Multimodal Statement Networks for Diagnostic Knowledge

Modeling and Integration

Wojciech Cholewa, Marcin Amarowicz, Pawel Chrzanowski and Tomasz Rogala

Institute of Fundamentals of Machinery Design , Silesian University of Technology,

Konarskiego 18A, Gliwice, 44-100, Poland

Keywords:

Multimodal Statement Networks, Diagnostic Reasoning, Knowledge Modeling, Graphical Models, System

REx, Consistency Checking.

Abstract:

This paper addresses selected aspects of diagnostic knowledge management. Not only does it introduce the

process of modeling with the use of multimodal statement networks, but it also presents methods of diagnostic

model development, nature of its individual layers and elaborates on the construction process of diagnostic

knowledge base along with its veriﬁcation. The main purpose of the presented method is to aid in design

of diagnostic systems for complex objects requiring inference process that takes places under condition of

uncertainty with partially contradictory knowledge being observable. Chosen elements were implemented in

free REx software developed in R language.

1 INTRODUCTION

Diagnostic knowledge modeling is a process of prepa-

ration and storage of knowledge for the needs of con-

structing information systems inter alia diagnostic ex-

pert systems. The process is expected to produce

transparent, comprehensible and consistent records of

diagnostic knowledge that consider the nature of se-

lected domain it pertains to. Since diagnostic infor-

mation may derive from multiple sources, the inte-

gration of knowledge is essential, its main purpose

being construction of a relevant knowledge base of

diagnostic systems. Both quality and availability of

knowledge sources play a crucial role in ensuring the

effectiveness of inference realized by the diagnostic

system. The integration requirement particularly ap-

plies to diagnostic systems designed for complex ob-

jects. The knowledge of complex objects may be ex-

tensive and be obtained from independent experts; it

may derive from various literary sources and be im-

plicitly represented, for instance in the form of data

acquired during diverse diagnostic experiments.

Integration of knowledge from multiple sources

requires development of new methods of diagnostic

knowledge modeling with application of easily inter-

pretable manners of knowledge representation. Fur-

thermore, the knowledge modeling process should en-

able independent experts’ contribution to diagnostic

system preparation as well as ensure that modiﬁca-

tions and knowledge updates are easily performed.

An introduction of a uniﬁed knowledge repre-

sentation for multiple types of knowledge sources

constitutes the key step to obtain a convenient tool

for knowledge modeling in general, and diagnostic

knowledge in particular. Unfortunately, it is a dif-

ﬁcult task because of heterogenous characteristic of

diagnostic data and commonly used heterogenous di-

agnostic models. Considering the nature of diagnostic

knowledge, the inference process should be based on

reasoning under uncertainty conditions. This, in turn,

requires that special forms of this knowledge repre-

sentation as well as appropriate inference algorithms

be implemented.

A wide range of techniques of diagnostic mod-

eling using different knowledge representation and

inference algorithms is known to be currently avail-

able. A scope of them can be ﬁnd in (Korbicz et al.,

2004). Apart from process models (object models)

used in model based diagnosis, typical diagnostic

models which mapping observed measurement sig-

nals into qualitative description of technical states of

process can be generally classiﬁed in terms of their

transparency into two groups. The ﬁrst group com-

prises black box models which structure and parame-

ters are generally identiﬁed during the machine learn-

ing process and can not be tuned or updated without

restarting learning process. Neural networks classiﬁer

is an example. Parameters of these models are gen-

140

Cholewa W., Amarowicz M., Chrzanowski P. and Rogala T..

Multimodal Statement Networks for Diagnostic Knowledge Modeling and Integration.

DOI: 10.5220/0004519501400147

In Proceedings of the International Conference on Knowledge Engineering and Ontology Development (KEOD-2013), pages 140-147

ISBN: 978-989-8565-81-5

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

erally hard to understand and their integration with

explicit expressed knowledge (e.g. opinions) is often

impossible. The second group is the part of transpar-

ent diagnostic models (e.g. graphical models) (Cow-

ell et al., 2003), (Kohler and Friedman, 2009), which

structure and parameters may have own physical in-

terpretation or contractual meaning and which make it

easier to integrate knowledge from multiple sources.

