Toward Robust and Accurate Myoelectric Controller Design Based on Multiobjective Optimization Using Evolutionary Computation

Myoelectric pattern recognition holds significant importance in designing control strategies for a range of applications, including upper limb prostheses and biorobotic hand movement systems. It serves as a crucial aspect in formulating effective control strategies for these applications. This study presents a novel method for creating a robust and accurate electromyogram (EMG)-based controller. The approach involves leveraging a kernelized support vector machine (SVM) classifier to interpret surface electromyography (sEMG) signals and accurately deduce muscle movements. The primary objective in designing the classifier is to minimize false movements specifically during the "rest" position of the controller, thereby optimizing the overall performance of the EMG-based controller (EBC). To achieve this, the training algorithm of the supervised learning system is formulated as a problem of constrained multiobjective optimization. For tuning the hyperparameters of SVM, we employ the nondominated sorting genetic algorithm II (NSGA-II), which is an elitist multiobjective evolutionary algorithm (MOEA). Experimental results are presented using a dataset comprising sEMG signals obtained from 11 subjects at five different positions of the upper limb. Furthermore, the performance of the trained models based on the two-objective metrics, namely classification accuracy and false-negative have been evaluated on two different test sets to examine the generalization capability of the proposed training approach while implementing limb-position invariant EMG classification. The results presented clearly demonstrate that the proposed approach offers increased flexibility to the designer in selecting classifier parameters, enabling them to optimize the robustness and accuracy of the EBC more effectively.