FUZZY DIAGNOSIS MODULE BASED ON INTERVAL

FUZZY LOGIC:

OIL ANALYSIS APPLICATION

Antonio Sala

Systems Engineering and Control Dept., Univ. Polit

ecnica de Valencia

Cno. Vera s/n, 46022 VALENCIA, SPAIN

Bernardo Tormos, Vicente Maci

an, Emilio Royo

CMT Motores T

ermicos, Univ. Polit

ecnica de Valencia

Cno. Vera s/n, 46022 VALENCIA, SPAIN

Keywords:

fuzzy expert systems, fuzzy diagnosis, uncertain reasoning, interval fuzzy logic.

Abstract:

This paper presents the basic characteristics of a prototype fuzzy expert system for condition monitoring

applications, in particular, oil analysis in Diesel engines. The system allows for reasoning under absent or

imprecise measurements, providing with an interval-valued diagnostic of the suspected severity of a particular

fault. A set of so-called metarules complements the basic fault dictionary for ﬁne tuning, allowing extra

functionality.

1 INTRODUCTION

In diagnosis of industrial processes, there is a signif-

icant practical interest in developing technologies for

a more effective handling of the information available

to ease the procedures of inspection and maintenance

(I/M) by means of greater automation.

Computer-aided diagnosis is one of the earliest

ﬁelds of applications of artiﬁcial intelligence tools

(Russell and Norvig, 2003). Logic and statistical in-

ference have been tried in previous applications.

Indeed, the full diagnostic problem under uncer-

tain data would need to be considered in a proba-

bilistic framework. Taking into account the “gradu-

alness” of the symptoms and possible diagnostics of

varying degree of severity (captured by fuzzy logic),

the most complete approach would be setting up a

continous Bayesian network. This paradigm arose in

the last decade as a probabilistic alternative to rea-

soning, superior to truth-maintenance approaches in

some cases (see (Russell and Norvig, 2003) and ref-

erences therein). However, inference on this para-

digm is intractable in a general case (NP-hard). If the

amount of uncertainty is low (if a signiﬁcant subset

of the possible measurements is always obtained and

the “determinism” of the underlying system is accept-

able), then fuzzy logic-based approaches to reason-

ing may be a viable solution in practice. This is the

case of some industrial diagnosis problems, such as

oil analysis, to which the system in development is

targeted.

Logic uncertainty can be accommodated by pos-

sibility theory (Cayrac et al., 1996), or by interval-

valued fuzzy logic (Entemann, 2000). The second ap-

proach is the one followed in this work. Other works

discussing condition monitoring (diagnostic and su-

pervision tasks) using fuzzy logic are, for instance

(Carrasco and et. al., 2004; Chang and Chang, 2003).

Condition monitoring can also be dealt with with

model-based approaches (Isermann and Ball

e, 1997),

if enough quantitative descriptions of the system are

available.

This paper presents the structure of a fuzzy infer-

ence module that incorporates some innovations eas-

ing the setting up of rules and improving the quality

of the ﬁnal diagnostic conclusions. In particular, the

use of interval fuzzy logic, the methodology to deal

with exceptions and the possibility of expressing dif-

ferent alternatives for the same diagnostic and, if they

do not aggree, ﬁring a fuzzy contradiction warning.

An application of the system is presently being

tried on an oil analysis task whose main requirements

appear in (Maci

an et al., 1999).

This paper presents, in two sections the structure of

the fuzzy condition monitoring module being devel-

oped and the key concepts of the oil analysis applica-

tion in which the possibilities of the system are being

tested.

Sala A., Tormos B., Macián V. and Royo E. (2005).

FUZZY DIAGNOSIS MODULE BASED ON INTERVAL FUZZY LOGIC: OIL ANALYSIS APPLICATION.

In Proceedings of the Second International Conference on Informatics in Control, Automation and Robotics, pages 85-90

DOI: 10.5220/0001161200850090

 SciTePress

2 THE FUZZY CONDITION

MONITORING MODULE

The presented fuzzy condition monitoring module is

structured in the following main submodules:

• measurement preprocessing

• fuzzy rule base inference

• postprocessing of the conclusions

2.1 Measurement preprocessing

Raw data from sensors may need some sort of pre-

processing prior to rule evaluation. Indeed, if non-

linearity inversion, statistical calculations, dynamic

processing, etc. is carried out beforehand, then the

subsequent rules will be simpler. This preprocessing

is, however, application-speciﬁc in most cases (see

Section 3).

