Dimensionality reduction in the control of quasi-static force production tasks in humans

Motor Control;Model Reduction;Principal Component Analysis;Non-negative Matrix Factorization

Speidel, Nathaniel Christopher ; LaValle ; Steven M. ; López-Ortiz ; Citlali ; Dullerud ; Geir E.

关键词: Motor Control;    Model Reduction;    Principal Component Analysis;    Non-negative Matrix Factorization;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/101099/SPEIDEL-THESIS-2018.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Human muscle activation patterns have proven difficult to characterize due to the large number of degrees of freedom present in the system. As a result, efforts to effectively reduce the number of degrees of freedom used to characterize this system have become an important area of research. The underlying characteristic behind the reduction in dimensionality is the ability to group together individual degrees of freedom (typically muscles) together to create new variables that act as the input to the system. In these experiments, two common grouping methods are explored: principal component analysis (PCA) and non-negative matrix factorization (NMF) subjected to a generalized Akaike information criterion (AIC) to serve as a quality of fit estimator. These are used to group the muscle activity of individuals during quasi-static force production tasks to synthesize reduced-order models that account for 90\% of the muscle activity contributing to the task. Regression techniques are then utilized to obtain mathematical models that describe the system's behavior. Ultimately, we show that the system's response is a multi-valued function and linear combinations of predetermined functions perform poorly in terms of goodness of the fit at capturing the data. However, a Fourier series parameterization through time yields promising results in terms of validity of studying reduced order models and how they may be used to further study robotic systems or form movement characterization and rehabilitation infrastructure.