A combined statistical detection and qualitative fault isolation scheme for abrupt faults in dynamic systems

Model-based fault detection and isolation (FDI) of physical systems uses models to infer faults from the observed behavior of the system. We develop a robust model-based FDI scheme for abrupt faults in process components of a continuous dynamic system. In this work, a component fault is modeled as a change in the value of a corresponding generic physical parameter in the bond graph representation of the system model. Abrupt faults are changes that occur at time scales much faster than the nominal system dynamics, and cause a transient response in the system. This work builds on a qualitative model-based fault isolation engine, named TRANSCEND, that analyzes the fault in terms of the dynamic behavior of the transient response immediately after the point of fault occurrence. TRANSCEND is implicitly robust against model uncertainties. However, the mapping from numeric measurement data to symbolic feature values determines the extent to which the scheme can accommodate measurement uncertainty. This is the signal-to-symbol transformation problem, and it must be addressed to ensure quantifiable FDI performance.; A signal-to-symbol transformation scheme is developed for TRANSCEND . A key aspect to the solution is the explicit decoupling of fault detection from symbolic analysis. Transient detection is based on a discrete wavelet transform representation of the data, and a statistical decision function in the transform domain. Fault isolation requires extracting the transient characteristics in terms of symbolic magnitude and derivatives around the point of fault occurrence. Therefore, symbol generation must be initiated as close as possible to the fault onset. By estimating the point of fault occurrence through analysis of the decision function, we can (partially) correct for the delay between fault occurrence and detection.; We demonstrate the approach through simulation experiments for a damped spring-mass system and a vehicle suspension system. Since the detection and isolation techniques are decoupled, detection sensitivity is higher than would be possible with symbolic analysis alone. Fault isolation performance is evaluated in terms of accuracy and precision. We emphasize accuracy at the cost of precision for smaller faults. However, the use of robust estimators for the symbolic features ensures that fault isolation precision improves with increasing fault size.