Incomplete or absent measurements: In the case

of incomplete information, the measurements can be

given in interval form, and fuzzy reasoning will be

carried out via generalisation of ordinary rules to

interval-valued logic values, as described in Section

2.3, giving rise to an interval output of estimated

severities.

2.2 Fuzzy inference module

The fuzzy inference module has been also built with

different submodules. It has a “variable deﬁnition

submodule”, a “rulebase deﬁnition submodule” and

an “inference engine”.

Variable deﬁnition. The name, operating range and

applicable ”concepts” (fuzzy sets) on the variables are

deﬁned via a suitable syntax. An example appears

below on the variable ”CU” (copper concentration):

CU NORMAL 5 1 CU HIGH 5 20

CU VERYHIGH 50 150

the ﬁrst number deﬁning the support of the fuzzy set,

the second one deﬁning the core. The last line de-

ﬁnes, for instance, that the concentration of copper

is not ”very high” if it is below 50 ppm, and then,

gradually starts to be considered ”very high” up to

150 ppm where it is considered 100% abnormal, in-

dicating that rules related to this concept would ﬁre

a ”severe” fault. In the intermediate ranges, the rules

would conclude a fault with an ”intermediate” sever-

ity.

Rulebase deﬁnition. The rulebase is deﬁned by

means of a set of rules in the form:

Disorder Symptom-List END

They conform the core of the rulebase, and the ele-

ments in the symptom list are assumed to be linked

by an ”AND” connective. For examples, see Section

3. Inference is carried out by evaluating the minimum

of the severity of the symptoms in the symptom list of

a particular disorder, and assigning that value to the

severity of the associated disorder (see later). If some

”OR” connectives were to be used, it can be done by

means of the metarules to be deﬁned.

Symptom relevance modiﬁer. Each symptom may

be affected by a coefﬁcient indicating that its presence

conﬁrms the fault, but its absence (in the presence of

the rest) indicates a milder severity.

Metarules. The basic rules can be used to detect

particular situations of interest, with an estimated in-

terval of severity as a result. These situations are,

many times, the ultimate faults to be detected.

However, there are occasions where they must be

combined. This combination may have a logical in-

terpretation in terms of AND, OR, NOT; in this case,

the so-called metarules are introduced to handle the

situation. One possible structure is:

DISORDER IF LOGIC-EXPRESSION

where LOGIC-EXPRESSION is any user-deﬁned

combination of symptoms or previously inferred

atomic disorders, with conjunction, disjunction and

negation operators. This rules will be denoted as MIF

rules.

Note that the basic rules could have been embed-

ded into this syntax. However, the proposed one al-

lows for the incorporation of relevance modiﬁers that

might be more cumbersome in the middle of a logic

expression.

Another type of metarule is the one in the form:

DISCARD Disorder IF LOGIC-EXPRESSION

used, for instance, to discard a ”general” fault if a

more speciﬁc situation sharing the same symptoms

(plus some other ones) is encountered, or to set up

“exceptions” to rules. Again, in Section 3, some ex-

amples appear. Other types of metarules (UNION and

ALTERNATIVE) will be later described.

2.3 Inference Methodology

This section discusses the inference mechanism used

to evaluate the presented rules and metarules. The

core of the inference system works with fuzzy interval

uncertain propositions. In these propositions, their

truth value is an interval [ν, π], where 0 ≤ ν ≤ 1

and 0 ≤ π ≤ 1. It describes partial knowledge: the

minimum value it can attain given the present infor-

mation is called necessity ν and the maximum value

will be denoted as possibility π. If they do coincide,

ICINCO 2005 - INTELLIGENT CONTROL SYSTEMS AND OPTIMIZATION

then the proposition can be considered an ordinary

fuzzy proposition. The interpretation of possibility

and necessity is related but not equivalent to other ap-

proaches such as (Cayrac et al., 1996).

Connectives. Connectives allow combination of

atomic propositions into more complex ones by con-

junction, disjunction and negation. Assuming a par-

ticular connective for a fuzzy proposition is given

(such as T-norms for AND, T-conorms for OR, etc.

