Documente Academic
Documente Profesional
Documente Cultură
TABLE I
DEGREES OF SHARPNESS AND VAGUENESS FOR MAIN ELEMENTS IN PROCESS AUTOMATION. INTENSITY OF A PROPERTY: +++ LARGE; ++ MEDIUM; + SMALL
input variables, the process classes, the automation functions models based on physical laws result and one can distinguish
and the output variables. Sharp information is labeled to be the following cases of analytical process models (quantitative
“crisp” and vague information to be “fuzzy.” models):
A.1) equations with known parameters;
A. Input Variables A.2) equations with partially unknown parameters;
A.3) equations with unknown parameters.
With regard to the type of input signals precise and impre-
cise values are distinguished. Usual measurement equipment In practice, case A.2 appears more frequently than A.1 and
is designed to deliver signals with precise mean values and A.3. If the structure is known, the unknown parameters can
low standard deviations. Low cost sensors may have imprecise be determined experimentally by parameter estimation and
outputs with biased values and large standard deviations. identification methods, e.g., [9]. Hence, quantitative process
Nondirectly measurable variables like slip, material stress, models result from proper combination of theoretical and
or comfort or quality measures can only be calculated or experimental modeling. The field of process modeling has
estimated based on other measurement variables and analytical reached a fairly mature status and is packaged in attractive
models and are therefore imprecise to some extent (This is also software tools.
known as sensor fusion). Also linguistic values of a human However, if the internal behavior cannot be expressed by
operator may be processed, like the color of fire in a cement physical laws in the form of equations, as for “not well
kiln, a brush fire of a dc motor or the type of noise of an defined” processes like drying, vulcanizing, special chemical
engine. Hence, the degree of precision and imprecision of the reactions or biological processes, some qualitative information
input variables may vary within fairly broad ranges. on the causalities may be expressed as rules:
(1)
B. Process Classes The condition part then contains as inputs the memberships
The description of the static and dynamic behavior in form of facts to the premise and possibly their logical connection
of mathematical models plays a crucial role as well for the by AND, OR and NEGATION, and the conclusion part de-
design of the technical processes as for a systematic design of scribes the logical consequence on the output. Then one can
high performance control systems. As an example continuous distinguish heuristic process models (qualitative models):
processes with continuous input and output signals are con- B.1) rules with known conditions, known conclusions and
sidered. If the internal process behavior can be described by known relations between conditions and conclusions;
basic physical principles, theoretical modeling can be applied B.2) rules with unknown logical connections within the
by stating balance equations, physical-chemical state-equations condition, but known conclusion;
(constitutive relations) and phenomenological laws. Then the B.3) rules with known condition and unknown conclusion.
structure of the mathematical model is obtained either as a Hence, a gradual decrease of knowledge on the process
lumped parameter or distributed parameter system (ordinary can be observed from analytical process models, case A.1,
or partial differential equations). By this way quantitative to heuristic process models, case B.3. This is also shown in
ISERMANN: FUZZY LOGIC APPLICATIONS 223
TABLE II
DEGREES OF VAGUENESS FOR PROCESS AUTOMATION FUNCTIONS WITH A FUZZY LOGIC APPROACH. 11 VAGUE; 1 LESS VAGUE; CRISP; 0 NOT DEFINED
Fig. 6. Signal flow in a fuzzy rule-based system with two levels of rules and max/min operations and singletons as inputs.
and the fuzzy sets of the conclusion. If the rule is inter- rules grow. By special assumptions the computations can be
preted by the Mamdani-implication and the compositional simplified considerably [20], [21] for example by singletons
rule of inference by Zadeh [3] is applied, the following at the input and output or universal sets at the output.
steps can be distinguished: Fig. 6 shows the signal flow for a fuzzy-rule-based system
1) matching of the facts with the rule premises (determina- with two levels of rules and singletons as inputs. In this case
tion of the degrees of fulfillment of the facts to the rule fuzzification and matching reduces to directly determining the
premises fit values); degree of fulfillment of the premises. The evaluation of the
2) if the rule contains logic connections between several rule conclusion then consists of applying the respective -
premises by fuzzy AND or fuzzy OR the evaluation norms and -conorms (e.g., min-max or prod-sum operation)
is performed by t-norms or t-conorms (The result gives for the logic operations in the premises and determination of
then the degree of fulfillment of the facts to the complete the fuzzy sets of the conclusion.
premise); The chaining of rules can be performed with or without
3) the evaluation of the resulting conclusion follows the defuzzification. In Fig. 6 no defuzzification is considered
max-min composition [3]. (The degree of fulfillment of between the two levels of rules.
the premise restricts the fuzzy set of the conclusion.) According to the knowledge base the rules can be repre-
• accumulation of all conclusion fuzzy sets if several rules sented by a network, building up several levels. For fuzzy
contribute to the output (a union operation yields a global control usually only one level of rules is required. However,
fuzzy set of the output). fault diagnosis leads to tree structures with several levels. The
• defuzzification to obtain crisp outputs (various defuzzifi- contributions of all rules to the last level have then to be
cation methods can be used, as, e.g., center of gravity, accumulated, e.g., by a union operation. This usually results in
maximum-height, and mean of maximum, to obtain a combined fuzzy sets. A crisp output as a final result is obtained
crisp numerical output value). by defuzzification, if required. In the next two sections different
The computational effort of a fuzzy-rule-based system in- kinds of using fuzzy logic approaches in control systems are
creases strongly if the number of fuzzy sets and the number of considered for both normal and special operating conditions.
