Sensor Fault Detection using Adaptive Modified Extended Kalman Filter Based on Data Fusion Technique

Abstract

This paper investigates the application of data fusion technique to enhance the sensor fault detection and diagnosis. The extended Kalman filter (EKF) is used to fuse the process measurement sensor data. The usual approach in the classical EKF implementation, however, is based on the constant diagonal matrices for the process and measurement covariance. This inflexible constant covariance set-up which employs the ideal white noise model assumption for describing the process and measurement noises causes the EKF algorithm to diverge or at best converge to a large bound even it the EKF model is perfectly tuned. This paper presents an adaptive modified extended Kalman filter (AMEKF) algorithm to prevent the filter divergence leading to an improved EKF estimation. The performances of the resulting sensor fault detection system are demonstrated an a simulated continuous stirred tank reactor (CSTR) benchmark case study for drift in calibration (bias error) and drift in degradation.