In this paper we focus on transparent forms of knowl-

edge representation for complex objects treating only

black box models as a source of data reconstruction

or a source of preprocessed input data.

Aspects of diagnostic knowledge management

concerned with modeling, storaging, updating and

veriﬁcation of consistency of modeled knowledge are

main objective of this paper. The general model-

ing process is conducted with the use of multimodal

statement networks introduced in (Cholewa, 2010),

including, inter alia, types of network layers, meth-

ods of layer interoperability, as well as construction

and development of models. Furthermore, methods

of veriﬁcation of consistency of modeled knowledge

were addressed as well. Multimodal models enrich

the available range of methods (Cholewa, 2010). To

name a few, examples may include multiaspect mod-

els (Skupnik, 2009) or models with context inference

(Timoﬁejczuk, 2012).

Finally, the authors presented free software (REx)

(Cholewa et al., 2011), which provides a possibility to

perform actions described in this paper that are related

to modeling and management of diagnostic knowl-

edge.

2 MULTIMODAL STATEMENT

NETWORKS

One of the convenient methods of information rep-

resentation in expert systems includes a statement

which is equivalent to assertion on recognition of an

indicative expression resulting from observed facts,

or representing an opinion. One may distinguish be-

tween simple and complex statements. Simple state-

ments are presented in the form of a pair < c, v >,

where c is the content of the statement, and v is the

value (e.g. logical value) of the statement. Com-

plex statements, however, are presented in the form

of < c, v >, where c is a set of possible variants of

the statement content, and v is a set of values of sub-

sequent variants of its content. Provided the state-

ment content is constant, and that the set of variants

of statement content is an exhaustive set of mutually

excluding variants of statement content, then the com-

plex statements may be represented by means of ade-

quate sets of simple statements.

In a set of statements, one can observe a set of

statements with known values, and a set of statements

with unknown values. The statements of known val-

ues include primary statements whose values were di-

rectly determined by external processes, such as mea-

surement data or a user input, as well as constant

statements whose values were arbitrarily assumed by

e.g. a knowledge engineer. In turn, a set of state-

ments of unknown values includes secondary state-

ments whose values are strictly contingent on other

statement values, and are not directly deﬁned by ex-

ternal processes as well as isolated statements inde-

pendent of other statements. The main purpose of

inference is to determine values of secondary state-

ments for desired values of primary statements. A di-

vision into the sets of primary and secondary state-

ments is subject to change depending on data avail-

able from external sources at a given moment of time.

For instance in the case of diagnosis process sec-

ondary statements are related mainly to fault or mal-

function and primary statements to observed process

variables or user input. In the case of root cause

recognition, the secondary statements are related to

the unknown causes of observed faults (primary state-

ments). Because the process of inference is bidirec-

tional the ﬁrst case of inference and the second one

can be conducted in the same model.

Statements may be studied as approximate state-

ments provided their content and/or values are ap-

proximate. Approximate values of statements can be

deﬁned as degrees of truth or degrees of belief in the

truth of statements. The value of approximate state-

ments s is to be considered as a point value b(s),

or as an interval value, e.g. b(s) = [0.6, 0.9]. Such

an approach also allows for considering the point

value as a special form of an interval value, e.g. for

b(s) = 0.3 the interval value shall be represented as

b(s) = [0.3, 0.3].

In (Cholewa, 2010) the author introduced con-

cepts of multimodal statement networks (multimodal

model) where selected nodes represent statements.

