Académique Documents
Professionnel Documents
Culture Documents
Abstract
In this paper an overview is given of the various ways fuzzy logic can be used to improve industrial control. The
application of fuzzy logic in control is illustrated by two case studies. The first example shows how fuzzy logic,
incorporated in the hardware of an industrial controller, helps to finetune a PID controller, without the operator having
any a priori knowledge of the system to be controlled. The second example is from process industry. Here, fuzzy logic
supervisory control is implemented in software and enhances the operation of a sintering oven through a subtle
combination of priority management and deviation-controlled timing. Finally the key areas of research in fuzzy logic
control are discussed: First the paper discusses how fuzzy logic control can be combined with other methods. By properly
separating the a priori model knowledge of the process under control, a hierarchy of nonlinear control methods is
established. After a short discussion of how to optimize an intuitive fuzzy rule base, we show that it is possible to derive
conditions for asymptotic stability and robustness of fuzzy logic controllers, using classical nonlinear analysis.
human-like reasoning with automata. This feature that do not allow for real-time evaluation of the
makes it possible to combine very different types of system. In addition not all parameters can be meas-
information and use these to control real-life pro- ured with the desired accuracy.
cesses. The many different strategies to achieve this (2) A second possible approach is to use the
control further complicate a complete review of FL process identification model. In this approach a dy-
control. namic multi-variable model for control is adopted.
Three main reasons can be given for the present This model is identified by fitting it to the real
popularity and therefore for the many different process. Using statistics, the model is modified until
applications of FL control in industry. In the first its response closely approximates that of the real
place FL can be easily combined with existing process. A disadvantage in utilizing this approach
methods. In addition process control often deals is that it is often not allowed to accumulate the data
with intrinsic uncertainties and/or ambiguities, due necessary for making the identification, or that the
to change in parameters (starting materials, fuel time necessary to make enough statistics is not
mixture, external environment) and/or difficult and available.
indirect measurements. Classically these uncertain- (3) In the third place one can exploit fuzzy logic
ties are to be handled by human operators. A final to model the control behavior of an expert operator
advantage of applying FL is its suitability for rapid (or even of more than one operator). This method
prototyping, yielding a shorter time-to-market. can be used independent of the availability of the
In Section 2 we will first discuss the various process for experiments. Of course it is necessary
options for designing an industrial controller and that the operators have gained enough expertise as
then focus on fuzzy control, highlighting those fea- to control the process and that they are able and
tures that are characteristic for fuzzy control. After willing to express their expertise to the designers of
reviewing when and how to employ FL, two case a FL controller. Like in the previous methodology,
studies are described. In Section 3 the hardware it is not necessary to have an in-depth understand-
architecture of the ES100 industrial controller of ing of the interrelations of all parameters.
O M R O N is described. Next, in Section 4, the con-
trol of a sintering oven by means of FL control 2.2. Fuzzy control
software is explained. In Section 5 we show how FL
control is embedded in other control methods and Although fuzzy controllers are based on the
finally review in Section 6 the stability and robust- modeling of knowledge using linguistic or fuzzy
ness of FL controllers. variables, this does not imply that for measuring
process parameters low accuracy and cheap sensors
could be used. In the experience of the author fuzzy
2. Industrial controller design logic does not reduce the need for accurate
measurements and sensors. The advantage of fuzzy
2.1. Basic control design options logic lies in its ability to use qualitative knowledge
(control rules) to quantify controller actions. The
In the design stage of an industrial controller application of fuzzy logic in designing a controller
basically three different methodologies can be has distinct advantages compared to other methods:
followed: (a) Many input- and output variables can be
(1) The most accurate approach is to use the handled simultaneously.
physical model of the system. This looks as the (b) All knowledge rules in the fuzzy expert
most elegant solution. However only the most system apply simultaneously; inferencing can easily
simple processes can be controlled in this way. The be mapped on a multi-processor system. Even
reasons are obvious: In realistic industrial control conflicting rules can be accommodated by the
problems we deal with many parameters, interac- formalism.
tions that are not always transparent or have (c) An important data reduction is brought
strongly nonlinear relationships, and timescales about by reducing the infinite number of values
A.J. van der Wal / Fuzzy Sets and Systems 74 (1995) 33 41 35
Table 1
Subgroups for fuzzy logic control
1 2 3 4
Burning zone Preheater Composition of Fuel mix
starting material
DEVIATION 1 ~
CALCULATION INTERVAL
MAX
PRIORITYMANAGEMENT]
M1N
I
PROCESS CONSTRAINTS [DEVII " ~
&
STATE INDICES
Fig. 6. Timing calculation.
