11-10-2012, 01:27 PM
Extended Multiple Model Adaptive Estimation for the Detection of
Sensor and Actuator Faults
Extended Multiple.pdf (Size: 2 MB / Downloads: 67)
ABSTRACT
A combination of the multiple model adaptive estimation
method (MMAE) with an extended Kalman filter (EKF) is
considered. The ability of the MMAE method to detect faults
based on a predefined hypothesis and the parameter-estimating
ability of an EKF results in a very efficient fault detection
approach. This so-called extended multiple model adaptive
estimation method (EMMAE) has been investigated on a
nonlinear model of an aircraft. The results show that it is a
very capable method to detect faults of various types.
INTRODUCTION
T he multiple model adaptive estimation (MMAE)
method [1] is based on a bank of parallel Kalman filters
(KF), each tuned to describe a particular fault status of the
system. The output of each KF is then weighted by its
corresponding probability based on the measurement
history.
The MMAE method is a good choice for the detection of
actuator as well as sensor faults, as long as the expected
faults can be hypothesized by a reasonable number of
Kalman filters. However, the number of addressable faults is
rather restricted, and this method reaches its limits as soon
as the actual occurring fault does not closely match the
predefined fault hypothesis. This may occur when an
actuator is stuck at an unknown position (here called "lockin-
place" fault) that affects the dynamics of the system. We
know from KF theory that we have to consider all
systematic errors, however, since lock-in-place faults cannot
be predicted, they may have detrimental effects on the filter
performance. Due to the biased residual, the KF provides a
wrong estimation of the state variables, which causes severe
problems with the probability calculation.
SIMULATION RESULTS
The fault detection with the EMMAE method was tested
on a nonlinear aircraft model of tenth order. Simulations
were performed in Matlab/Simulink on an open-loop control
architecture, where an input sinusoidal signal excited the
process. For test purposes two (redundant) ailerons were
"equipped" with an EKF. Fig. 4 depicts the first simulation
in which two ailerons are excited by a sine signal. After 4.5
s the second aileron gets stuck at an offset position. Fig. 4.b
clearly shows that the roll rate is still estimated correctly
despite the fault. Fig. 4.c shows the estimation of the aileron
bias. Up to the aileron fault the estimation follows the actual
aileron movement.
CONCLUSION
The MMAE method has been combined with a parameterestimating
EKF in order to extend the class of detectable
faults. Every fault is described by two parameters, one
having binary characteristic, estimated by the MMAE part of
the EMMAE algorithm and the other one being a (timevarying)
real value estimated by the EKF part of the
EMMAE algorithm.
Simulations show the ability of the EMMAE method to
detect faults that can be described by these two parameters.
The proposed EMMAE method will be further tested in a
closed-loop control architecture, and robustness properties
of the FDI algorithm will also be investigated.