The network structure is a multilayer one and is de-

ﬁned as a directed hypergraph described by ordered

triple:

< V, E, Γ >, (1)

where V is the set of all hypergraph vertices, E is the

set of hypergraph edges, and Γ is the set of hyper-

graph modes representing selected layers of the net-

work (Heath and Sioson, 2009). The network layers

are deﬁned as

= {V

, E

}, Γ

= {V

, E

}, (2)

where V

⊂ V , V

⊂ V , E

⊂ E, E

⊂ E. It is as-

sumed that card(V

∩V

0) and it is estimated that

MultimodalStatementNetworksforDiagnosticKnowledgeModelingandIntegration

141

the covering degree of vertices of component models

is not signiﬁcant card(V

∩ V

) << card(V ). Mul-

timodal networks consider different types of interac-

tions which occur between elements of this network.

While implementing the belief networks, the cor-

rect determination of values of conditional probabil-

ity tables poses the greatest difﬁculty. This proce-

dure may be simpliﬁed by introduction of knowledge

representation in the form of necessary and sufﬁcient

conditions. If the belief in the truth of statement s

is always accompanied by the belief in the truth of

statement s

, however, not necessarily inversely, then

statement s

constitutes the sufﬁcient condition for

statement s

, whereas statement s

is a necessary con-

dition for statement s

Assuming that the degree of belief in the truth of

and s

statements have the following value:

b(s

) ∈ [0, 1], b(s

) ∈ [0, 1]. (3)

The information on s

being a sufﬁcient condition for

may take the following form:

b(s

) 6 b(s

). (4)

If the complete belief in the truth of s

statement is

b(s

) = 1, then the result would be b(s

) = 1, in other

words, the complete belief in the truth of statement

. What is more, considering the absence of belief in

the truth of statement s

, i.e. b(s

) = 0, only a trivial

conclusion may be formulated, i.e. b(s

) = [0, 1]. The

necessary and sufﬁcient conditions may be presented

in the form of graphs Fig.1, where the directed edge

deﬁnes the increased value of the degree of belief in

the truth of the statements.

Figure 1: A graph presenting statement s

as a sufﬁcient

condition for statement s

, and statement s

as a necessary

condition for statement s

The studied expert systems should enable infer-

ence even in imprecise, incomplete and, at times,

even contradictory information environment. In such

a case approximate knowledge may be represented as

approximate necessary and sufﬁcient conditions with

the application of permissible deviation δ:

b(s

) − δ 6 b(s

), where δ > 0 (5)

The deviation value δ may be considered equal for

all conditions, or be individually considered for each

statement, or be individual δ

p,n

for each condition de-

ﬁned for a pair of statements {s

, s

b(s

) − δ

p,n

6 b(s

), where δ

p,n

> 0 (6)

The advantage of multimodal statement networks

is that selected nodes of the network may simul-

taneously occur in many layers. Furthermore, one

may also implement different types of component net-

works based on necessary and sufﬁcient conditions

(called approximate statement networks), or bayesian

networks. In order to obtain the ﬁnal results of infer-

ence, it is essential that appropriate methods of results

aggregation for the nodes shared by a number of lay-

ers be used.

3 DIAGNOSTIC KNOWLEDGE

MODELING

Declarative diagnostic knowledge may manifest na-

ture of uncertain knowledge. Thus, the sources of

this uncertainty is to be given particular attention as

the information included in diagnostic signals may be

irrelevant and insufﬁciently correlating to the object

state (Cempel and Natke, 2011). It may also result

from measurement errors or incomplete information

caused by a faulty measurement path. Knowledge de-

ﬁned by independent experts may be ambiguous, e.g.

a semantic range of notions used by different experts

may be various, or partially contradictory when ex-

perts’ opinions on the same topic differ; it may also be

quite frequently articulated in an approximate man-

ner.

Modeling of diagnostic knowledge using multi-

modal statement networks may involve:

• considering the need and the purpose of division

of a complex diagnostic model into a set of sim-

ple models represented by layers of a multimodal

model. This stage should emphasize the improve-

ment of modeled knowledge transparency, or en-

sure a possibility for selected fragments of domain

knowledge to be analyzed in an independent man-

ner by a group of independent experts.

