Evaluation of the effectiveness of PCA and ICA in improving muscle movement recognition from raw EMG signals

Evaluation of the effectiveness of PCA and ICA in improving muscle movement recognition from raw EMG signals

In the last decade, the development of classification models through machine learning for the control of multifunctional prosthetic devices has been increasing. Electromyography (EMG) are recordings produced by muscle fibers naturally when performing movements; if modeled, they could play a more act...

Ful tanımlama

Kaydedildi:
Detaylı Bibliyografya
Asıl Yazarlar: Amezquita Garcia, Jose Alejandro, Bravo Zanoguera, Miguel Enrique, Fabian Natanael
Materyal Türü: Online
Dil:spa
Baskı/Yayın Bilgisi: Universidad Autónoma de Baja California 2023
Online Erişim:https://recit.uabc.mx/index.php/revista/article/view/318
Etiketler: Etiketle
Etiket eklenmemiş, İlk siz ekleyin!
id recit-article-318
record_format ojs
institution RECIT
collection OJS
language spa
format Online
author Amezquita Garcia, Jose Alejandro
Bravo Zanoguera, Miguel Enrique
Fabian Natanael
spellingShingle Amezquita Garcia, Jose Alejandro
Bravo Zanoguera, Miguel Enrique
Fabian Natanael
Evaluation of the effectiveness of PCA and ICA in improving muscle movement recognition from raw EMG signals
author_facet Amezquita Garcia, Jose Alejandro
Bravo Zanoguera, Miguel Enrique
Fabian Natanael
author_sort Amezquita Garcia, Jose Alejandro
title Evaluation of the effectiveness of PCA and ICA in improving muscle movement recognition from raw EMG signals
title_short Evaluation of the effectiveness of PCA and ICA in improving muscle movement recognition from raw EMG signals
title_full Evaluation of the effectiveness of PCA and ICA in improving muscle movement recognition from raw EMG signals
title_fullStr Evaluation of the effectiveness of PCA and ICA in improving muscle movement recognition from raw EMG signals
title_full_unstemmed Evaluation of the effectiveness of PCA and ICA in improving muscle movement recognition from raw EMG signals
title_sort evaluation of the effectiveness of pca and ica in improving muscle movement recognition from raw emg signals
description In the last decade, the development of classification models through machine learning for the control of multifunctional prosthetic devices has been increasing. Electromyography (EMG) are recordings produced by muscle fibers naturally when performing movements; if modeled, they could play a more active role in this type of control. These signals are used to control devices/applications. The problem with these models is the stochastic nature of the signal, the variability between subjects and the inherent cross-communication that makes them inaccurate when faced with a high number of movements. The stochastic nature and variability of the signal are already widely studied, however, there are still no definitive results that describe generalizable movement classification models. Here, two databases available on the CapgMyo network and the Ninapro project are studied, their characteristics are evaluated, with the objective of investigating the variability of the muscle signal between subjects, the factors that modify it and how the use of analysis affects principal components (PCA) and independent component analysis (ICA) to EMG information in classification models. A comparison was made between the results in terms of recognition percentages of classic machine learning methods such as linear discriminant analysis (LDA) and quadratic analysis (QDA) using transformation techniques to new spaces introducing the possibility of performing a dimensionality reduction. with PCA and ICA, algorithms usually used to solve problems such as blind source separation (BSS), which is applicable to the phenomenon presented in muscle signals and their acquisition through surface electrodes. The results can be evaluated through the recognition percentage of the classification models created, these show that for raw EMG signals the PCA and ICA methods are useful to perform a reduction in the dimensionality of the data without providing a significant increase in the recognition percentages. It was shown that the recognition percentages in the classification of movements for the Capgmyo database were higher thanks to the characteristics that define it, a higher recognition percentage was obtained ranging from 72.5% to 87.9% with QDA, and 82.8% to 90% for QDA with PCA. The main contribution is the evaluation of the effectiveness of algorithms such as PCA and ICA in machine learning tasks with raw EMG data. Future work is to lay the foundations to reduce the effects of cross-communication in EMG recordings.
publisher Universidad Autónoma de Baja California
publishDate 2023
url https://recit.uabc.mx/index.php/revista/article/view/318
_version_ 1792095386987397120