Fig. 5. Priority management in the fuzzy control of a sintering
oven.
The priority manager will determine a weight
factor that is closer to one, the more the higher
- reduce the energy consumption, priority goals are approximated (Fig. 5). Within
- reduce the costs of replacing fire-proof cladding, each rule group the deviations between actual and
- reduce the CO and NOx exhaust. target values are calculated and used in two differ-
Because of the high number of quantities and the ent ways: In the first place the deviations are fuzzi-
complexity of the sintering process, it is clear that fled and used as input to the fuzzy control rule base.
the simple fuzzy control strategies discussed so In the second place the deviations are translated
far will not suffice. Instead, one divides the control- into a normalized value DEVI ~ [ - 1, 1] that is
ler in a number of different groups: In the case of used to define the update frequency of the fuzzy
the sintering oven, these groups are listed in inference result to the output block.
Table 1. In Fig. 6 the effect of DEVI on the timing is
In each subgroup a fuzzy priority manager deter- shown: the more the absolute value of the control
mines the extent to which controller actions will be value deviates from the setpoint, the more frequent
executed in order to realize the different (sub)goals. the controller takes action. In combination with the
The numbers used in the table refer to the relative priority management the controller signals are gen-
priorities within each group: (1) has higher priority erated in a straightforward though subtle way. Fi-
than (2) and so on. nally the control block takes care of the desired
38 A.J, van der Wal / Fuzzy Sets and Systems 74 (1995) 33--41
action, taking into account the current status and In a recent publication Zhao et al. [8] describe
the constraints of the process. a robust hydraulic servo system that has very fast
The break-even for the return on investment of response and high positioning accuracy under dif-
the fuzzy logic controller was already reached after ferent load conditions. The system under control is
only one year. Apart from the direct advantages nonlinear because the natural frequency and damp-
mentioned above, the improvement of the control ing vary with load, stroke length and ending
system forced the staff to review all measuring position of the piston. The proposed fuzzy state
methodologies and optimizing all subsystems. controller uses fuzzy sets to represent the dynamic
Therefore special attention had to be paid to the characteristics of the piston movement. From these
oven setup. Generally this yielded an increased measured data optimum gain constants for the
awareness and knowledge with all staff involved in proportional error, the velocity and acceleration
the project. Another consequence of the incorpora- can be derived, assuming a third-order linear sys-
tion of fuzzy logic control into the oven has been tem model. Center of gravity defuzzification then
observed in practice. Because the installation is effectively acts as a linear interpolation scheme on
more than 90% of the time productive, the risk exists the measured data. The resulting robust fuzzy state
that operators may lose their experience and that controller is a very elegant example of how fuzzy
new operators cannot reach the skills needed pre- logic can be used to control a nonlinear system,
viously. Countermeasures to prevent this develop- using a linear model.
ment has been taken by constructing on-line simula- In the ES100 controller (Section 3), classical PID
tion programs, through which the operator can train control is combined with a fuzzy expert system,
control by mimicking the fuzzy logic controller. which makes the resultant controller nonlinear,
more robust and having better characteristics com-
pared to a linear PID controller.
