A Particle Filtering Framework for Fault Detection and Identification (FDI) – ICSL

A Particle Filtering Framework for Fault Detection and Identification (FDI)

Our fault diagnosis procedure fuses and utilizes the information present in a feature vector (observations) with the objective of determining the operational condition (state) of a system and the causes for deviations from desired behavioral patterns. From a nonlinear Bayesian state estimation standpoint, this task may be accomplished by the use of a Particle Filter-based module built upon the nonlinear dynamic state model.


Where fb, ft and ht are non-linear mappings, xd(t) is a collection of Boolean states associated with the presence of a particular operational condition in the system (normal operation, fault type #1, #2, etc.), xc(t) is a set of continuous-valued states that describe the evolution of the system given those operational conditions, n(t) is zero-mean i.i.d. uniform white noise and w(t), v(t) are non-Gaussian distributions that characterize the process and feature noise signals, respectively.

One particular advantage of our Particle Filtering approach is the ability to characterize the evolution in time of the above mentioned nonlinear model through modification of the probability masses associated with each particle, as new feature information is received. Furthermore, the output of the fault diagnosis module, defined as the current expectation of each Boolean state, provides a recursively updated estimation of the probability for each fault condition considered in the analysis. In addition, pdf estimates for the system continuous-valued states provide swift transition to failure prognosis algorithms, which is a primary advantage for a Particle Filter-based diagnosis framework.