226 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART A: SYSTEMS AND HUMANS, VOL. 28, NO. 2, MARCH 1998
(3)
(4)
V. FUZZY-FEEDBACK CONTROL FOR
NORMAL OPERATING CONDITIONS
With regard to automatic feedback control normal operat-
A similar procedure follows for multiple inputs, leading
ing conditions are understood where automatic operation is
to simple vector multiplications.
possible without manual control.
6) Drastic nonlinear characteristics can mainly be reached
by distributing the output attributes unequally over the
A. Fuzzy Control in the Basic Level (Direct Fuzzy Control) universe of discourse.
Before discussing more general aspects of fuzzy control
applications some practical aspects of fuzzy control are first B. Criteria for the Application of Fuzzy
considered. For the design of a fuzzy controller the signal Control in the Basic Level
flow of a general fuzzy-rule-based system as shown in Fig. 5
To discuss the potentials of fuzzy controllers in the basic
reduces to a rule activation block with one rule level, as shown
level (direct control) one has to consider the following basic
in Fig. 7. The two inputs and are, e.g., the control
control engineering viewpoints.
deviation and its derivative
1) The proper design of the static and dynamic behavior
(2) of the process is the most important prerequisite for
reaching a good control performance.
and are assumed to be crisp values. 2) The engineering design includes the right specification
The membership functions of the inputs (attributes) are and design of the actuator and the measurement equip-
usually symmetric triangles with equal distribution over the ment, such that the process can at least be controlled
universe of discourse or ramps, such that the sum is always manually.
one. The formulation of the attributes of the output variable 3) The solution of these problems presupposes the exis-
shows frequently more variety [22]. Now a special controller tence of models of the process, at least rough static
configuration is considered. The effect of choosing different characteristics. Hence, simple models frequently exist
output attributes like rectangles, triangles, ramps, etc., has been for important processes.
investigated in [21]. For two ramps as inputs and and 4) In most cases linear PI- and PID-controllers are ap-
center of gravity defuzzification, the analysis of the controller’s plied and the parameters are tuned by tuning rules.
transfer behavior yielded the following results. Mathematical models are not required for the selection
1) In selecting the output attributes the position of their of the controller type and the parameter tuning. A
center of gravity is mainly important. rough understanding of the process’ control behavior is
2) The shape, the width of the output attributes and the sufficient in most cases.
inference method (max-min or max-prod.) are of minor 5) With modern microprocessor-based controllers it is no
importance. They influence the interpolation between the problem to generate a static nonlinear behavior like non-
centers. linear characteristics, dead zones, direction dependent
3) The width of rectangles and triangles has no influence, as control actions, etc.
long as the membership functions do not overlap each Fuzzy-rule-based controllers as in Fig. 7 and Fig. 8(a) gen-
other. erate a three-dimensional static mapping of the output in
4) Overlapping triangles or ramps (mountains) generate S- dependence on the two inputs (look-up-table). Depending on
shaped nonlinear characteristics which are, however, not the selection of the membership functions this mapping may
much different from a linear one. become nonlinear and depending on the number of inputs
5) Formulation of the linguistic output attributes as sin- multidimensional. By adding dynamic elements at the input
gletons leads to a drastic reduction of the computational or the output (e.g., derivative or integrating terms) the fuzzy
effort. The activation of an output attribute which controller gets dynamic properties.
ISERMANN: FUZZY LOGIC APPLICATIONS 227
(a) (b)
(c) (d)
(e) (f)
(g) (h)
(i) (j)
Fig. 8. Control system structures with different fuzzy control components (hybrid classical/fuzzy control). F: fuzzification; I: inference; and D: defuzzification.
Nonlinear characteristic surfaces can of course also be The control engineering aspects for the application of fuzzy
generated directly. The advantage of the fuzzy-rule-based controllers are still under investigation. Some advantages and
approach lies in the formulation of the input-output behavior disadvantages can be summarized as follows.
in form of linguistically expressed rules. By this way the 1) Nonlinear processes. For processes which can be lin-
practical experience of human operators can be implemented earized and processes with analytically known nonlinear
in another form than by (simple) equations, which may be behavior, such as nonlinear characteristics or hysteresis,
easier to understand for some operators and nonexperts. many solutions do exist with classical control sys-
228 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART A: SYSTEMS AND HUMANS, VOL. 28, NO. 2, MARCH 1998
tems like PID-, state space controllers with observers, TABLE III
cancellation-, deadbeat-, minimum-variance and predic- CONTROLLER STRUCTURE IN DEPENDENCE
ON THE PROCESS BEHAVIOR AND SIGNALS
tive controllers [23]. In most cases only rough process
model knowledge is required, especially, if self-tuning
or adaptive control methods are applied [9].
For processes where no sufficient control perfor-
mance can be reached with PI- and PID-controllers
and parameter adjustment by tuning rules, fuzzy control
may be an alternative. This is especially valid for
processes with difficultly understandable behavior, if
only qualitative knowledge is available and strongly
nonlinear properties do exist. Some examples are milling
auxiliary controller and a main controller [Fig. 8(c)] [23]. The
plants, glass melting, cement kilns, food processing.
auxiliary controller can, e.g., be a classical P- or PI-controller
2) Knowledge of human operators. If manual control of
and the main controller can be a fuzzy controller which is
a difficult process is possible the operator knowledge
especially designed for the nonlinear behavior of process part
may be formulated in fuzzy rules with linguistic inputs
and outputs (This is in many cases not an easy task).
Another interesting possibility is a cascade control system
The parametrization can be improved especially by
with a fuzzy main controller for two different control variables
trial and error and also additional inputs can be used,
[Fig. 8(d)]. Here fuzzy rules decide which one of both control
e.g., in the sense of feedforward actions. However,
variables is in action with a smooth transition from one to the
the fine parametrization may take a long time, since
other. An example is the velocity and distance control in car
much more parameters have to be adjusted than for
driving [31].
