Skip to Main Content
 

Global Search Box

 
 
 
 

Files

File List


Khademi_Dissertation_2632638.pdf (15.98 MB)

ETD Abstract Container

Abstract Header

Design and Optimization of Locomotion Mode Recognition for Lower-Limb Amputees with Prostheses

Khademi, Gholamreza
http://rave.ohiolink.edu/etdc/view?acc_num=csu1568747409603973

Abstract Details

2019, Doctor of Philosophy in Engineering, Cleveland State University, Washkewicz College of Engineering.
Lower-limb prostheses feature a high-level control system, referred to as locomotion mode recognition (LMR), which enables seamless human-prosthesis-environment interactions. This dissertation has five aims to show the application of feature selection (FS), optimization, and sensor fusion for the development of optimal predictive LMR systems. Aim 1 is to develop a new hybrid evolutionary algorithm (EA) with enhanced exploration and exploitation abilities for optimizing the LMR design parameters. Our hybrid EA features three new components, including feature sharing among candidate solutions, local search via gradient descent, and random mutation and re-initialization. Statistical tests demonstrate the competitive performance of our method in comparison to 12 other optimization algorithms. Aim 2 is to investigate the application of an EA combined with filter and wrapper FS methods for building accurate LMR for transfemoral amputees while, at the same time, eliminating unneeded features. Experimental gait data collected from the residual limb is used to train various linear and nonlinear classifiers with optimally selected time- and frequency-domain features. Aim 3 is to find an optimal Pareto front, a set of equally preferable non-dominated feature subsets, using multi-objective optimization (MOO) to trade off between complexity and accuracy in LMR. We propose a new gradient-based multi-objective embedded FS method by incorporating an elastic net in multilayer perceptron (MLP) neural network training. Our proposed method shows competitive performance compared to four variants of EA-based MOOs on the basis of normalized hypervolume and relative coverage. Aim 4 is to develop a general sensor / feature optimization framework for LMR that effectively eliminates not only irrelevant or redundant features but also unneeded measurement signals while still maintaining performance. We apply the proposed framework for locomotion mode prediction of several transfemoral individuals with a powered knee-ankle prosthesis. We categorize misclassification based on their impact on the user's walking stability and comfort. A deep neural network trained with an optimal subset found by the optimization framework results in 1.98% steady-state and 4.09% transitional error rates, while only using approximately 41% and 53% of the available features and sensors, respectively. Aims 2-4 would potentially lead to less frequent clinical visits needed for sensor replacement and calibration, which may save health care costs and the prosthesis user's time and energy. Aim 5 is to develop an ensembling LMR system with environmental awareness. We incorporate the terrain conditions (e.g., level walking, stair ascent, and stair descent) in the LMR design to provide valuable subject-independent prior information about the upcoming locomotion modes. We construct an efficient deep convolutional neural network through transfer learning from RGB images captured during various walking tasks. Predicted terrain is combined with mechanical data to design an ensembling algorithm for constructing robust, reliable, and non-delayed LMR. Our sensor fusion approach significantly decreases the confusion between locomotion modes, which make LMR systems viable for real-world conditions.
Dan Simon (Advisor)
Hanz Richter (Committee Member)
Antonie van den Bogert (Committee Member)
John P. Holcomb (Committee Member)
Yiying Fan (Committee Member)
291 p.
Electrical Engineering
Optimization; Locomotion mode recognition; Powered Prosthesis; Multi-objective optimization; Convolutional neural networks; Inertial measurement units; Impedance control; Lower-limb amputees

Recommended Citations

Citations

  • Khademi, G. (2019). Design and Optimization of Locomotion Mode Recognition for Lower-Limb Amputees with Prostheses [Doctoral dissertation, Cleveland State University]. OhioLINK Electronic Theses and Dissertations Center. http://rave.ohiolink.edu/etdc/view?acc_num=csu1568747409603973

    APA Style (7th edition)

  • Khademi, Gholamreza. Design and Optimization of Locomotion Mode Recognition for Lower-Limb Amputees with Prostheses. 2019. Cleveland State University, Doctoral dissertation. OhioLINK Electronic Theses and Dissertations Center, http://rave.ohiolink.edu/etdc/view?acc_num=csu1568747409603973.

    MLA Style (8th edition)

  • Khademi, Gholamreza. "Design and Optimization of Locomotion Mode Recognition for Lower-Limb Amputees with Prostheses." Doctoral dissertation, Cleveland State University, 2019. http://rave.ohiolink.edu/etdc/view?acc_num=csu1568747409603973

    Chicago Manual of Style (17th edition)

csu1568747409603973
338
© 2019, all rights reserved.
This open access ETD is published by Cleveland State University and OhioLINK.