• agreeing on strategy of multimodal model devel-

opment (describing modeled knowledge by means

of a model),

• developing layers of a multimodal model using

available representations of statement network

layers (developing structure and determining pa-

rameters of selected layers of the model),

• validating model layers (studying layers with the

use of sensitivity analysis, measuring conditional

independence, analyzing syntactical correctness

of connected knowledge),

• synthesizing layers of the multimodal model,

KEOD2013-InternationalConferenceonKnowledgeEngineeringandOntologyDevelopment

142

• tuning parameters of the multimodal model, in-

cluding parameters related to synthesis of selected

model layers.

The stages described above assumed that knowl-

edge modeling is realized with a so-called closed

world assumption, i.e. a ﬁnite set of available state-

ments as well as content variants of these statements.

Application of a method of a domain knowledge

modeling, considering general multimodal model

structure, is strictly dependent on a number of fac-

tors. It may be partially dependent on the structure of

a diagnosed object as well as subject to availability of

knowledge sources pertaining to that object. A need

of appropriate knowledge de-fragmentation facilitat-

ing interpretation and further editions may also affect

the division.

One strategy of diagnostic modeling is the use of a

multiscale model which constitutes a special form of

a multimodal model with its layers representing diag-

nostic knowledge at various detail levels, and thus, re-

sembling a hierarchical structure. An example is mul-

tiscale model in which one of the layers represents a

particular fragment of a subsystem, whereas another

layer represents a component of this subsystem, e.g.

hydraulic system and a centrifugal pump. Yet another

example is a model whose layers represent uneven de-

gree of granulation of the domain knowledge in ques-

tion. For instance, while the ﬁrst layer describing

general knowledge derived from ISO directives con-

cerns operation of a given object and permissible level

of relative vibration in bearing support structures for

given machine types, the second layer includes de-

tailed, speciﬁc knowledge of the selected rotor bear-

ing support structure.

Diagnostic knowledge modeling with a use of

multimodal models is a simpliﬁed modeling process,

and goes beyond the mentioned strategies of diagnos-

tic knowledge modeling with multiaspect and multi-

scale models. Figure 2 presents an example of multi-

modal network with fragments of multiscale and mul-

tiaspect networks. A diagnostic model of a rotor sup-

ported by 2 hydrostatic bearings may include the fol-

lowing exemplary layers:

• L0 layer represents a portion of knowledge con-

cerning a general technical condition of an ob-

ject as well as a technical condition of the object

subsystems, e.g. statements with the following

content: technical condition of object ={normal,

faulty}; technical condition of hydraulic subsys-

tem ={normal, faulty}, etc.

• L1 layer describes general diagnostic knowl-

edge of the object, for instance, value of com-

ponent frequency corresponding to rotor rota-

tional speed ={high, medium and low}, level

of dynamic unbalance ={permissible, exceeded,

impermissible},

• L2 describes general knowledge of a chosen sub-

system e.g. hydraulic system. A layer may con-

tain statements concerning values of stream mass

levels of a working ﬂuid {low, medium, high},

pump rotational speed {low, high},

• L2.1 and L2.3 layers present diagnostic knowl-

edge of chosen components of hydraulic system

described in L2 layer (pumps, an electro hydraulic

actuator); exemplary statements include operation

temperature of pumps measured on the housing

{low, regular, high}, leakage in pump on the suc-

tion site {no, yes},

• L3 describes knowledge of the object functional

state that is recognized on the basis of selected

primary statements such as shaft rotational speed

{constant, quasi variable, variable}, functional

state {operation mode number 1, operation mode

number 2}.

The presented model only illustrates the structure

of a multimodal model. Depending on the require-

ments, this model may be extended or re-constructed.

It is also possible to develop layers that would in-

clude, inter alia, diagnostic knowledge for remaining

components subsets, physical states of chosen work-

ing ﬂuids. Furthermore, additional layers may supple-

ment the already available knowledge, however, may

be also introducing different opinions by independent

experts. What is so essential about layers of the de-

veloped multimodal model is that not only does one

observe relatively low covering degree of vertices in

subsequent layers of the model, but also that it, to a

large extent, refers to the speciﬁc knowledge of the

examined object. In the case of modeling of a class of

signiﬁcantly similar objects for which general domain

knowledge may be distinguished, the knowledge may

be also a subject of ontological modeling.

