Journal of Bionic Engineering (2023) 20:2619–2632 https://doi.org/10.1007/s42235-023-00419-w

An Optimization System for Intent Recognition Based on an Improved KNN Algorithm with Minimal Feature Set for Powered Knee Prosthesis

Yao Zhang1 · Xu Wang1 · Haohua Xiu2 · Lei Ren1,3 · Yang Han4 · Yongxin Ma1 · Wei Chen1 · Guowu Wei5 · Luquan Ren1

Haohua Xiu xiuhh@nbut.edu.cn * Lei Ren lren@jlu.edu.cn

1 Key Laboratory of Bionic Engineering, Ministry of Education, Jilin University, Changchun 130022,

People’s Republic of China

2 Robotics Institute of NBUT, Ningbo University of Technology, Ningbo 315211, People’s Republic of China

3 Department of Mechanical, Aerospace and Civil Engineering, The University of Manchester, Manchester M13 9PL, UK

4 School of Mechanical Science and Aerospace Engineering, Jilin University, Changchun 130022,

People’s Republic of China

5 School of Science, Engineering and Environment, University of Salford, Salford M5 4WT, UK

Abstract: In this article, a new optimization system that uses few features to recognize locomotion with high classification accuracy is proposed. The optimization system consists of three parts. First, the features of the mixed mechanical signal data are extracted from each analysis window of 200 ms after each foot contact event. Then, the Binary version of the hybrid Gray Wolf Optimization and Particle Swarm Optimization (BGWOPSO) algorithm is used to select features. And, the selected features are optimized and assigned different weights by the Biogeography-Based Optimization (BBO) algorithm. Finally, an improved K-Nearest Neighbor (KNN) classifier is employed for intention recognition. This classifier has the advantages of high accuracy, few parameters as well as low memory burden. Based on data from eight patients with transfemoral amputations, the optimization system is evaluated. The numerical results indicate that the proposed model can recognize nine daily locomotion modes (i.e., low-, mid-, and fast-speed level-ground walking, ramp ascent/decent, stair ascent/descent, and sit/

stand) by only seven features, with an accuracy of 96.66% ± 0.68%. As for real-time prediction on a powered knee prosthesis, the shortest prediction time is only 9.8 ms. These promising results reveal the potential of intention recognition based on the proposed system for high-level control of the prosthetic knee.

Keywords: Intent recognition · K-Nearest Neighbor algorithm · Powered knee prosthesis · Locomotion mode classification