PID-controllers. Another application is the imitation
of human controllers in the loop, as “fuzzy control D. Fuzzy Feedforward Control
involves a representation of what is basically a human If an external disturbance of a process can be measured
solution in the form of fuzzy if-then rules,” [3]. This before it acts on the output variable the control perfor-
kind of application is, e.g., interesting for car driving, mance with respect to this disturbance can be improved by
commanding aircraft, operating a crane or workshop feedforward control. Then the fuzzified measured disturbance
activities. Here the actions of a human can be formulated is just added to the premise of the rules of the fuzzy feed-
in rules and can be adapted by identification methods back controller [Fig. 8(e)]. Thus, proportional acting nonlinear
[21], [24]. Hence, fuzzy controllers can be used to feedforward control can be realized.
investigate human’s behavior.
3) Fast prototype design. A further advantage of the fuzzy E. Variable Control Structures with Fuzzy Weighting
control concept is a fast prototyping by some few rules,
if a classical standard controller is not suitable. However, The selection of classical controllers depends among others
the fine tuning for reaching a good control performance on process properties such as the transfer behavior of the plant,
usually lasts too long. the size of control deviations and the kind of disturbances [23].
Therefore the type of controller (and its parameters) should
A still holding disadvantage of direct fuzzy control is the
be altered, if such properties change. This is the case if for
difficulty of stability analysis. In fact, the proof of stability
example the size of control errors change drastically or if wide
for arbitrary fuzzy controllers is not a simple task due to
process operating ranges have to be governed.
their various structures and nonlinear characteristics. Several
Table III shows some examples for changing (classical)
approaches for stability analysis have been presented in recent
controller structures. Usually only one controller type with
years. [25]–[27] give an overview of existing methods. Most
robust properties is implemented or two controller types are
of them are adopted from nonlinear control theory, e.g.,
switched on and off alternatively. A fuzzy logic approach now
Lyapunov method, harmonic balance, Popov criterion etc. The
offers a gradual change from one controller type to another by
examples show that both time domain and frequency domain
weighting the control deviation by for controller 1 and
approaches are applied. It can be distinguished between meth-
for controller 2 based on the (e.g., identified) process
ods considering the fuzzy controller as nonlinear input-output
behavior [Fig. 8(f)]. Dependent on the process behavior rules
mapping, e.g., [8], and those regarding the internal structure of
can be established to weight the controllers action (and to
the fuzzy controller such as [28]–[30]. The latter approaches
change its parameters). One example is IF disturbances have
can also be used in a constructive manner in order to design
higher frequencies AND control deviation is medium THEN
robust fuzzy controllers.
decrease weight of PI-controller AND increase weight of
The next sections show other forms of applying fuzzy
PD-controller (see also [32]).
concepts for control systems.
C. Fuzzy Cascade Control Systems F. Fuzzy-Tuning of Classical Controller Parameters
The control performance of a single loop system can be In practice the parameters of P-, PI- and PID- con-
improved by expanding it to a cascade control system with an trollers are mostly adjusted by tuning rules based on
ISERMANN: FUZZY LOGIC APPLICATIONS 229
closed loop oscillation experiments with a P-controller try to use already available other sensor signals as inputs
(Ziegler–Nichols rule) or open loop transfer function mea- [see Fig. 8(i)]. If analytical relations are known the control
surement (Chien–Hrones–Reswick rule). In [24] and [33], a variable can be calculated, e.g., a heat stress from measured
parameter tuning procedure with if-then fuzzy rules [Fig. 8(h)] temperature variations. However, if mathematical models are
is shown for PI- and PID-controllers. The controller starts with unknown or too complicated, one can formulate rules. An
small values of the parameters. Measured features of the closed example is the room temperature heating control without
loop transient function such as the overshoot and the ratio rise indoor and outdoor sensors. The basic water temperature is
time/settling time are fuzzified and rules for the changes of the determined by the average temperatures over the year and
controller parameters allow to adjust the controller parameters during a day (yearly and daily temperature profile). This rough
after some few (2–8) step response experiments. One example forecast is corrected by AND connections in the premise
of the 24 tuning rules for a PI-controller is: with measured inputs as: average energy consumption of the
IF overshoot is too small AND settling ratio is alright previous day and present energy consumption [37].
THEN increase gain by 30% The tuning rules hold for
a large variety of process behaviors and also for stochastic I. Fuzzy Quality and Comfort Control
disturbances. The quality of products is mostly based on several proper-
ties. Examples are as follows.
G. Fuzzy Adaptive Control 1) Metal products: geometrical precision, roughness of sur-
Adaptive control systems adapt their control behavior to face, texture, stiffness, hardness.
the changing properties of the controlled processes and their 2) Ceramic products: geometry, color, surface, acoustic
signals. Most adaptive control systems can be divided into sound.
feedforward and feedback adaptive controllers [9]. Fig. 8(g). 3) Textile products: elasticity, strength, texture, color, pat-
shows a scheme for a feedforward adaptive control which tern, homogenity, bulking, skin reception, washing, and
is based on the fact that the changing properties of the iron properties.
plant can be grasped by the measurement of a signal 4) Cement: particle size and its distribution, specific sur-
on the process (e.g., mass flow, speed, load). If rules are face (Blaine number), sieve residues, early strength and
known how the controller parameters of a classical (or fuzzy strength after 28 days of binding.
controller) must be changed in dependence on the measured 5) Tyre: stiffness, traction, durability, noise, rain and aqua-
signal, a fuzzy rule adaptation can be realized. If the process planing sensitivity, temperature sensitivity.