4 KNOWLEDGE MANAGEMENT

Implementation of effective diagnostic systems in-

volves processing of large sets of diagnostic knowl-

edge. One of the key objectives is to devise efﬁcient

management methods. Among elements directly re-

lating to knowledge management for the needs of di-

agnostic modeling one can distinguish the following

phases taking place in machine and process diagnos-

tics:

• diagnostic knowledge acquisition,

• veriﬁcation of acquired knowledge,

MultimodalStatementNetworksforDiagnosticKnowledgeModelingandIntegration

143

Figure 2: An exemplary structure of a multimodal model.

L0 - general layer, L1- layer with general knowledge, L2-

layer with a chosen subsystem, e.g. hydraulic system, L2.1-

centrifugal pump, L2.3-electro hydraulic valve, L3-layer

with description of functional state.

• connection of knowledge to a knowledge base,

• application of collected knowledge in the diagnos-

tic process,

• evaluation of sets of collected data, modiﬁcations

and updates entry.

The manner in which knowledge is represented

has to correspond to the method of inference being

in line with algorithms, and, simultaneously, it has

to allow knowledge engineers to add, modify and

delete base components. Modifying the knowledge

base may be required in the case of contradictory in-

formation/errors, or when new elements have to be

connected to the available set.

In order to effectively use knowledge in the form

of multimodal networks, appropriate sets of consid-

ered statements, also referred to as thesauri, are being

developed under a closed world assumption. Each el-

ement of this set, i.e. an individual statement, is usu-

ally deﬁned by a chosen set of attributes facilitating

statement management. With an exception of state-

ment content and value assigned to it, each statement

may be described by, among others, an identiﬁer, an

abbreviation, key words, statement versions, author,

etc. The following statement may serve as an exam-

ple:

id: glst-053,

content: excessive vib. in bearing node N5,

version: 001,

id-source: mast-023,

value: 1,

abbreviation: vibration-n5,

key-words: vibration, bearing,

author: mamarowicz.

Complex statements are additionally deﬁned by

content variants of the statement. Some of the at-

tributes are optional and do not signiﬁcantly affect the

quality of communication with thesaurus. Nonethe-

less, some of them, for instance statement identiﬁer

and version, are crucial for proper collection and ap-

plication of particular statements.

Explanation systems also aid in handling of the

statement sets (Cholewa, 2004). Their main purpose

is to determine chosen terminology that is used in

construction of particular statements. A developed

system of explanations, also referred to as a dictio-

nary, comprises explanations of chosen deﬁnitions re-

lating to given elements as well as subsets of moni-

tored objects, possible faults, fault symptoms, diag-

nostic methods and techniques, etc. Additionally, it is

possible to create explanations to the content of each

statement in order to obtain a more detailed statement,

or to draw particular attention to chosen speciﬁc con-

texts of use of a statement in the process of construc-

tion and application of diagnostic models.

Statements may be used in developing multimodal

statement networks which e.g. could present knowl-

edge describing known cause-and-effect dependen-

cies related to addressed problems. Deﬁnition of par-

ticular networks involves, above all, tasks of deter-

mining topological structure, i.e. deﬁning particular

nodes and relations occurring between them. There-

fore, visualization of the developed statement net-

works is so crucial. For individual networks, it boils

down to visualization of graphs, and it is important

that an optimal algorithm of node distribution be cho-

sen so that a clear form of network, i.e. one without

broken edges, can be obtained (Tamassia, 2007). In

the case of multimodal models a three dimensional

visualization may pose an interesting solution, shown

in Fig. 3. It allows for simultaneous browsing of all

model layers and easy identiﬁcation of shared nodes.

A number of solutions devised for multimodal mod-

els include additional virtual layer onto which all

nodes used for construction of component networks

were distributed. This layer facilitates information ex-

change as well as data input and network data reading.

KEOD2013-InternationalConferenceonKnowledgeEngineeringandOntologyDevelopment

144

Figure 3: An example of multimodal network visualization,

in system REx.

