Toward Minimal-Sensing Locomotion Mode Recognition for a Powered Knee-Ankle Prosthesis

Objective: Locomotion mode recognition (LMR) enables seamless and natural transitions between low-level control systems in a powered prosthesis. We present a new optimization framework for LMR that eliminates irrelevant or redundant features and measurement signals while still maintaining performance. Methods: We use multi-objective biogeography-based optimization to find a compromise solution between performance and the minimization of feature set size. Experimental data are collected from four transfemoral users walking with a powered knee-ankle prosthesis. We compare the performance of LMR systems trained with the optimal feature subsets and with the full feature set using a deep neural network classifier across six locomotion modes: standing, flat-ground walking, stair up/down, and ramp up/down. Results: Statistical tests indicate that classifier performance using the optimal feature subsets is statistically equal to that using the full feature set. The LMR trained with an optimal subset results in the 1.98% steady-state and 4.09% transitional error rates, while only using approximately 41% and 53% of the available features and sensors, respectively. Conclusion: Results thus indicate the capability of the proposed framework to achieve simultaneously accurate and low-complex LMR systems for transfemoral individuals with powered prostheses. Significance: This framework would potentially lead to less frequent clinical visits needed for sensor replacement and calibrations, which may save health care costs and the prosthesis user's time and energy.