Locally Linear Neuro-Fuzzy Estimate of the Prosthetic Knee Angle and Its Validation in a Robotic Simulator

Abstract

Here, we present a low-power magnetic measurement system based on only two Hall-effect elements and a permanent magnet integrated into a smart knee prosthesis to accurately measure knee flexion–extension. The smart prosthesis was tested in a robotic knee simulator that provides squat movements and different patterns of recorded gait from subjects. The squat movements were used to build linear and locally linear neuro-fuzzy estimators to translate the magnetic measurements into knee flexion angle. The simulated gait patterns then used to validate the models. The locally linear neuro-fuzzy estimator showed a clear benefit against linear regression models, by sequentially splitting the measurement space into subregions and find local models for each subregion. The obtained root mean square errors on test data were lower than 1.3° for the neuro-fuzzy estimates representing less than 3% of range of rotation. The result was compared with the estimates from a previously designed configuration of three 2-D anisotropic magnetoresistive (AMR) sensors tested in the same setup. We showed that by using the neuro-fuzzy model for two Hall-effect elements, similar performance to the AMR-based estimator can be obtained while the power consumption can be reduced more than three folds.