Generating an adaptive and robust walking pattern for prosthetic ankle-foot utilizing a nonlinear autoregressive network with exogenous inputs / Hamza Al Kouzbary

Hamza, Al Kouzbary (2021) Generating an adaptive and robust walking pattern for prosthetic ankle-foot utilizing a nonlinear autoregressive network with exogenous inputs / Hamza Al Kouzbary. Masters thesis, Universiti Malaya.

	Image (JPEG) (The Candidate’s Agreement) Restricted to Repository staff only Download (679Kb)
Preview	PDF (Thesis M.A) Download (1275Kb) | Preview

Abstract

Many challenges are associated with the development of powered lower limb prostheses, ranging from their mechanical design to their control system. Many studies on the use of control algorithms in the field of rehabilitation robotics have attempted to mimic the behavior of an intact lower limb with different walking speeds over diverse terrains and used different control structures and logic to achieve this overarching goal. Recently, most of these studies tend to use a hierarchical control structure with three control levels. This three-level control structure has at least one element of discrete transition properties that requires many sensors to improve classification accuracy. However, these sensors also lead to higher computational load and costs. In this study, a developed artificial neural network capable of generating dynamic control signals of the missing foot, user-independent and free-mode method using minimum sensory feedback signals was proposed to eliminate the need to switch among different controllers. A database was constructed using four OPAL wearable devices (Mobility Lab, APDM Inc., USA) for seven able-bodied subjects. The gait of each subject at three ambulation speeds during ground-level walking was recorded to train a non-linear autoregressive network with an exogenous input recurrent neural network (NARX RNN) for estimating foot orientation (angular position) in the sagittal plane using shank angular velocity as external input. The trained NARX RNN estimated the foot orientation of all subjects at different walking speeds over a flat terrain with an average root mean square error of 2.1°±1.7°. The minimum correlation between the estimated and measured values was 86% which indicated the high similarity between the estimated and measured foot trajectories. Moreover, results of the t-test show that the error is normally distributed with a high certainty level (0.88 minimum p-value). In addition, the extreme value distribution of the measured and estimated data of all subjects were quite identical which indicates probabilistic consistency of the model. NARX RNN capability to generate the dynamic control signals for different walking cadences will reduce the risk of amputees falling or stumbling when a wrong classification occurs using the conventional three-level controllers.

Actions (For repository staff only : Login required)