A stochastic linear neural network-based differential evolution algorithm for optimizing reluctance actuator

Abstract

Reluctance actuators encounter significant challenges in active gravity compensation due to nonlinear effects including dynamic stiffness, leakage flux, and fringing flux, which cause unstable thrust output during operation. This study focuses on suppressing thrust fluctuations in open-loop reluctance actuator gravity compensation systems through design parameter optimization. We first establish a parametric optimization model based on equivalent magnetic circuit principles to characterize nonlinear thrust dynamics. To address limitations in conventional optimization approaches, we develop a differential evolution algorithm enhanced by a stochastic linear neural network (DE-SLNN). This hybrid method combines conventional linear neural networks (CLNN) with a stochastic dynamic opposite learning mechanism (SDOLM) to strengthen DE’s search capability. Comprehensive validation using CEC 2014 benchmark confirms DE-SLNN’s accelerated convergence speed and superior search performance. Application to reluctance actuator optimization demonstrates DE-SLNN’s ability to maintain full-stroke thrust output with a small deviation of less than 2.69% for stable gravity compensation, as confirmed by thrust distribution analysis. Comparative validation with finite element analysis (FEA) reveals thrust error control within 0.16%, verifying the optimization model’s precision. Experimental results demonstrate the framework’s effectiveness in mitigating nonlinearities and achieving stable thrust output within the actuator’s operational stroke.