Page 1083 - The Mechatronics Handbook

frequency domain features of the sensor signal such as mean, variance, skewness, kurtosis, crest factor,
or power in a specified frequency band. Choosing appropriate monitoring indices is crucial. Ideally the
monitoring indices should be: (i) sensitive to the system/process health conditions, (ii) insensitive to the
working conditions, and (iii) cost effective. Once a monitoring index is obtained, the monitoring function
is accomplished by comparing the value obtained during system operation to a previously determined
threshold, or baseline, value. In practice, this comparison process can be quite involved. There are a
number of feature-based monitoring methods including pattern recognition, fuzzy systems, decision
trees, expert systems, and neural networks.
Fault detection and identification (FDI) process in dynamic systems could be achieved by analytical
methods such as detection filters, generalized likelihood ratio (which uses Kalman filter to sense discrep-
ancies in system response), and multiple mode method (which requires dynamic model of the system and
could be an issue due to uncertainty in the dynamic model) (Chow and Willsky, 1984).
As mentioned above, the system failures could be detected and identified by investigating the difference
between various functions of the observed sensor information and the expected values of these functions.
In case of failure, there will be a difference between the observed and the expected behavior of the system,
otherwise they will be in agreement within a defined threshold. The threshold test could be performed
on the instantaneous readings of sensors, or on the moving average of the readings to reduce noise.
In a sensor voting system, the difference of the outputs of several sensors and each component (sensor
or actuator) is included in at least one algebraic relation. When a component fails, the relations including
that component will not hold and the relations that exclude that component will hold. For a voting system
to be fail-safe and detect the presence of a failure, at least two components are required. For a voting
system to be fail-operational and identify the failure, at least three components are required, e.g., three
sensors to measure the same quantity (directly or indirectly). As Chow and Willsky (1984) pointed out, for
the detection and identification of a single failure among m components at least (m - 1) relations are required
(more relations are preferred for better performance in the presence of noise).

39.4 Intelligent Fault Detection Techniques

The fault tolerant control (robust control and decision-making process) should include allowable per-
formance degradation in the failed state, criticality and likelihood of the failure, urgency of response to
failure, tradeoffs between correctness and speed of response, normal range of system uncertainty, dis-
turbance environment, component reliability vs. redundancy, maintenance goals (mean-time-to-failure,
mean-time-to-repair, maintenance-hour/operation-hour, etc.), system architecture, limits of manual inter-
vention, and life-cycle costs (Stengel, 1991).
Fault detection could be achieved by redundancy in sensing (measurement) and actuation, parallel
redundancy (e.g., dual sensors or actuators), analytical redundancy, and artificial intelligence (expert
systems, artificial neural network, or integration of both techniques) combined with redundancy.
Stengel (1991) classified the analytical redundancy into direct and temporal redundancy. Direct redun-
dancy consists of algebraic relationship among instantaneous outputs of sensors and is useful for sensor
failure detection, but not for actuator failure detection. Temporal redundancy includes the relationship
among histories of sensor outputs and actuator inputs (also comparison of the outputs of dissimilar
sensors at different times). Temporal redundancy could be used for both sensor and actuator FDI, e.g.,
a sensor voting system with mixed displacement and velocity sensors could detect failures of both types
of sensors. The computational complexity of temporal redundancy is higher compared to the direct
redundancy case as it requires the dynamics of the system.
An expert system embodies in a computer the knowledge-based component of an expert skill in such
a manner that the system can generate intelligent actions and advice and can, when required, justify to
the user its line of reasoning. In general, an expert system is composed of three parts: an inference engine,
a human-machine interface, and a knowledge base. The inference engine is the knowledge processor and is
modeled after the expert’s reasoning. The engine works with available information on a particular problem,
coupled with the knowledge stored in the knowledge base to draw conclusions or recommendations.

©2002 CRC Press LLC