5 FORMAL VERIFICATION OF

KNOWLEDGE

A formal veriﬁcation of diagnostic knowledge is

an important stage of construction of a knowledge

base. A syntactic veriﬁcation, which consists in for-

mal searching of typical errors in adequately formal-

ized and stored knowledge in data base, is a com-

monly applied method of verifying diagnostic knowl-

edge. Among the methods of formal syntactic veriﬁ-

cation of knowledge base, one may enumerate meth-

ods basing on error detection through knowledge base

searching, as well as methods using parasyntactic

logic (Nguyen, 2005). These processes most fre-

quently consist in elimination of redundancies, loops,

excess, e.g. subsumed rules or rules not affecting

the distinguishing between secondary statements, etc.

For the majority of knowledge management systems

it is a stage which, depending on knowledge repre-

sentation method, can be realized in an algorithmic

manner. Systems with rule-based knowledge serve as

an example.

Further methods of knowledge veriﬁcation in-

clude examination of structure of stored knowledge.

This applies to knowledge representation that refers

to tree or network structures to which graphical net-

works discussed in this paper, may also be assigned.

Syntactic veriﬁcation of such structures is simpler.

For instance, considering the layers represented by

means of belief networks due to non-cyclical and di-

rected graph structure, it is not possible to add knowl-

edge which could lead to looping of inference algo-

rithm, which could take places in regular rule-based

systems. Another element that facilitates knowledge

veriﬁcation is a convenient visualization tool of the

network structure as well as appropriate navigating

tools. This is particularly crucial when dealing with

spatial networks such as multimodal statement net-

works.

In the case of ﬂat network or tree models, iden-

tiﬁcation of looping rules is not necessary provided.

Inference in the system takes place within the closed

world assumption, and a deﬁned set of statement con-

tents is semantically coherent. The accuracy of the

variant semantics of chosen statements is subject to

knowledge management using thesaurus discussed in

Section 4. This, however, does not exclude a neces-

sity of knowledge veriﬁcation in terms of its seman-

tics. For diagnostic systems based on knowledge one

may distinguish the following causes of semantic in-

accuracy and knowledge inconsistency:

• knowledge was deﬁned on the basis of learning

data that included errors, for instance contradic-

tory information. This may lead to an inaccurate

structure of mutual relations between statements

that can result in emergence of associations be-

ing statistically independent, or to deﬁnition of

inaccurate values of graphical model parameters.

An example includes inaccurate values of condi-

tional probabilities tables for layers represented

by means of belief networks. And for these values

the graphical model does not manifest expected

sensitivity to changes in values in selected pri-

mary statements (use of relevant associations),

• data on the object as well as primary statements on

the object are not consistent in the temporal sense,

as a result, a part of data obtained during the sys-

tem operation is not consistent with other data or

other statements. Correct deﬁnition of knowledge

in inference for inconsistent information in the

temporal sense may lead to contradictory conclu-

sions. Additionally, such inconsistency may also

result from errors of spatial inconsistency of data

that are caused by, among others, wrong identi-

ﬁers on diagnostic signal tags,

• knowledge of relations between statements was

deﬁned on the basis of independent knowledge

sources, e.g. experts whose opinions may be par-

tially contradictory.

Further discussions concern formal methods of

identifying knowledge inconsistency relating to the

last two mentioned causes.

