07-07-2012, 10:36 AM
A Multi-Model Approach for Detection and Isolation of Sensor and Process Faults for a Heat Exchanger
[attachment=27040]
Introduction
Components, sensors and actuators in physical systems
are often subjected to unexpected and unpermitted deviations
fiom acceptabldusualktandard conditions, the so-called faults.
In order to face the possible loss of the system overall
performance, which may present hazards to personnel or lead
to unacceptable economic loss, many research efforts in the
field of process supervision,, fault detection and diagnosis have
been made [1,2,3,4,5,6]. The purpose is not only to determine
if a fault is present in a system (fault detection), but also the
determination of kind and location (fault isolation) or
determination of size and time varying behavior (fault
identification) of the fault. Most schemes consist of two levels,
Figure 1 : Model-based FDI a symptom generation part and a diagnostic part (Fig. 1).
Multi-Model Based Symptom Generation
The task of a fault detection scheme is to observe
possible deviations of the actual plant behavior from the
nominal process behavior (Fig. 1). The nominal behavior can
be derived from a process model. Significant symptoms are
differences between estimations and measurements of states
and outputs, of time constants and static gains etc.
Multi-Model Approach
The main idea of multi-model approach is to diagnose
these deviations by calculating only the difference between
the actual plant measurements J and their estimates from the
model bank.
-r = y - i (5)
These differences are called residuals. The residuals have the
propem of being approximately zero in the fault-free case and
they deviate from zero if a fault occurs on a measured signal or
in the process itself. In order to isolate the source of the fault, a
set of residuals is needed and the overall process is therefore
decomposed into a set of subprocesses. Each of the residuals
should only be sensitive to faults in the used measurements and
in the underlying subprocess and is insensitive to faults in other
measurements or parts of the process. Hence, the pattern of
deflected and undeflected residuals indicates the source of the
fault. The task is to generate a set of structured residuals [6],
where each fault leads to a unique pattern. The pattern can be
depicted in an incidence matrix shown in Table 1.
Water-steam heat exchanger
This is a tubular heat exchanger which heats up the water from
temperature T,, to T3,. The water flow rate highly influences
the static and dynamic behavior of the heat exchanger,
especially for low flow rates as can be seen from Fig.7. With
increasing flow-rate, the volume of water which is heated up in
W1 increases also. If the steam flow rate is constant, the output
temperature decreases because more water is heated up with
the same amount of steam. The static gain with respect to F,,
decreases for the same reason and the time constant with
respect to F, , increases, due to the higher heat capacity of the
larger volume of water.
Conclusions
A multi model approach was presented for the
detection and isolation of process and sensor faults for a
nonlinear system. The approach was successfully implemented
on an industrial scale thermal plant. Eight signals were used to
generate six residuals, which enable detection of fifteen faults.
The Approach is based on decomposition of the process into
several subprocesses and recording of all important
measurements. This may be a drawback of the approach, if too
much additional sensors are required.
[attachment=27040]
Introduction
Components, sensors and actuators in physical systems
are often subjected to unexpected and unpermitted deviations
fiom acceptabldusualktandard conditions, the so-called faults.
In order to face the possible loss of the system overall
performance, which may present hazards to personnel or lead
to unacceptable economic loss, many research efforts in the
field of process supervision,, fault detection and diagnosis have
been made [1,2,3,4,5,6]. The purpose is not only to determine
if a fault is present in a system (fault detection), but also the
determination of kind and location (fault isolation) or
determination of size and time varying behavior (fault
identification) of the fault. Most schemes consist of two levels,
Figure 1 : Model-based FDI a symptom generation part and a diagnostic part (Fig. 1).
Multi-Model Based Symptom Generation
The task of a fault detection scheme is to observe
possible deviations of the actual plant behavior from the
nominal process behavior (Fig. 1). The nominal behavior can
be derived from a process model. Significant symptoms are
differences between estimations and measurements of states
and outputs, of time constants and static gains etc.
Multi-Model Approach
The main idea of multi-model approach is to diagnose
these deviations by calculating only the difference between
the actual plant measurements J and their estimates from the
model bank.
-r = y - i (5)
These differences are called residuals. The residuals have the
propem of being approximately zero in the fault-free case and
they deviate from zero if a fault occurs on a measured signal or
in the process itself. In order to isolate the source of the fault, a
set of residuals is needed and the overall process is therefore
decomposed into a set of subprocesses. Each of the residuals
should only be sensitive to faults in the used measurements and
in the underlying subprocess and is insensitive to faults in other
measurements or parts of the process. Hence, the pattern of
deflected and undeflected residuals indicates the source of the
fault. The task is to generate a set of structured residuals [6],
where each fault leads to a unique pattern. The pattern can be
depicted in an incidence matrix shown in Table 1.
Water-steam heat exchanger
This is a tubular heat exchanger which heats up the water from
temperature T,, to T3,. The water flow rate highly influences
the static and dynamic behavior of the heat exchanger,
especially for low flow rates as can be seen from Fig.7. With
increasing flow-rate, the volume of water which is heated up in
W1 increases also. If the steam flow rate is constant, the output
temperature decreases because more water is heated up with
the same amount of steam. The static gain with respect to F,,
decreases for the same reason and the time constant with
respect to F, , increases, due to the higher heat capacity of the
larger volume of water.
Conclusions
A multi model approach was presented for the
detection and isolation of process and sensor faults for a
nonlinear system. The approach was successfully implemented
on an industrial scale thermal plant. Eight signals were used to
generate six residuals, which enable detection of fifteen faults.
The Approach is based on decomposition of the process into
several subprocesses and recording of all important
measurements. This may be a drawback of the approach, if too
much additional sensors are required.