behavior cannot be determined directly by the measurement The various product properties which accumulate to an
of external process signals, feedback adaptive control has to overall quality measure can be influenced during processing by
be used [Fig. 8(h)]. The changing properties of the process changing of reference values of control loops, reaction times,
have to be determined by the measurement of different internal load variables, ingredients, etc. As frequently no analytical
control loop signals. Classical feedback adaptive control is models are known, but only rules, a multivariable quality fuzzy
then based on the identification of a mathematical process controller with quality attributes as inputs and manipulated
model (model identification adaptive control) or a comparison process variables as outputs can be formulated [Fig. 8(b)].
with a reference model (model reference adaptive control) A similar situation can be found for the control of comfort
with the goal to design a controller automatically. A fuzzy where the human perception plays a role. Examples are as
feedback adaptive controller does not use a mathematical follows:
process model but rules for the adaptation of the controller 1) The comfort of the climate in air conditioning where
based on measured closed loop signals. For example the the temperature, humidity and kind of labor of human
controlled and manipulated signals are used to determine beings follow certain rules.
measures for the control performance. After the fuzzification 2) The comfort of the passengers during car driving where
of these measures tuning rules of the controller determine the received frequency spectra of accelerations can be
the adjustment of the parameters of a classical or a fuzzy formulated as rules with the adjustable shock absorbers
controller. as outputs and road conditions and accelerations as
A further possibility results from using a fuzzy model of inputs. In [38] the simulation of a fuzzy controller is
the process. The expert knowledge is here used to model the shown where rules are given for the adjustment of the
process by rules and measured signals to refine the model shock absorber dependent on the relative speed of the
with identification procedures [34]–[36]. The controller design damper and the speed of the chassis.
then is performed automatically. The controller types may be 3) The comfort of the start-up of automobiles with manip-
classified as predictive controller [35], PID-controller [36], ulation of the clutch and gear shifting [39].
feedback linearization, etc. or fuzzy controllers. This results
then in adaptive control with fuzzy models. J. Neuro-Fuzzy Control
An increasing activity can be observed in the field of neuro-
H. Fuzzy Control of Nonmeasurable Quantities fuzzy control systems [40]–[42]. The combination of fuzzy
Sometimes a control variable is not measurable or a sensor logic and neural networks integrates the advantages of both
should be saved because of economic reasons. Then one can approaches in one controller. On the one hand, the controllers
230 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART A: SYSTEMS AND HUMANS, VOL. 28, NO. 2, MARCH 1998
can be designed in the fuzzy sense using heuristic knowledge, additional manipulated variable and controlled variable
on the other hand, they can be trained from learning data by [Fig. 8(j)].
means of algorithms developed for neural networks. Various The discussion in Sections III and IV has shown that the
training algorithms have been presented for both on-line and potentials for fuzzy logic approaches in the feedback control
off-line training. Two general strategies in neuro-fuzzy control domain are mainly
can be distinguished: The most simple proposals use two 1) fuzzy control in the basic level for nonlinear processes
separate components (one fuzzy system, one neural network) with only qualitative knowledge or if imitation of human
in a parallel control structure. However, approaches utilizing controllers is required;
the structural equivalence between certain types of fuzzy 2) addition of fuzzy control to classical control (cascade-
systems and artificial neural networks [43]–[45] seem to be control, feedforward control, variable control structures);
more promising. One single controller can be trained from data 3) fuzzy control in the adaptation level;
(interpretation as neural network) and still remain transparent 4) fuzzy control for nonmeasurable quantities, quality or
due to the internal fuzzy-rule structure (interpretation as fuzzy comfort measures;
systems). 5) fuzzy control for abnormal operating conditions and
large-scale dynamic effects.
Hence, fuzzy control approaches may be especially advan-
VI. FUZZY-CONTROL FOR SPECIAL OPERATING tageous as an addition to classical control systems, forming
CONDITIONS AND LARGE-SCALE DYNAMIC EFFECTS hybrid classical fuzzy control systems. Then different fuzzy
control components can be specified, as shown in Fig. 8.
Many processes are automatically controlled for normal
operating conditions with sufficient control performance. With
regard to feedback control systems normal operating condi- VII. SUPERVISION, FAULT DETECTION AND DIAGNOSIS
tions mean that stable or not too sluggish control is possible
with a fixed controller or an adaptive controller (mostly A. Knowledge-Based Fault Detection
feedforward adaptation of controller parameters). This is for Within automatic control of technical systems supervisory
example the case for the control of flow or pressure, heat functions serve to indicate undesired or unpermitted process
exchangers or steam boilers within a range of about 20% states and to take appropriate actions in order to maintain the
to 100% of the load. Especially for small loads closed loop operation and to avoid damages or accidents. The following
operation with the normal controller is frequently not possible functions can be distinguished:
as the gain of the process and its time constants increase too a) monitoring: measurable variables are checked with re-
much or the valves operate in nearby closed position with gard to tolerances and alarms are generated for the operator.
undefined characteristics. b) automatic protection: in the case of a dangerous process
Other examples of abnormal process behavior are state, the monitoring function initiates automatically an appro-
1) the start-up and shutdown phase of engines, machines, priate counteraction.
industrial processes; c) supervision with fault diagnosis: based on measured
2) mode changes in chemical processes; variables features are calculated, symptoms are generated via
3) spinning of cars outside of normal tire/road slip (unstable change detection, a fault diagnosis is performed and decisions
steering behavior); are made for counteractions.
4) large slip of tires and blocking of wheels after heavy Fig. 9 shows an overall scheme of knowledge based fault
braking; detection and diagnosis. The main tasks can be subdivided in
5) stalling of an aircraft after falling short of speed; fault detection by analytic and heuristic symptom generation
6) control of processes after failures (emergency control). and fault diagnosis [47], [48].
