A KCMAC-PD controller with reduced memory and optimized mapping for the torque control of electric load simulator

Abstract

An electric load simulator requires high demand on control precision and dynamic performance due to its inherent non-linearity and external interference of surplus torque. The cerebellar model articulation controller (CMAC) is a simple, fast and promising neural network with good performance. However, there still exist some problems in the CMAC network, such as large memory, over-learning and complex mapping. This paper introduces the kernel method in CMAC to form KCMAC, by using a third-order B-spline as a kernel function, so that the mapping of CMAC is transferred from the feature space to the kernel space. This method may effectively reduce the storage space as well as the computational complexity. A compound controller with KCMAC and PD (proportional–derivative) is designed with improvement on learning speed for the torque control of an electric load simulator. Compared with the conventional CMAC-PD control strategy, the KCMAC-PD has improved the control precision by 40.4%, 40.8%, 14.1% and 30.5% at a loading frequency of 0.5 Hz, 1 Hz, 1.5 Hz and 2 Hz in the experiments, respectively. The dynamic simulation and experimental results of KCMAC-PD show that this control strategy may ensure loading precision and avoid over-learning of CMAC. They also demonstrate that KCMAC has ability to smooth control output and restrain external disturbances.