5. Combining fuzzy control with other methods Neural networks can be exploited in optimizing
crude, fuzzy-rule-based controllers, as e.g. derived
In Fig. 1 we already showed a classical PID from directly modeling operator experience. In
controller that was applied in parallel to a fuzzy Fig. 7 a typical result of this technique is shown.
controller. The question that we will try to answer The picture is produced with the AutoMF tool
in this section is how to combine different [2] and shows the result of tuning a neural net-
(non)linear controllers in a meaningful way. Apart work to a nonlinear function through learning 200
from the basic control design options as mentioned data patterns. The network models fuzzy know-
in Section 2.1, the designer of an industrial control- ledge via Gaussian Membership Functions (MF).
ler is always facing the problem how to combine the The mean and variance of each M F are encoded as
different controlling strategies in an efficient and synaptic weights. First the position and width of
cost effective way. Especially when facing nonlinear the MFs for each of the input and output variables
control, the designer sees himself confronted with is determined using a fuzzy c-means clustering
a large number of controlling principles, each hav- algorithm [ 1]. Next finetuning is achieved by stan-
ing a number of advantages and disadvantages. dard backpropagation of the errors. The combina-
However a general theoretical framework is lack- tion of learning capability of neural networks and
ing, so that a choice mostly is made in a heuristic the approximate linguistic modeling proves to be
way based on previous experience. very powerful.
In the field of nonlinear control very little has The underlying strategy in making use of various
been published on how the designer of a controller (non)linear methods that can be used for modeling
can take maximum benefit of the different novel real world systems is generally heuristic. In order to
control methodologies (fuzzy logic, neural net- make this process more structured, we propose the
works and genetic algorithms) at hand, their rela- following hierarchical approach:
tion and on how they should be combined with In a first step we use the accurate physical model
classical ( mostly linear) control methods. of the system under control. In this way it would be
A.J. van der Wal / Fuzzy Sets and Systems 74 (1995) 33 41 39
2 2
1 1
U 0 U 0
-1 -1
-2 -2
7a 7b
2 2
1.5 1.5
e2 1 e 2
1
0,5 0.5
0 0
70 7d
Fig. 7. Results of AutoMF nonlinear function estimation of the simulated control surface f ( x , y ) = sin(x) + cos(y) on
x,y ~ [0, 6] x [0, 6]. The upper two figures show the results after clustering (a) and after finetuning (b). The lower two figures show the
squared deviation from the exact result after clustering (c) and after finetuning (d). Finetuning was achieved by training with 200
patterns.
possible to include all essential parameters, signifi- ing on a particular property of the system, without
cant variables and their time derivatives into the changing anything else. Thus fuzzy logic enables
model. Also symmetry (both space and time) and the designer to apply the principle of separation of
elementary conservation laws are then guaranteed concerns: each controller output can be optimized
to be modeled. Linearization should only be done in different areas in state space independently. This
when appropriate as far as robustness is concerned, is in contrast with e.g. a neural network (NN)
and care should be taken not to eliminate essential controller. Here, a change in weights in one of
nonlinearities. In this way we avoid having to teach the neurons basically affects all controller out-
the system by repeated exposure to (input, output) puts. Also if a neural network is trained with
pairs what the underlying physics is. (input, output) pairs, all neurons are updated:
The second step is to incorporate practical ex- No localization exists. Therefore a pure N N ap-
perience-based knowledge. Fuzzy logic is the meth- proach should only be considered to design con-
odology of choice to model operator knowledge or trollers when structural understanding is absent,
supervisory control strategies. Also one can use and sufficient training data can be generated. As
fuzzy logic control in a very 'localized' way, focus- a third step we may employ N N to form so-called
40 A.,L van der Wal / Fuzzy Sets and Systems 74 (1995) 33 41
neuro-fuzzy (NF) systems. In NF systems the learn- control problems such as helicopter flight control,
ing capability of the NN is used to finetune the without the existence of a complete stability theory.
