Model predictive control (MPC) has been widely applied in industry, especia
lly in the refining industry. As all feedback controllers require correct s
ensor measurements, unreliable sensors can cause the MPC controller to move
the process in an erroneous manner. Data validation of sensor measurements
is a prerequisite in applying advanced control, particularly multivariable
control which depends on many sensors. However, little research work is av
ailable on how feedback controllers like MPC complicate the task of sensor
validation and process fault diagnosis. In theory, a controller can transfe
r the effect of a sensor fault in a controlled variable to the manipulated
variables. In this paper, principal component analysis (PCA) is applied to
detect, identify and reconstruct faulty sensors in a simulated FCC unit. A
base PCA model is generated by perturbing the process throughout the operat
ing region. Performance of MPC with and without data validation is compared
. The same base PCA model is applied to detect and identify dynamic process
faults. We demonstrate that process faults can be detected and diagnosed a
t an early stage. (C) 2001 Elsevier Science Ltd. All rights reserved.