(Weber, 1983) in a non-intervalic fuzzy logic) in the

fuzzy-uncertain framework, the resulting interval will

be evaluated with interval arithmetic, i.e., given a

fuzzy connective C : [0, 1] × [0, 1] → [0, 1], Y =

C(X

, X

), its interval version is

[ν

, π

] =

[ min

∈ [ν

, π

]

∈ [ν

, π

]

C(µ

, µ

), max

∈ [ν

, π

]

∈ [ν

, π

]

C(µ

, µ

)]

(1)

Regarding negation, the truth degree of ¬p

is de-

ﬁned as the interval [ν = 1 − π

, π = 1 − ν

For instance, using the minimum and maximum as

conjunction and disjunction operators, the intervalisa-

tion of AND and OR operations is:

[ν

, π

]∧[ν

, π

] = [min(ν

, ν

), min(π

, π

)] (2)

[ν

, π

] ∨[ν

, π

] = [max(ν

, ν

), max(π

, π

)](3)

Let us consider now how inference is carried out in

the basic rules and in the metarules.

Basic Rules: The AND intervalic operation (2) is

used (or its trivial generalisation to more intervals).

Furthermore, if a particular symptom has an “irrele-

vance factor” ρ its membership value µ is transformed

to ρ ∗ (1− µ)+µ before carrying out the interval con-

junction.

In fact, the implemented approach considers the

so-called inference error (Sala and Albertos, 2001)

so that given a logic value µ, the inference error is

e = 1 − µ (extended to interval arithmetic). Then,

given a list of q symptoms in a rule, the overall infer-

ence error is:

E =

i=1

)

(4)

So with p = 2 the methodology could be denoted as

Euclidean inference. With p → ∞, the result is the

Those expression apply when the propositions refer

to independent, non-interactive variables. As in ordinary

interval arithmetic, the above expression (1) in complex

propositions yields conservative (excessively uncertain) re-

sults with correlation and multi-incidence in the arguments.

same as the interval AND (using minimum) described

above. By ﬁxing the value of p the user may specify a

different behaviour (the lower p is, the more severity

is substracted due to partially non-ﬁred symptoms).

The formula (4) can be generalised to intervals, ob-

taining the lowest inference error by using the norm

of the minimum inference error of each symptom, and

the highest one with the maxima of the inference error

intervals. Membership values are recovered from the

resulting inference error ﬁgures by means of a nega-

tion formula.

Metarules: The logic operations will use the above

intervalar expressions when evaluating any LOGIC-

EXPRESSION in MIF metarules.

In the DISCARD metarules, the interval-arithmetic

substraction will be used, i.e.:

DISCARD [ν

, π

] IF [ν

, π

]

will give as a result the interval [ν

′

, π

′

= max(0, ν

− π

) π

′

= max(0, π

− ν

)

Membership value transformations. In some cases,

one would like to introduce a set of rules detecting

conditions for an intermediate fault and different con-

ditions for a severe one (say, F1):

IF COND1 THEN F1 INTERMEDIATE

IF COND2 THEN F1 SEVERE

(5)

To deal with this case, the tool under discussion al-

lows linear transformations of membership via the so-

called UNION metarule:

UNION FAULTNAME

IDENT1 l11 l12 IDENT2 l21 l22 ...

The coefﬁcients l

and l

deﬁne a linear transfor-

mation µ = l

∗ (1 − µ) + l

∗ µ to be carried out

on the membership of identiﬁer ”IDENT-i”. After-

wards, an interval-logic OR is carried out. Obviously,

to combine a particular condition with no member-

ship transformation, the setting l

= 0 and l

= 1

must be used

. For instance, the above case (5) would

be encoded by:

UNION F1

COND1 0 0.5 COND2 0.5 1

The linear mapping actually implemented is not exactly

the one above. It is:

′

∗ (1 − µ) + l

∗ µ µ ≥ 0.02

0 µ < 0.02

(6)

In that way, it can be speciﬁed that an intermediate severity

fault must be suspected if any nonzero activation of a partic-

ular condition holds, but no ﬁring will occur if none of the

conditions are active above a signiﬁcant threshold (0.02).

FUZZY DIAGNOSIS MODULE BASED ON INTERVAL FUZZY LOGIC: OIL ANALYSIS APPLICATION

Alternatives. A closely related situation arises

when several alternatives for diagnosing the same

fault exist. If all measurements were available, all

of them should produce the same result so specifying

only one of them in the rulebase will do. However, to

improve results accounting for missing or imprecise

measurements, several of these alternative rules may

be intentionally speciﬁed. In that case, to allow com-

bining different alternatives into one diagnostic, the

intersection of the intervals produced by inference on

each of them will be the produced conclusion of the

inference.

