Publicação

Study and application of computational techniques to identify anatomical back postures

Ver documento

Detalhes bibliográficos
Resumo:Currently, the incidence of occupational diseases has been increasing, both in administrative and in production posts, being the musculoskeletal diseases, one of the most common and affect, mainly, the spine. Considering the above-mentioned problem, it is important to develop strategies that enable the recognition and prediction of postures performed by workers in their workplaces. To this end, this thesis proposes a pipeline that aims to build a Machine Learning model that may be capable of recognizing, in real time, the upper body postures of the users. The above described framework provides a model capable of recognizing 6 different static postures (back frontal bending, left lateral back bending, right lateral back bending, neck frontal bending, working overhead, and finally the standing posture). In addition to these static postures, the model is also capable of recognizing the transitions between the neutral posture and each of the other five. Several feature selection algorithms have been tested to find the biomechanical features that best distinguish the different postures. The algorithms tested were the Principal Component Analysis, the Forward Sequential Selection, the Minimum Redundancy Maximum Relevance (mRMR) and the ReliefF. Also, several classification algorithms were tested (Support Vector Machines, K-nearest Neighbors, TreeBagger, Discriminant Analysis, Convolutional Neural Network and Feed Forward Neural Network). These algorithms were trained and tested with data from 50 healthy subjects who volunteered to acquire data realizing the above-mentioned postures. The best results were obtained using the SVM classifier with the quadratic kernel and using the characteristics selected by the mRMR algorithm. The classification model showed promising results during cross-validation, more specifically, it presented an MCC value of 0.973. As future work, other types of sensors should be integrated, and other features should be tested to improve posture classifier performance.
Autores principais:Alpoim, Luís Filipe Beites
Assunto:Machine learning Feature selection Occupational diseases Posture classification Doenças ocupacionais Classificação de posturas Engenharia e Tecnologia::Engenharia Médica
Ano:2019
País:Portugal
Tipo de documento:dissertação de mestrado
Tipo de acesso:acesso aberto
Instituição associada:Universidade do Minho
Idioma:inglês
Origem:RepositóriUM - Universidade do Minho
Descrição
Resumo:Currently, the incidence of occupational diseases has been increasing, both in administrative and in production posts, being the musculoskeletal diseases, one of the most common and affect, mainly, the spine. Considering the above-mentioned problem, it is important to develop strategies that enable the recognition and prediction of postures performed by workers in their workplaces. To this end, this thesis proposes a pipeline that aims to build a Machine Learning model that may be capable of recognizing, in real time, the upper body postures of the users. The above described framework provides a model capable of recognizing 6 different static postures (back frontal bending, left lateral back bending, right lateral back bending, neck frontal bending, working overhead, and finally the standing posture). In addition to these static postures, the model is also capable of recognizing the transitions between the neutral posture and each of the other five. Several feature selection algorithms have been tested to find the biomechanical features that best distinguish the different postures. The algorithms tested were the Principal Component Analysis, the Forward Sequential Selection, the Minimum Redundancy Maximum Relevance (mRMR) and the ReliefF. Also, several classification algorithms were tested (Support Vector Machines, K-nearest Neighbors, TreeBagger, Discriminant Analysis, Convolutional Neural Network and Feed Forward Neural Network). These algorithms were trained and tested with data from 50 healthy subjects who volunteered to acquire data realizing the above-mentioned postures. The best results were obtained using the SVM classifier with the quadratic kernel and using the characteristics selected by the mRMR algorithm. The classification model showed promising results during cross-validation, more specifically, it presented an MCC value of 0.973. As future work, other types of sensors should be integrated, and other features should be tested to improve posture classifier performance.