Fault detection and isolation for a three tank system based on a bilinear model of the supervised process

Time varying innovation generators combined with generalized likelihood ratio (GLR) tests are designed for detection and isolation of faults in a three tank system. This diagnosis system is based on a bilinear model of the supervised process. It is shown to work properly in a larger working range than a fault detection and isolation (FDI) system based on a linear model. As the faults enter in a bilinear way in the model, achieving exact decoupling of the residuals with respect to some of the faults is not possible. One has to resort to approximation methods such as the approach developed in Patton and Chen (1993). The whole FDI system is designed and tuned on the basis of a simulation of the three tank system. Next it is applied to actual pilot plant data and it is shown to perform well. To be able to detect temporary faults (namely fault appearance and disappearance) with the GLR test, a strategy based on the use of two Kalman filters running in parallel is used