This is implemented in the current tool by an ALT

metarule, with a syntax similar to the union metarule:

ALT FAULTNAME

IDENTIFIER1 l11 l12 IDENTIFIER2 l21 l22 ...

allowing also a linear membership transformation (6)

before the interval intersection is calculated.

For instance, let us assume that, after the mem-

bership transformations, if any, alternative A yields

[ν

, π

] and alternative B yields [ν

, π

] as estimated

severity intervals. If π

> ν

, then the intersection is

not empty and the following estimated severity inter-

val is produced:

[max(ν

, ν

), min(π

, π

)]

Otherwise, the system outputs the interval [π

, ν

]

ﬂagged with a contradiction warning, as the intersec-

tion is empty. If ν

− π

is small, then the contradic-

tion level is small and the above interval can be ac-

cepted as an orientative result. If it is a large number,

it means that different alternatives give totally differ-

ent results so an error in the rulebase deﬁnition or a

fault in one of the measurement devices providing the

data must be suspected.

The results of the inference is a list of truth values

of the disorders. Those truth values are to be inter-

preted as the “severities” (from incipient to severe) of

the associated disorders.

Post-processing. The output of the expert sys-

tem (interval of estimated severity) is translated onto

a summarised statement. Each value of severity

is mapped to a linguistic tag “negligible, incipient,

medium, severe”, deﬁning an interval of severities for

each tag partitioning the full [0,1] range. If both ex-

treme severities of the conclusions have the same tag,

then a conclusion in the form:

Fault FAULTNAME is TAG (MINI-

MUM,MAXIMUM severity)

is extracted. Otherwise, the produced sentence is:

Fault FAULTNAME severity might range from

TAG(MIN severity) to TAG(MAX severity)

3 OIL ANALYSIS APPLICATION

Oil analysis is a key technology in predictive mainte-

nance of industrial Diesel engines. Indeed, by deter-

mining the amount of wear particles, the composition

of them, and other chemicals in the oil, a sensible set

of rules can be cast to allow a reasonably accurate pre-

diction of the oil condition and/or some likely engine

malfunction.

Expert systems based on binary logic have been de-

veloped for the application (Maci

an et al., 2000), but

the use of a fuzzy logic inference engine is considered

advantageous and it is being evaluated.

An application of the above ideas is under develop-

ment at this moment. Let us discuss some issues on

its development.

Preprocessing. When acquiring information from

a particular engine, some observations have the same

meaning for all engines to be diagnosed. However,

other ones need the use of historical data to gener-

ate ”normalised” deviations taking into account sta-

tistical information for a particular engine brand or

model, or a particular unit with special characteris-

tics. In this way, the rulebase conception can be more

general (applied without modiﬁcation to a larger num-

ber of cases) if the data are suitably scaled and dis-

placed prior to inference or, equivalently, fuzzy sets

are modiﬁed according to the particular engine being

diagnosed.

In some measurements, the procedure involves nor-

malising the deviation from the mean expressing it in

variance units, and generating an adimensional quan-

tity. The statistical data are calculated from a database

of previous analysis, classiﬁed in brand, model (and

also from historical records from the same engine).

Other variables are transformed to ”engineering”

units, having a more suitable interpretation than the

raw readings (for instance, oil viscosities are ex-

pressed as a percentage of a reference value from

fresh oil characteristics, instead of the centiStoke

measurement).

Also, in order to consider real engine behaviour,

oil consumption and fresh oil additions are consid-

ered leading to obtain a compensated wear element

concentration more representative of engine status.

Knowledge base. At this moment, the team is in

process of acquiring and reﬁning a knowledge base

with diagnosis rules.

The basics of the knowledge to be incorporated on

the expert system lie on the following facts.

System is focused on automotive engines diagno-

sis (trucks, buses and general and road construction

equipment), and so, the different parameters to mea-

sure are (Maci

an et al., 1999):

ICINCO 2005 - INTELLIGENT CONTROL SYSTEMS AND OPTIMIZATION

• Oil properties: viscosity, Total Base Number

(TBN) and detergency.

• Oil contaminants: Insoluble compounds, fuel dilu-

tion, soot, ingested dust (silicon), water and glycol.

• Metallic elements: iron, copper, lead, chrome, alu-

minum, tin, nickel, sodium and boron.