Detection of inconsistency, that may occur as a re-

sult of measurement errors, spatial or temporal dis-

crepancies of collected data, may be identiﬁed using

statement networks as well. Examples include state-

ment networks whose layers are represented by means

MultimodalStatementNetworksforDiagnosticKnowledgeModelingandIntegration

145

of layers described through belief networks. Provided

a graphical model is deﬁned correctly, it is possible to

detect contradictory information that leads to incon-

sistency in networks of this type. Since it may result

from errors, error not necessarily being always the

root cause, for instance, they may trigger consider-

ably rare events analyzed by developed domain mod-

els, it is then the task of inference module to raise a

notiﬁcation of such an occurrence. For more informa-

tion regarding the mentioned methods, authors refer

to, among others, (Jensen and Nielsen, 2007), (Kohler

and Friedman, 2009). Similar solutions are identi-

ﬁable in networks represented by approximate state-

ment networks. Contradictory information is simulta-

neously identiﬁed with a particular value of parameter

δ in (5) and (6) occurring between the statements in

which the inconsistency is detected in the process of

inference. The δ parameter values are useful for iden-

tifying potentially contradictory information. Subse-

quently, in order to determine whether given incon-

sistency is triggered by inaccurate data or not, such

an analysis is to be carried out independently for each

individual network layer.

Detection of contradictory information estab-

lished on the basis of various opinions of independent

experts, is quite frequently considered a secondary re-

quirement (Nguyen, 2005). The primary requirement

is to develop a set of aggregated expert opinions. It

is essential to note, however, that a proper method of

opinion aggregation of a large group of experts should

also facilitate identiﬁcation of the sources of conﬂict.

Any information pertaining to potential conﬂict may

be useful in developing adequate aggregating func-

tions, for instance weights assigned to given experts.

Approximate statement networks also enable easy

identiﬁcation of inconsistency between opinions ex-

pressed by various experts. If the information newly

added to the system (a change in values of selected

primary statements) does not affect the value of pa-

rameter δ, where each layer is studied independently,

but it triggers a change in this value once the selected

layers have been studied, it may indicate potential in-

consistency in the layers in question. In addition, in-

consistency may also be detected on the basis of asso-

ciation analysis between statements described in dif-

ferent layers of multimodal networks.

6 REX SYSTEM

The research version of REx system, which operates

in R language, was designed and developed not only

for the purpose of modeling of diagnostic knowledge

in the form of multimodal network statements, but

also to support veriﬁcation, testing and management

of diagnostic knowledge. The essential system com-

ponents include:

• Bulletin Board - a component that collects and

shares information on values of selected state-

ments,

• Data Uploader - a module responsible for infor-

mation/data exchange with other systems,

• Explanation System - a component that collects

and shares explanations for selected statements,

• Model and Benchmark Repository - a module that

facilitates collection of thesauri and multimodal

models,

• User GUI - Graphic User Interface that enables

communication between the system and users not

familiar with R language syntax,

• Data Acquisitor - a module responsible for acqui-

sition of data from a technical object,

• Data Simulator - a module that simulates opera-

tion of a technical object,

• Knowledge Acquisitor - a module responsible for

acquisition of diagnostics knowledge.

A prototype version of REx system includes

Bulletin Board, Data Uploader, Model, Benchmark

Repository and User GUI. The further components

are in a testing phase with an emphasis on their practi-

cal application. Examples include a module that sim-

ulates operation of a technical object, and a module

that acquires data from the technical object. REx can

be used for own solutions within the R environment.

REx package offers a possibility of:

• deﬁning thesauri,

• deﬁning subsequent layers, their structures and re-

quired parameters,

• entering data to input nodes,

• performing inference.

One implemented an algorithm for determination

of results of network calculations that considers dif-

ferent types of networks stretched onto selected layers

(of belief networks and approximate statement net-

works).

The installation pack can be downloaded at

http://www.ipkm.polsl.pl/index.php?n=Projekty.Rex.

7 SUMMARY

Design and development of diagnostic systems with

the use of multimodal statement networks requires

KEOD2013-InternationalConferenceonKnowledgeEngineeringandOntologyDevelopment

146

implementation of a method of diagnostic knowledge

modeling that allows for integrating knowledge from

various sources. It also emphasizes the nature of mod-

eled knowledge which may be approximate, incom-

plete, inconsistent, or partially contradictory. Intro-

duced networks offer a possibility of modeling knowl-

edge with a participation of independent experts who

can develop selected layers of diagnostic models.

Graphical models and multimodal statement networks

require application of a number of additional tools in-

tended for diagnostic knowledge management. These

include the tools discussed in this paper, tools used for

semantic and syntactic veriﬁcation of implemented

knowledge, as well as tools enhancing composition of

a dictionary of statements and statement management.

