Particle filtering on hybrid dynamical systems for sensor fault detection

Abstract

Sensor faults may be modelled as discrete random variables jointly distributed with process states and measurements, resulting in a hybrid (continuous and discrete) dynamical system. State estimation of such systems is challenging due to the combinatorial explosions following propagation of the discrete states. For this reason, a new algorithm is proposed which leverages the structure of the particle filter to incorporate discrete states representing sensor health. The discrete states characterize either normal or faulty conditions, with unique faults (e.g., stuck, out-of-range, etc.) as individual states. These become additional states to estimate which are subsequently used for fault detection and diagnosis. The method is evaluated using a simulated continuous stirred tank reactor benchmark and diagnoses sensor faults with an average sensitivity > 90% for two fault types and > 70% for three fault types.