Robust State Estimation of Induction Motor using Desensitized Rank Kalman Filter

The state estimation accuracy of the nonlinear induction motor is restricted by parameter uncertainties. To reduce state estimation error sensitivities to parameter uncertainties, we proposed a novel desensitized rank Kalman filter (DRKF) based on a new sensitivity definition. A novel desensitized cost function for the deterministic sampling methods is designed to obtain an optimal gain matrix. The sensitivity propagation is summarized for deterministic sampling methods. Based on the rank sample rule, the sensitivity propagation method is given, and the DRKF algorithm is derived. The induction motor with two uncertain stator and rotor resistances are simulated to demonstrate that the proposed DRKF has an excellent performance on mitigating the negative effects of parameter uncertainties.

I. Introduction

To reduce the cost of the induction motor and improve the reliability, some key technologies of sensorless control strategy have been developed for induction motor [1]–[3]. One of the essential methods is estimating the motor position and speed to feedback the closed-loop speed control, in which the state and observation system is a highly nonlinear model. One of the popular technologies for motor speed sensorless control is the nonlinear Kalman filtering (KF) class, which includes the extended Kalman filter (EKF), unscented Kalman filter (UKF), cubature Kalman filter (CKF), rank Kalman filter (RKF), ensemble Kalman filter and particle filter [4]–[7]. Kim et al. first used the EKF to obtain the state estimations in the speed sensorless vector control [8], and then many variants of the EKF are successfully applied in several sensorless AC drives [2], [9]-[10]. However, the EKF algorithm approximates the nonlinear function by using the Taylor series and ignoring the higher-order term. The linearization of the motor models will bring unexpected biases and decrease estimation accuracy. Then, some deterministic sample methods, such as UKF, CKF and RKF, are proposed to approximate the probability density function by using the propagation of sigma points [4], [6] and achieve higher accuracy in motor speed estimation [1], [11]-[13].