Genetic Algorithm-Optimized Feature Selection for sEMG-IMU Fusion Improves Intention Detection in AI-Driven Robotic Walking System

Abstract

The increasing demand for responsive and intuitive assistive walking devices, driven by an aging population, underscores the need for intelligent systems powered by emerging machine learning (ML) technologies. This study introduces a novel feature fusion framework based on the Nondominated Sorting Genetic Algorithm II (NSGA-II) to fuse surface electromyography (sEMG) signals with inertial measurement unit (IMU) data and a high-level control architecture, enabling accurate and robust motion intention detection for robotic assistive walking systems. The proposed feature fusion method consistently outperformed statistical filter-based techniques such as mutual information (MI), minimum redundancy maximum relevance (MRMR), correlation coefficient (CC), and Fisher score (FS). It significantly improved the classification metrics of random forest (RF), K-nearest neighbour (KNN), and support vector machine (SVM) classifiers across varying feature counts. For example, the feature fusion algorithm improved RF’s accuracy by 6.74%, 7.67%, 6.35%, and 3.60% and enhanced precision by 6.77%, 7.67%, 6.36%, and 3.61% relative to FS, CC, MRMR, and MI, respectively. Similarly, the algorithm increased RF’s recall by 6.79%, 7.71%, 6.38%, and 3.62%. The proposed feature fusion and high-level and low-level control frameworks were implemented on SoloWalk for real-time interaction, enabling participants to perform daily walking activities. Real-time validation confirmed system stability across gait patterns and user variations, demonstrating its effectiveness in assistive walking robots.