Other measurements could be performed upon the oil

sample, but with these basic parameters a good diag-

nosis can be achieved. Systems developed for other

types of engines could choose other parameters tak-

ing into account the particularities of these types of

engines.

Let us consider, as an example, the kind of knowl-

edge involved on the dust contamination detection.

Silica and silicates are present at high concentra-

tions in natural soils and dusts. It is for this reason that

silicon is used as the most important indicator of dust

entry into an engine. There have been several studies

done on the causes of premature wear in components

and results vary from study to study but one thing is

clear: external contamination of lube oil by silicon is

a major cause of accelerated wear. Particles of air-

borne dust vary in size, shape and abrasive properties

and in an engine the ingress of atmospheric dust takes

place primarily through the air intake. Those parti-

cles, not retained by ﬁlters, and similar in size to the

oil ﬁlm clearance in main lubricated parts of the en-

gine do the maximum damage. Once the dust particle

has entered an oil ﬁlm it forms a direct link between

the two surfaces, nullifying the effect of the oil ﬁlm,

thus, the immediate consequence is a ”scratching” of

the surface as the particle is dragged and rolled across

the surfaces. The second and potentially more serious

problem is that once the dust particle is introduced

in between the two surfaces, it changes the loading of

the surface from an even distribution to a load concen-

trated on the particle with a huge increase in pressure

at this point. The increase in pressure causes a de-

ﬂection of the surface, which will eventually result in

metal fatigue and the surface breaking up. As soon as

a dust entry problem occurs there is an increase in the

silicon concentration into the oil and an acceleration

of the wear pattern. As long as the oil samples are

being taken at regular intervals in the correct manner,

the dust entry will be detected at a very early stage. If

an effective corrective action is taken, the life of the

component will be signiﬁcantly increased, reducing

maintenance costs.

Diagnosis of silicon contamination (dust inges-

tion):

• If normal wear patterns combine with high silicon

readings in oil analysis, there are three possibilities:

a silicone sealant, grease or additive is in use mix-

ing with engine oil, an accidental contamination of

the sample has occurred or dust ingestion is in the

ﬁrst stage and no others wear patterns are present

yet (too lucky situation). Action recommend to the

maintenance technicians must be to check if an ad-

ditive, grease or sealant has been used recently on

the engine and make sure that the correct sampling

technique was used. An inspection of the air ad-

mission system on the engine will be necessary if

previous action results negative.

• Increased engine top-end wear (iron, chromium, or

aluminium concentration rises up). This increased

engine top-end wear is caused by airborne dust that

has been drawn into the combustion chamber being

forced down between the ring, piston and cylinder.

Dust origin is caused by a defective air cleaner or

a damaged induction system. Actions to be taken

by maintenance technicians are inspect the air ﬁlter

element thoroughly, and check its seals and support

frame for damage and distortion and check too the

pleats for damage. If there is any doubt about a

ﬁlter element, it should always be changed. If the

leak was found, it is necessary to repair the leak and

determine the condition of the engine by checking

compression or blowby.

• Increased engine bottom-end wear (lead, tin or cop-

per concentration rises up). This situation indi-

cates that dirt is basically getting into the lube oil

directly and not past the pistons and rings. The

likely sources are: leaking seals, defective breather,

damaged seal on oil ﬁller cap or dipstick, or dirty

storage containers and/or top-up containers. Rec-

ommended action to be taken by technicians must

be that any dust that is in the oil will be pumped

through the oil ﬁlter before entering the bearings.

Therefore, the ﬁrst step is to examine the oil ﬁlter

looking for dust contamination or bearing mater-

ial. If excessive dust is found, thoroughly check all

seals and breathers, etc. Check the oil storage con-

tainers and top-up containers for ﬁnding the source

of contamination.

In the syntax of implemented prototype tool, the

rules are:

CONTS1 SI NOT NORMAL END

WEAR_1 IF

FE NOT NORMAL or CR NOT NORMAL

or AL NOT NORMAL

WEAR_2 IF

PB NOT NORMAL or SN NOT NORMAL

or CU NOT NORMAL

CONTS2 IF CONTS1 and WEAR_1

CONTS3 IF CONTS1 and WEAR_2

SILICON_CONTAMINATION IF CONTS1

DUST_INGESTION IF CONTS2 or CONTS3

As another example, water problems can be divided

into two different sources: an external water contam-

ination or refrigerant leakage, in each case a different

FUZZY DIAGNOSIS MODULE BASED ON INTERVAL FUZZY LOGIC: OIL ANALYSIS APPLICATION

behaviour is presented. Additionally, water can evap-

orated and not to be present in oil. For this case, other

ﬁngerprints, that remain in oil when water evaporated

must be found, such as: sodium (NA), boron (BO), its

ratio (NABO) or glycol (GLIC) (its absence would ﬁre

rule CONTW3 at most 70%).