In these cases the process behavior is frequently not well 1) Analytic Symptom Generation: The analytical knowl-
known because some physical laws do not hold any more or edge on the process is used to produce quantitative, analytical
are not known (e.g., change from turbulent to laminar flow in information. Based on measured process variables a data
heat exchangers, rotating of combustion engines with irregular processing has to be performed to generate first characteristic
speed, chattering in metal cutting processes). Hence, there is values by
need to control processes with large uncertainties and large- 1) limit value checking of directly measurable signals.
scale dynamic effects [46]. The usual way for governing the Characteristic values are exceeded signal tolerances.
abnormal process is manual control with experienced operators 2) signal analysis of directly measurable signals by use of
by bringing the process back into normal or safe conditions. signal models like correlation functions, frequency spec-
Now a special fuzzy controller can be designed by making tra, autoregressive moving average models (ARMA).
use of the rules gained by the operator [Fig. 8(j)]. It is typical Characteristic values are, e.g., variances, amplitudes,
for these abnormal operating conditions that the manipulation frequencies or model parameters.
has to be performed differently in comparison to the normal 3) process analysis by using mathematical process mod-
operation, e.g., by using fast and large actions as in the case els together with parameter estimation, state estimation
of spinning cars or the recovery of an aircraft from the post- and parity equation methods. Characteristic values are
stalling phase. The special fuzzy controller may then use an parameters, state-variables or residuals.
ISERMANN: FUZZY LOGIC APPLICATIONS 231
In some cases special features can be extracted from these Three important model-based fault detection methods and
characteristic values, e.g., physical defined process coefficients their residuals are:
or special filtered or transformed residuals. These features 1) parameter estimation: changes of parameter estimates
are then compared with the normal features of the nonfaulty ;
process and methods of change detection and classification are 2) state estimation: changes of state estimates or
applied. The resulting changes (discrepancies) of the described output errors ;
direct measured signals, signal models or process models are 3) parity equations: output error or polynomial error
then analytic symptoms of faults. .
2) Heuristic Symptom Generation: In addition to the symp- For more details, see [47]. The measured or estimated
tom generation with quantifiable information heuristic symp- quantities like signals, parameters, state-variables or residuals
toms can be produced by using qualitative information from are usually stochastic variables with mean value and
human operators. Through human observation and inspection variance
heuristic characteristic values in form of special noise, color,
smell, vibration, wear and tear, etc. are obtained. The process (5)
history in form of performed maintenance, repair, former
faults, and lifetime, load measures constitute a further source as normal values for the nonfaulty process. Analytic symptoms
of heuristic information. Statistical data (e.g., MTBF, fault are then obtained as changes
probabilities) achieved from experience with the same or sim-
ilar processes can be added. By this way heuristic symptoms
(6)
are generated which can be represented as linguistic variables
(e.g., small, medium, large) or as vague numbers (e.g., around with reference to the normal values. Fig. 10(a) shows an ex-
a certain value). ample for a stochastic symptom If a fixed threshold
is used, a compromise can be made between the detection of
small faults and false alarms because of short term exceeding
B. Model-Based Fault Detection Methods [see Fig 10(b)]. Methods of change detection which estimate
Different approaches for fault detection by using mathemat- the change of the mean value and relate it to the standard
ical models were developed in the last 20 years. The task deviation may improve the (binary) decision. Examples are
consists in the detection of faults in the technical process likelihood-ratio test, Bayes-decision, run-sum test, two probe
including actuators and sensors by measuring the available -test. Now a triangle-membership function is specified
input and output variables and . The process is with the center as the mean and the lower width or
considered to operate in open loop. A distinction can be because for a Gaussian distribution 68% or 95% of all
made between static and dynamic, linear and nonlinear process values lie within those intervals [Fig. 10(c)], thus obtaining
models. a combined symptom representation of the mean and the
232 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART A: SYSTEMS AND HUMANS, VOL. 28, NO. 2, MARCH 1998
chaining a rule, the facts are matched with the premise and 5) fault decision by maximal possibility;
the conclusion is drawn based on the logical consequence 6) defuzzification to obtain a crisp size of the most possible
(modus ponens). Therefore with the symptoms as inputs fault.
the possible faults are determined using the heuristic Multiple rule activation from all symptoms to all faults
causalities. In general the symptoms have to be considered (vertical procedure) [53]:
as uncertain facts. Therefore a representation of all observed 1) chaining by
symptoms as membership functions of fuzzy sets in the
interval [0, 1] is feasible, especially in unified form. multiple composition to obtain overall relations
4) Approximate Reasoning with Fuzzy Logic: With the in form of a matrix ;
structure of a fault symptom tree a fuzzy-rule-based system 2) matching of with all current symptoms
with multiple levels of rules can be established, as shown ;
in Fig. 6. The symptoms of the features are now 3) accumulation to obtain the global possibility for each
represented by fuzzy sets with linguistic fault;
4) fault decision by maximal possibility;
meanings like “normal,” “less increased,” “much increased,”
5) defuzzification to obtain a crisp size of the most possible
etc. Then the following analogies to Fig. 6 hold:
fault.
— symptoms In both approaches, a defuzzification can be performed after
— events each level or not.
To reduce the computational effort, simplifying assumptions
— faults may help. The following simplifications are possible:
In contrast to fuzzy control in the basic level fuzzy fault 1) singletons for the symptoms: matching is reduced to
diagnosis differs by direct fuzzification of crisp values;
2) singletons for the events and faults: the conclusions are
1) inputs are mostly no crisp measurements, but detected
singletons with a height of the matching degree between
symptoms represented as fuzzy sets;
facts and rule premise;
2) not only one level of rules does exist, but mostly several
3) universal fuzzy sets for the events and faults: the con-
levels;
clusion is a universal set with height of the matching
3) frequently it is difficult to specify the membership
degree between facts and rule premise. But then the size
functions of the intermediate events because of very
of the event or fault cannot be determined.
vague knowledge.