MFs of an estimated FL knowledge base, as in The often-heard complaint from control engin-
AutoMF. eers that fuzzy control (and generally all nonlinear
Finally, if there are not sufficient training data to control) does not provide in a straightforward way
finetune our neurofuzzy controller, we have a final criteria for stability is however true. In the very
tool: the genetic algorithm (GA). A typical GA inspiring article 'Why is fuzzy control robust?', Pok
requires the existence (and knowledge) of an evalu- and Xu [4] lay the foundation for a more general
ation function that rates the 'chromosomes' (e.g. stability theory for fuzzy control. By considering
vectors consisting of neural weights). Optimization the system trajectories in phase space and making
is achieved by a survival-of-the-fittest strategy for a number of assumptions on the system to be con-
the chromosomes. So if no training is possible, it trolled, they derive conditions for asymptotic stab-
may be possible to generate an appropriate (neural) ility and robustness.
fuzzy controller if one can validate the different Their line of thought is very elementary and
combinations of neural weights. therefore attractive. Starting with a two-input fuzzy
In going from the 'exact' model, via a fuzzy, controller consisting of only 4 control rules:
linguistic, approximate model, to the learning capa-
bilities of NN and GA, we are increasingly dealing I F ( E = P ) and (R = P) then U = P ,
with methods that are highly nonlinear and relying I F ( E = P ) and (R -- N) then U = Z ,
on emperical data. NN and GA rely strongly on
statistical methods and learning is inherently slow. I F ( E = N ) and (R -- P) then U = Z ,
On the other hand, 'exact' physical models may be I F ( E = N ) and (R = N) then U = N .
impractical, as it is not always possible to evaluate
the algorithm in real time. Here P, N and Z denote linguistic values 'positive',
Of all these methods fuzzy logic clearly is very 'negative' and 'zero', respectively. The fuzzy vari-
attractive because of its ability to focus on problem ables E and R represent the weighted error and
areas and thus to separate different concerns. In weighted rate of change of the error and both are
addition a fuzzy logic controller has recognizable represented by two linear membership functions
structure and outperforms analytical methods in (P,N) each on a bounded universe of discourse.
execution time and memory requirements. The controller output variable U takes three lin-
guistic values (N, Z, P) and is represented by three
triangular MFs.
6. Stability and robustness of fuzzy controllers Using this very elementary fuzzy controller to
stabilize a second-order linear process with low
The initial scepticism with which fuzzy logic was damping of 0.1 and a unity process gain, the
initially met when Lotfi Zadeh first introduced his authors show by simulation that the controller is
theory in 1965, is only matched by the reluctance of capable of stabilizing the system even when the
the control engineering community in Europe and damping factor is varied from 0 to 1.5 and the
the US to accept fuzzy-logic-based control as a valu- process gain varies between 1 and 3. System re-
able novel methodology for developing nonlinear sponse is virtually identical under these variations,
control. The early work of Mamdani [3] (introduc- indicating its robustness. In addition it is shown
ing fuzzy logic control) has given rise to a multitude that the same controller is also adaptable to oper-
of practical applications in industrial control in Ja- ate satisfactorily in an open-loop unstable process.
pan and China. Of these the well-known voice- Overall setting time equals the rise time without
commanded helicopter and car experiments by oscillation and is therefore substantially shorter
Sugeno [6] are outstanding examples of fuzzy con- than in the case of PID control.
trol. At the same time it is a beautiful illustration of The subsequent stability analysis of the fuzzy
a concept that works, even in solving very hard controller is straightforward using standard
A.J. van der Wal / Fuzzy Sets and Systems 74 (1995) 33-41 41
Lyapunov theory, but very laborious, because model for the computer experiments. This will in
many different regions in phase space have to be practice seldom be the case. Bearing in mind how-
considered. The analysis shows that the fuzzy con- ever that under strict conditions often a reasonable
troller effectively operates as a variable structure analytical approximation of the system under con-
controller with a boundary layer [5] of varying trol can be made, it can be concluded that a stabil-
thickness. The robustness and steady-state accu- ity analysis of a fuzzy controller can be made. In
racy of the controller can be expressed in terms of multi-variable systems however the work will
the thickness of the boundary layer. In addition it is quickly become tedious or even impossible.
possible to formulate criteria for asymptotic stabil- These developments in stability analysis show
ity of the controlled system. that although fuzzy logic controllers deal with lin-
guistic variables and rule-based control algorithms,
they can be treated within the general framework of
7. Conclusions nonlinear stability analysis.