Finally, to take into account a speciﬁc situation

such as a refrigerant leakage with greats amounts of

copper from tube wear caused by water surface corro-

sion, an speciﬁc rule is deﬁned too (CONTW4). In the

syntax of the implemented tool, the rules are:

CONTW1 WATER NOT NORMAL END

CONTW2 GLIC NOT NORMAL END

CONTW4 CU VERYHIGH END

CONTW3

NA NOT NORMAL

BO NOT NORMAL

NABO ABNORMAL

GLYC NOT NORMAL 0.7

END

WATERFAULT IF

CONTW1 or CONTW2 or CONTW3 or CONTW4

DISCARD WEAR IF CONTW4

So, based on the ideas from the above rules, a full

rulebase is being built at this moment.

4 CONCLUSIONS

This paper presents a prototype fuzzy expert sys-

tem for oil diagnosis. The ﬂexibility of this sys-

tem is greater than those of binary rules, allowing

for gradation of diagnosis. Also, reﬁnements such

as interval-valued memberships, membership trans-

formations, exceptions (discarding), unions and alter-

natives are included. In this way, the diagnostic ca-

pabilities and the readability of the rule base improve

substantially.

The oil analysis application in consideration pro-

vides a quite complete set of measurements from

which expert rules can be asserted with a reasonable

reliability. However, for suitable diagnosis on a par-

ticular engine, a pre-processing module is essential:

this module incorporates records of similar engines

(same brand, model, history of the one being moni-

tored, fresh oil characteristics, analytical calculations,

etc.) so that the fuzzy set deﬁnitions are adapted for

each case.

The full system is, at this moment, in development

and prototype testing stage (comparing with human

experts’ conclusions and those from preexisting ad

hoc software based on binary logic), but its prelimi-

nary results are promising.

ACKNOWLEDGEMENTS

The research work presented in this paper is ﬁ-

nanced by project FEDER/MEC DPI-2004-07332-

C02-02 (Spain).

REFERENCES

Carrasco, E. and et. al., J. R. (2004). Diagnosis of acidiﬁca-

tion states in an anaerobic wastewater treatment plant

using a fuzzy-based expert system. Control Engineer-

ing Practice, 12(1):59–64.

Cayrac, D., Dubois, D., and Prade, H. (1996). Handling

uncertainty with possibility theory and fuzzy sets in a

satellite fault diagnosis application. IEEE Trans. on

Fuzzy Systems, 4(3):251–269.

Chang, S.-Y. and Chang, C.-T. (2003). A fuzzy-logic based

fault diagnosis strategy for process control loops.

Chemical Engineering Science, 58(15):3395–3411.

Entemann, C. (2000). A fuzzy logic with interval truth val-

ues. Fuzzy Sets and Systems, 113:161–183.

Isermann, R. and Ball

e, P. (1997). Trends in the applica-

tion of model-based fault detection and diagnosis of

technical processes. Control Engineering Practice,

5(5):709–719.

Maci

an, V., Lerma, M., and Tormos, B. (1999). Oil analysis

evaluation for an engines fault diagnosis system. SAE

Papers, 1999-01-1515.

Maci

an, V., Tormos, B., and Lerma, M. (2000). Knowledge

based systems for predictive maintenance of diesel en-

gines. In Proc. Euromaintenance Conf., volume 1,

pages 49–54. Swedish Maintenance Society-ENFMS.

Russell, S. and Norvig, P. (2003). Artiﬁcial Intelligence: a

modern approach. Prentice-Hall, 2nd edition.

Sala, A. and Albertos, P. (2001). Inference error minimisa-

tion: Fuzzy modelling of ambiguous functions. Fuzzy

Sets and Systems, 121(1):95–111.

Weber, S. (1983). A general concept of fuzzy connectives,

negations and implications based on T-norms and T-

conorms. Fuzzy Sets and Systems, 11:115–134.

ICINCO 2005 - INTELLIGENT CONTROL SYSTEMS AND OPTIMIZATION