In the cases of singletons or universal sets for events
The approximate reasoning follows the steps described in
and faults the conclusions are identical with the membership
Section IV: fuzzification or matching, rule activation with eval-
degree of the facts with the logical operation in the premise.
uation by the compositional rule of inference, accumulation
This means the whole procedure reduces to an aggregation of
and defuzzification. The dimensions of the fuzzy-rule-based
the facts with appropriate t-norms and t-conorms, and chaining
system is given by
of the results [47]. Fuzzy fault diagnosis can be expanded to
— number of symptoms time dependent faults and symptoms, including incipient and
— number of levels intermittent faults and multiple faults [53].
— number of rules per level 5) Neuro-Fuzzy Systems for Diagnosis: The knowledge
— number of faults acquisition and tuning for fault diagnosis usually takes
The overall dimension may therefore blow up strongly even much effort and time. Therefore adaptation mechanisms of
for small components or processes. Therefore the software the establishment of rules and membership functions are
implementation is important. Two procedures to perform the required. Considering now the similarity between fuzzy logic
reasoning are known. and neuronal networks adaptive neuro-fuzzy systems can be
Sequential rule activation from level to level developed [54]–[59]. In [54] it is shown that a combination
(horizontal procedure): of the bounded operator for fuzzy-AND and the bounded sum
for fuzzy-OR leads to an adjustable ANDOR-operator ( -
1) matching of symptoms with the premises of rules
operator), which can be approximated by a sigmoidal activity
of level (including their logical operations) to function. Furthermore, by applying Gaussian membership
determine the events ; functions adjustable symptom membership functions can be
2) matching of events with the obtained. By measurement of faults and symptoms the neuro-
fuzzy diagnosis system is adapted by a backpropagation
premises of rules of level to determine the events
method, see also [55].
;
3) matching of events with the faults fuzzy subsets VIII. CONCLUSION
; Fuzzy logic provides a systematic framework to process
4) accumulation of fault fuzzy subsets to obtain the global vague variables and vague knowledge. As the vagueness
possibility for each fault; in automation generally increases with increasing functional
234 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART A: SYSTEMS AND HUMANS, VOL. 28, NO. 2, MARCH 1998
hierarchical level the potentials of fuzzy logic approaches [19] R. Yager, Essentials of Fuzzy Modeling And Control. New York: Wiley,
naturally grow with higher levels. Fuzzy control in the basic 1994.
[20] B. Kosko, Neural Networks and Fuzzy Systems. London, U.K.:
level may be attractive for processes where only qualitative Prentice-Hall, 1992.
knowledge is available and strongly nonlinear properties do [21] B.-M. Pfeiffer, Einsatz von Fuzzy-Logik in lernfähigen digitalen
exist, or if human control has to be imitated. Interesting Regelsystemen. Fortschr.-Ber. VDI-Reihe 8 Nr. 500. Düsseldorf:
VDI-Verlag, 1995.
possibilities for fuzzy control exist in the higher control [22] , Selbsteinstellende klassische Regler mit Fuzzy-Logik,” Automa-
levels, as e.g., cascade control, variable control structures tisierungstechnik, vol. 42, pp. 69–73, 1994.
[23] R. Isermann, Digital Control Systems. New York: Springer-Verlag,
with fuzzy weighting, fuzzy controller parameter tuning, fuzzy 1989, vol. 1–2.
adaptive control, fuzzy control of nonmeasurable quantities [24] B.-M. Pfeiffer and R. Isermann, “Fuzzy logic applications in learning
and fuzzy quality and comfort control. Another application digital control systems,” in Proc. 3rd European Congr. Fuzzy Intelligent
Techniques Soft Computing, Aachen, Germany, 1995, pp. 847–852.
area is the fuzzy control for special (abnormal) conditions [25] G. Bretthauer and H.-P. Opitz, “Stability of fuzzy systems—a sur-
where today mostly only manual control is applicable. Thus vey,” in Proc. European Congr. Fuzzy Intelligent Technologies, Aachen,
hybrid classical/fuzzy control system structures result. Germany, 1994, pp. 283–290.
[26] J. P. Marin and A. Titli, “Comparative analysis of stability methods
The supervision of processes requires the treatment of for fuzzy controllers,” in Proc. European Congr. Fuzzy Intelligent
quantitative and qualitative knowledge. Here fuzzy approaches Technologies, Aachen, Germany, 1994, pp. 1183–1187.
are especially attractive for symptom generation with fuzzy [27] W. J. Jansen et al., “Stability properties of fuzzy controllers,” J. A, vol.
36, no. 3, pp. 27–37, 1995.
thresholds, linguistically described observed symptoms and [28] H. Kiendl and J. J. Rüger, “Stability of fuzzy control systems using
the approximate reasoning with multilevel fuzzy-rule-based facet functions,” Fuzzy Sets Syst., vol. 70, no. 2–3, pp. 275–285, 1995.
systems for fault-symptom tree structures. [29] H. O. Wang, K. Tanaka, and M. Griffin, “An analytical framework
of fuzzy modeling and control of nonlinear systems: Stability and
Further applications of fuzzy logic methods may arise design issues,” in Proc. American Control Conf., Evanston, IL, 1995,
within the qualitative knowledge processing of intelligent pp. 2272–2276.
or multiobjective control systems including higher level con- [30] T. A. Johansen, “Fuzzy model based control: Stability, robustness, and
performance issues,” IEEE Trans. Fuzzy Syst., vol. 2, pp. 221–234, 1994.
trol functions and learning. Especially the combination with [31] S. Germann and R. Isermann, “Nonlinear distance and cruise control for
nonlinear approximation (identification) and classification by passenger cars,” in Proc. IFAC-Workshop Advances Automotive Control,
artificial neural networks may open many new possibilities. Ascona, Switzerland, Mar. 13–17, 1995, pp. 203–208.