The presented aspects are being currently developed

within the available software REx devised in R pro-

gramming language. REx environment facilitates fur-

ther research into advanced applications and develop-

ment of multimodal statement networks for the pur-

pose of modeling of diagnostic knowledge, and re-

search into methodology of interoperability between

multimodal model layers that would consider incon-

sistency occurring in both data and knowledge.

ACKNOWLEDGEMENTS

Described herein are selected results of study, sup-

ported partly from the budget of Research Task No.

4 implemented under The National Centre for Re-

search and Development in Poland and ENERGA SA

strategic program of scientiﬁc research and develop-

ment entitled ,,Advanced technologies of generating

energy.”

REFERENCES

Cempel, C. and Natke, H. (2011). Model-Aided Diagno-

sis of Mechanical Systems: Fundamentals, Detection,

Localization, Assessment. Springer, Berlin.

Cholewa, W. (2004). Expert systems in technical diagnos-

tics. In Korbicz, J., Ko

scielny, J. M., Kowalczuk, Z.,

and Cholewa, W., editors, Fault Diagnosis. Models,

Artiﬁcial Intelligence, Applications, pages 591–632.

Springer-Verlag, Berlin.

Cholewa, W. (2010). Multimodal statements networks for

diagnostic applications. In Sas, P., Bergen, B. (eds.)

Proceedings of the International Conference on Noise

and Vibration Engineering ISMA 2010, September

20-22, pp. 817-830. Katholique Universiteit Lueven,

Lueven (2010).

Cholewa, W., Rogala, T., Chrzanowski, P., and Amarowicz,

M. (2011). Statement networks development environ-

ment Rex. In Computational collective intelligence.

Technologies and applications. ICCCI 2011, Gdynia,

Poland, September 21-23, 2011. Proceedings. Pt 2.

Eds: P. Je¸drzejewicz, N.T. Nguyen, K. Hoang. pp. 30-

39, (LNCS vol. 6923). Springer.

Cowell, R. G., Dawid, P. A., Lauritzen, S. L., and Spiegel-

halter, D. J. (2003). Probabilistic Networks and Ex-

pert Systems. Springer, New York.

Heath, L. and Sioson, A. (2009). Multimodal networks:

Structure and operations. IEEE/ACM Trans. Com-

putational Biology and Bioninformatics, 6(2):pp.321–

332.

Jensen, F. and Nielsen, T. (2007). Bayesian networks and

decision graphs. Springer, Berlin.

Kohler, D. and Friedman, N. (2009). Probabilistic graph-

ical models: principles and techniques. MIT Press,

Cambridge.

Korbicz, J., Ko

scielny, J. M., Kowalczuk, Z., and Cholewa,

W. (2004). Fault Diagnosis. Models, Artiﬁcial Intelli-

gence, Applications. Springer-Verlag, Berlin.

Nguyen, N. (2005). Processing inconsistency of knowledge

on semantic level. Journal of Universal Computer Sci-

ence, 11(2):pp. 285–302.

Skupnik, D. (2009). An application example of a multi-

aspect diagnostic model. In VIII International Techni-

cal Systems Degradation Seminar. TSD International

Seminar, Liptovsky Mikulas, 15-18 April 2009. Ed.

J. Ma¸czak. Warsaw University of Technology [et al.].

Warszawa, 2009, pp. 86-87.

Tamassia, R. (2007). Handbook of Graph Drawing And

Visualization. Discrete Mathematics And Its Applica-

tions. Chapman & Hall/CRC, London.

Timoﬁejczuk, A. (2012). Methodology context based

reasoning in technical diagnostics. In 5th Interna-

tional Congress on Technical Diagnostics. Diagnos-

tics’2012, Krak

ow, Poland, 3rd - 5th September 2012.

Abstracts., 2012, pp. 33-34.

MultimodalStatementNetworksforDiagnosticKnowledgeModelingandIntegration

147