven models in predicting two different types of motion: wrist flexion/extension and wrist adduction/abduction. Differently from the common literature set-up, in which single-unite surface electrodes are used, we acquire forearm EMG signal using a wearable device. Comparison between our results and similar literature study results is successively done to assess the validity of this set-up in wrist angle estimation experiment. Material and Methods. 6 healthy subjects are asked to complete two different wrist motion tasks: firstly, a flexion/extension task and in a second time an adduction/abduction task. Forearm muscles EMGs are acquired using OYMotion GForcePro+, while wrist cinematic ones, used as regression targets, are acquired using a next generation inertial measurement unit (NGIMU). Once the EMG signals are processed and segmented, 4 time-domain features (Hudgins’ set) and one frequency-domain features are extracted. These latter are used to train and validate the model outcoming from Random Forest (RF) algorithm, while a Convolutional Neural Network (CNN) is trained and validate using the processed EMG signals. Performance metrics are calculated to evaluate the estimation performances both in training and testing phase. Results and Discussion. The outcomes of this study show the CNN has better performances with respect to the RF algorithm, both in training and testing phase and in prediction of both wrist motion tasks. The lower performance in feature-based model (RF) can be attributed to the extraction of low discriminative features, despite Hudgins’ set is considered a gold standard approach in this type of studies. Moreover, comparing the obtained performance metrics with those found in similar literature studies, the implementation of wearable device is demonstrated to be a valid option to surface electrodes in wrist motion estimation experiment.
EMG-data Driven Model for Wrist Joint Motion Estimation
VERDE, GIACOMO
2020/2021
Abstract
Introduction. This thesis aims to investigate the performance of free-model techniques in continuous wrist motion estimation, using forearm muscle EMG signals. More in details, we compare the performance of feature-based models and signal-driven models in predicting two different types of motion: wrist flexion/extension and wrist adduction/abduction. Differently from the common literature set-up, in which single-unite surface electrodes are used, we acquire forearm EMG signal using a wearable device. Comparison between our results and similar literature study results is successively done to assess the validity of this set-up in wrist angle estimation experiment. Material and Methods. 6 healthy subjects are asked to complete two different wrist motion tasks: firstly, a flexion/extension task and in a second time an adduction/abduction task. Forearm muscles EMGs are acquired using OYMotion GForcePro+, while wrist cinematic ones, used as regression targets, are acquired using a next generation inertial measurement unit (NGIMU). Once the EMG signals are processed and segmented, 4 time-domain features (Hudgins’ set) and one frequency-domain features are extracted. These latter are used to train and validate the model outcoming from Random Forest (RF) algorithm, while a Convolutional Neural Network (CNN) is trained and validate using the processed EMG signals. Performance metrics are calculated to evaluate the estimation performances both in training and testing phase. Results and Discussion. The outcomes of this study show the CNN has better performances with respect to the RF algorithm, both in training and testing phase and in prediction of both wrist motion tasks. The lower performance in feature-based model (RF) can be attributed to the extraction of low discriminative features, despite Hudgins’ set is considered a gold standard approach in this type of studies. Moreover, comparing the obtained performance metrics with those found in similar literature studies, the implementation of wearable device is demonstrated to be a valid option to surface electrodes in wrist motion estimation experiment.File | Dimensione | Formato | |
---|---|---|---|
Thesis_GiacomoVerde.pdf
Open Access dal 26/08/2023
Descrizione: Documento tesi
Dimensione
2.69 MB
Formato
Adobe PDF
|
2.69 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.12075/8003