[32] R. Stenz and U. Kuhn, “Automation of a batch destillation column using
fuzzy and conventional control,” IEEE Trans. Contr. Syst. Technol., vol.
3, no. 2, pp. 171–176, 1995.
REFERENCES [33] B.-M. Pfeiffer and R. Isermann, “Selftuning of classical controllers with
fuzzy logic,” Math. Comput. Simul., vol. 37, pp. 101–110, 1994.
[1] L. A. Zadeh, “Outline of a new approach to the analysis of complex [34] L. X. Wang, Adaptive Fuzzy Systems and Control. Englewood Cliffs,
systems and decision processes,” IEEE Trans. Syst., Man, Cybern., vol. NJ: Prentice-Hall, 1994.
SMC-3, pp. 28–44, 1973. [35] R. Babuska, J. Sousa, and H. B. Verbruggen, “Model-based design of
[2] , “A rationale for a fuzzy control,” J. Dyn. Syst., Meas., Contr., fuzzy control systems,” in Proc. 3rd European Congr. Fuzzy Intelligent
vol. 94, Series G, pp. 3–4, 1972. Technologies, Aachen, Germany, 1995.
[3] , “The role of fuzzy logic in modeling and control,” Model., [36] M. Fischer, O. Nelles, and D. Füssel, “Tuning of PID-controllers for
Identif. Contr., vol. 15, no. 3, pp. 191–203, 1994. nonlinear processes based on local linear fuzzy models,” in Proc. 4th
[4] E. H. Mamdani and S. Assilian, “An experiment in linguistic synthesis European Congr. Fuzzy Intelligent Technologies, Aachen, Germany,
with a fuzzy logic controller,” Int. J. Man-Mach. Stud., vol. 7, pp. 1–13, 1996.
1975. [37] D. Pfannstiel, “Einsatz adaptiver und fuzzybasierter Regelungsstrategien
[5] L. Holmblad and I. I. Ostergaard, “Control of a cement kiln by fuzzy in der Heizungstechnik,” Automatisierungstechnische Praxis, vol. 37,
logic,” in Fuzzy Information and Decision-Processes, M. M. Gupta and pp. 42–48, 1995.
E. Sanchez, Eds. Amsterdam, The Netherlands: North-Holland, 1982, [38] A. Heinrich, J. Kraft, F. Dörrscheidt, M. Boll, “Fuzzy-Regelung eines
pp. 389–399. semiaktiven Pkw-Fahrwerks,” Automatisierungstechnische Praxis, vol.
[6] S. Yasunobu and S. Miyamoto, “Automatic train operation by predictive 36, pp. 23–30, 1995.
fuzzy control,” in Industry Applications of Fuzzy Control, M. Sugeno, [39] K. Pfeiffer and R. Isermann, “Automatic control of vehicle starting for
Ed. Amsterdam, The Netherlands: North-Holland, 1985, pp. 1–18. emission tests on engine dynamometers,” in Proc. American Control
[7] H. P. Preuß, “Fuzzy-control-heuristische Regelung mit unscharfer
Conf., Seattle, WA, 1995.
Logik,” Automatisierungstechnische Praxis, vol. 34, pp. 176–184, 1992. [40] H. Takagi, “Fusion technology of fuzzy theory and neural net-
[8] D. Driankov, H. Hellendoorn, and M. Reinfrank, An Introduction to
works—survey and future directions,” in Proc. Int. Conf. Fuzzy Logic
Fuzzy Control. Berlin, Germany: Springer-Verlag, 1993.
[9] R. Isermann, K.-H. Lachmann, and D. Matko, Adaptive Control Systems. Neural Networks, Iizuka, Japan, 1990, pp. 13–26.
[41] M. Brown and C. Harris, Neurofuzzy Adaptive Modeling and Control.
Hertfordshire, U.K.: Prentice-Hall, 1992.
[10] G. N. Saridis, Self Organizing Control Of Stochastic Systems. New Hertfordshire, U.K., Prentice-Hall, 1994.
York: Marcel Dekker, 1977. [42] H.-P. Preuß and V. Tresp, “Neuro-fuzzy,” Automatisierungstechnische
[11] G. N. Saridis and K. P. Valavanis, “Analytical design of intelligent Praxis, vol. 36, no. 5, pp. 10–24, 1994.
machines,” Automatica, vol. 24, pp. 123–133, 1988. [43] J. S. R. Jang and C. T. Sun, “Functional equivalance between radial basis
[12] K. J. Åström, “Intelligent control,” in Proc. European Control Conf., function networks and fuzzy inference systems,” IEEE Trans. Neural
Grenoble, France, 1991 Networks, vol. 4, no. 1, pp. 156–158, 1993.
[13] Handbook of Intelligent Control, D. A. White and D. A. Sofge, Eds. [44] K. Hunt, R. Haas, and R. Murray-Smith, “Extending the functional
New York: Van Nostrand Reinhold, 1992. equivalence of radial basis function networks and fuzzy inference
[14] P. Antaklis, Defining intelligent control. IEEE Trans. Contr. Syst. Tech- systems,” IEEE Trans. Neural Networks, vol. 7, no. 3, pp. 776–781,
nol., pp. 4–66, June 1994. 1996.