spelling recit-article-3182024-02-13T02:55:52Z Evaluation of the effectiveness of PCA and ICA in improving muscle movement recognition from raw EMG signals Evaluación de la eficacia de PCA e ICA en la mejora del reconocimiento de movimientos musculares a partir de señales EMG crudas Amezquita Garcia, Jose Alejandro Bravo Zanoguera, Miguel Enrique Fabian Natanael EMG Comunicación cruzada Aprendizaje Automático Minería de datos EMG Cross communication Machine Learning Data mining In the last decade, the development of classification models through machine learning for the control of multifunctional prosthetic devices has been increasing. Electromyography (EMG) are recordings produced by muscle fibers naturally when performing movements; if modeled, they could play a more active role in this type of control. These signals are used to control devices/applications. The problem with these models is the stochastic nature of the signal, the variability between subjects and the inherent cross-communication that makes them inaccurate when faced with a high number of movements. The stochastic nature and variability of the signal are already widely studied, however, there are still no definitive results that describe generalizable movement classification models. Here, two databases available on the CapgMyo network and the Ninapro project are studied, their characteristics are evaluated, with the objective of investigating the variability of the muscle signal between subjects, the factors that modify it and how the use of analysis affects principal components (PCA) and independent component analysis (ICA) to EMG information in classification models. A comparison was made between the results in terms of recognition percentages of classic machine learning methods such as linear discriminant analysis (LDA) and quadratic analysis (QDA) using transformation techniques to new spaces introducing the possibility of performing a dimensionality reduction. with PCA and ICA, algorithms usually used to solve problems such as blind source separation (BSS), which is applicable to the phenomenon presented in muscle signals and their acquisition through surface electrodes. The results can be evaluated through the recognition percentage of the classification models created, these show that for raw EMG signals the PCA and ICA methods are useful to perform a reduction in the dimensionality of the data without providing a significant increase in the recognition percentages. It was shown that the recognition percentages in the classification of movements for the Capgmyo database were higher thanks to the characteristics that define it, a higher recognition percentage was obtained ranging from 72.5% to 87.9% with QDA, and 82.8% to 90% for QDA with PCA. The main contribution is the evaluation of the effectiveness of algorithms such as PCA and ICA in machine learning tasks with raw EMG data. Future work is to lay the foundations to reduce the effects of cross-communication in EMG recordings. En la última década el desarrollo de modelos de clasificación a través de aprendizaje automático para control de dispositivos protésicos multifuncionales ha ido en aumento. La electromiografía (EMG) son registros producidos por las fibras musculares de forma natural al realizar movimientos, de modelarse podrían tener un papel de forma más activa en este tipo de control. Estas señales son utilizadas para control de dispositivos/aplicaciones, el problema con estos modelos es la naturaleza estocástica de la señal, la variabilidad entre sujetos y la comunicación cruzada inherente que los vuelve inexactos ante un número alto de movimientos. La naturaleza estocástica y la variabilidad de la señal ya son ampliamente estudiadas, sin embargo, no existen aún resultados definitivos que describan modelos de clasificación de movimientos generalizables. Aquí se estudian dos bases de datos disponibles en la red CapgMyo y the Ninapro project, se evalúan las características de estas, teniendo como objetivo investigar la variabilidad de la señal muscular entre sujetos, los factores que la modifican y como afecta el uso de análisis de componentes principales (PCA) y el análisis de componentes independientes (ICA) a la información del EMG en modelos de clasificación. Se realizó una comparación entre los resultados en términos de porcentajes de reconocimiento de métodos clásicos de aprendizaje automático como el análisis discriminante lineal (LDA) y el cuadrático (QDA) utilizando técnicas de trasformación a nuevos espacios introduciendo la posibilidad de realizar una reducción de la dimensionalidad con PCA e ICA, algoritmos usualmente utilizados para resolver problemas como la separación ciega de fuentes (BSS) que es aplicable al fenómeno presentado en  señales musculares y su adquisición a través de electrodos superficiales. Los resultados pueden evaluarse a través del porcentaje de reconocimiento de los modelos de clasificación creados, estos muestran que para señales crudas de EMG los métodos de PCA e ICA son útiles para realizar una reducción de la dimensionalidad de los datos sin aportar un aumento significativo en los porcentajes de reconocimiento. Se demostró que los porcentajes de reconocimiento en la clasificación de los movimientos para la base de datos Capgmyo fueron superiores gracias a las características que la definen, se obtuvo un mayor porcentaje de reconocimiento que va del 72.5% al 87.9% con QDA, y del 82.8 al 90% para QDA con PCA. La aportación principal es la evaluación de la eficacia de algoritmos como PCA e ICA en tareas de aprendizaje automático con datos crudos de EMG. Como trabajo futuro esta ir plasmando las bases para reducir los efectos de la comunicación cruzada en los registros de EMG. Universidad Autónoma de Baja California 2023-10-06 info:eu-repo/semantics/article info:eu-repo/semantics/publishedVersion application/pdf text/html application/zip https://recit.uabc.mx/index.php/revista/article/view/318 10.37636/recit.v6n4e318 REVISTA DE CIENCIAS TECNOLÓGICAS; Vol. 6 No. 4 (2023): October-December; e318 REVISTA DE CIENCIAS TECNOLÓGICAS; Vol. 6 Núm. 4 (2023): Octubre-Diciembre; e318 2594-1925 spa https://recit.uabc.mx/index.php/revista/article/view/318/485 https://recit.uabc.mx/index.php/revista/article/view/318/486 https://recit.uabc.mx/index.php/revista/article/view/318/487 Copyright (c) 2023 José Alejandro Amézquita García, Miguel Enrique Bravo Zanoguera, Fabián Natanael Murrieta-Rico https://creativecommons.org/licenses/by/4.0

Benzer Materyaller