[15] M. M. Gupta and N. K. Sinha, Intelligent Control Systems. New York: [45] S. Halgamuge, Advanced Methods for Fusion of Fuzzy Systems and
IEEE, 1996. Neural Networks in Intelligent Data Processing, Fortschr.-Ber. VDI
[16] Advances in Intelligent Control, C. J. Harris, Ed., London, U.K.: Taylor Reihe 10 Nr. 401. Düsseldorf, VDI-Verlag, 1996.
and Francis, 1994. [46] J. Perkins, “Trends in process systems engineering,” Model., Identif.
[17] R. Isermann, “Toward intelligent control of mechanical processes,” Control, vol. 15, no. 3, pp. 171–177, 1994.
Contr. Eng. Practice, vol. 1, no. 2, pp. 233–252, 1993. [47] R. Isermann, “Integration of fault detection and diagnosis methods,”
[18] W. Pedrycz, Fuzzy Control And Fuzzy Systems, 2nd ed. New York: in Proc. IFAC Symp. Fault Detection, Supervision, Safety for Processes,
Wiley, 1993. Espoo, Finland, June 13–16, 1994, pp. 597–612.
ISERMANN: FUZZY LOGIC APPLICATIONS 235
[48] , “Model based fault detection and diagnosis methods,” in Proc. [61] H. Kiendl, “Fuzzy-control,” Automatisierungstechnik, vol. 41, pp.
American Control Conf., Seattle, WA, June 21–23, 1995. A1–A8, 1993.
[49] R. Isermann and B. Freyermuth, “Process fault diagnosis based on [62] G. V. S. Raju and J. Zhow, “Adaptive hierarchical fuzzy controller,”
process model knowledge. Part I: Principles for fault diagnosis with IEEE Trans. Syst., Man, Cybern., vol. 24, p. 973, July/Aug. 1993.
parameter estimation. Part II: Case study experiments,” J. Dyn. Syst., [63] H. J. Zimmermann, Fuzzy Set Theory and Its Applications, 2nd ed.
Meas., Contr., vol. 113, p. 627, 1992. Boston: Kluwer, 1991.
[50] R. Isermann, “On the integration of fault detection methods via unified [64] J. van Amerongen, H. R. van Nauta Lemke, and J. C. T. van der
fuzzy symptom representation,” 1st European Congr. Fuzzy Intelligent Veen, “An autopilot for ships designed with fuzzy sets,” in IFAC/IFIP
Technologies, Aachen, Germany, Sept. 7–10, 1993, pp. 400–407. Conf. Digital Comp. Applications. Amsterdam, The Netherlands:
[51] W. S. Lee, D. L. Grosh, F. A. Tillmann, and C.H. Lie, “Fault tree North-Holland, 1977.
analysis, methods and applications—a review,” IEEE Trans. Rel., vol.
R-34, no. 3, pp. 194–202, 1985.
[52] R. Isermann and M. Ulieru, “Integrated fault detection and diagnosis,”
in IEEE/SMC Conf. Systems Engineering Service Humans, Le Touquet,
France, 1993.
[53] M. Ulieru, “Fuzzy logic in diagnostic decision: Possibilistic networks,”
Ph.D. dissertation, Tech. Univ. Darmstadt, Darmstadt, Germany, 1996.
[54] M. Ayoubi, “Fuzzy systems design based on a hybrid neural structure Rolf Isermann (M’92) received Dipl.-Ing degree
and application to the fault diagnosis of technical processes,” Contr. in mechanical engineering and the Dr.-Ing. degree,
Eng. Practice, vol. 4, no. 1, pp. 35–42, 1996. both from the University of Stuttgart, Stuttgart,
[55] R. Isermann and M. Ayoubi, “Fault detection and diagnosis with Germany, in 1962 and 1965, respectively.
neuro-fuzzy-systems,” in Proc. 4th European Congr. Fuzzy Intelligent He became Professor at the University of Stuttgart
Technologies, Aachen, Germany, 1996. in 1972. Since 1977, he has been Head of the
[56] H. Berenji, “Learning and tuning fuzzy logic controllers through rein- Laboratory of Control Engineering and Process Au-
forcements,” IEEE Trans. Neural Networks, vol. 3, no. 5, 1992. tomation at the Institute of Automatic Control, Tech-
[57] R. Kruse, J. Gebhardt, and F. Klawonn, Foundations of Fuzzy Systems. nical University of Darmstadt, Darmstadt, Germany.
New York: Wiley, 1994. His main area of research interest covers process
[58] T. Pfeufer and M. Ayoubi, “Fault diagnosis of electro-mechanical modeling and identification, digital control systems,
actuators using a neuro-fuzzy network,” in Proc. 3rd Workshop Fuzzy- adaptive control systems, fault diagnosis, and mechatronic systems. He has
Neuro-Systeme ’95, Gesellschaft für Informatik, Darmstadt, Germany, also written several books and papers on these topics.
1995. Dr. Isermann was the Chairman of the international program committees of
[59] H. Ushida, T. Yamaguchi, K. Goto and T. Takagi, “Fuzzy-neuro control several IFAC Symposia and of the 10th IFAC-World Congress on Automatic
using associative memories, and its applications,” Contr. Eng. Practice, Control in Munich, Germany, in 1987. He was awarded by the Dr. h.c. degree
vol. 2, no. 1, pp, 129–145, 1994. from L’Université Libre de Bruxelles, Bruxelles, Belgium, in 1989 and from
[60] J. H. Kim, K. C. Kim, and E. Chong, “Fuzzy precompensated PID- the University of Bucharest, Bucharest, Romania, in 1996, and received the
controllers,” IEEE Trans. Contr. Syst. Technol., vol. 2, pp. 406–411, highest scientific award from VDE in 1996. Since 1996, he has been IFAC
1994. Vice President (Executive Board).