SUPPORTING REUSE OF KNOWLEDGE OF FAILURES
THROUGH ONTOLOGY-BASED SEMANTIC SEARCH
Steven B. Kraines and Weisen Guo
Science Integration Program (Human), Department of Frontier Sciences and Science Integration
Division of Project Coordination, The University of Tokyo, 5-1-5 Kashiwa-no-ha, Kashiwa
277-8568 Chiba, Japan
Keywords: Knowledge representation, Ontology, Semantic Search, Domain knowledge, Knowledge reuse, Failure
knowledge.
Abstract: In order to increase the effectiveness of sharing and reusing knowledge about failures, we have applied the
expert knowledge ontology-based semantic search system, EKOSS, to the “failure knowledge database”. An
ontology based on description logics is used as a formalized knowledge representation language for creating
semantic statements describing 212 of the JST failure cases. Using the EKOSS reasoner, the similarity be-
tween a search statement giving the conditions of a new project or project design and the semantic state-
ments describing failure cases can be quantified by inferring specific semantic relationships between entities
involved in the case. The corpus of semantic statements is described, and results of applying the EKOSS
semantic search to the semantic statements are analyzed. Finally, the effectiveness of the SCINTENG on-
tology for expressing the underlying failure mechanisms of the cases is discussed.
1 INTRODUCTION
Civil engineering projects and other applications of
technology have become increasingly knowledge
intensive. Institutions involved in construction, oper-
ation, and/or disposal of engineering artifacts such as
buildings, bridges and airplanes, are held responsible
for wide-ranging safety and environmental issues.
Failures to meet promised standards can seriously
impact public perception of the institution.
Many failures that plague institutions today
could have been avoided if the knowledge of similar
previous failures had been available and the relation-
ship to the project concerned had been recognized
(Tamura, 2003). The importance of “reusing” know-
ledge obtained from the analysis of actual failures
and the elucidation of the mechanisms leading to the
failures is widely recognized (Brown, 2007; Darling-
ton and Booker, 2006; Wintle and Pargeter, 2005).
One attempt to structure the knowledge from past
failures so as to facilitate their reuse in the assess-
ment of safety and potential failure mechanisms of
new technology applications is the “failure know-
ledge database” (hereafter failure DB) (Hatamura et
al., 2003a; Hatamura et al., 2003b; JST, 2010).
EKOSS (expert knowledge ontology-based se-
mantic search) is a Web-based system for supporting
the computer-aided sharing, discovery and integra-
tion of knowledge resources (Kraines et al., 2006a).
Building on the premise that knowledge sharing and
discovery could benefit greatly if human creators of
knowledge resources created computer-interpretable
descriptors of those resources (Gerstein et al., 2007;
Marcondes, 2004), EKOSS provides knowledge
creators with tools enabling them to easily construct
computer-understandable descriptors, which we call
“semantic statements”, that describe their resources
using ontologies based on description logics (DL) as
knowledge representation languages. The semantic
statement authoring tools incorporate several auto-
matic support functions such as NLP of the source
text in order to reduce the cognitive overhead for the
human authors. EKOSS then uses the semantic
statements to provide knowledge-intensive services
such as semantic search and knowledge mining.
Semantic statements that are authored directly by
the creators of the knowledge resources are expected
to be more accurate and semantically rich than de-
scriptors that are generated automatically using NLP
(Berners-Lee and Hendler, 2001; Blake and Rendall,
2006; Buckingham Shum et al., 2007). In particular,
a computer-mediated searching and matching system
having access to EKOSS semantic statements could
use logical and rule-based inference to provide
164
B. Kraines S. and Guo W..
SUPPORTING REUSE OF KNOWLEDGE OF FAILURES THROUGH ONTOLOGY-BASED SEMANTIC SEARCH.
DOI: 10.5220/0003068001640169
In Proceedings of the International Conference on Knowledge Management and Information Sharing (KMIS-2010), pages 164-169
ISBN: 978-989-8425-30-0
Copyright
c
 2010 SCITEPRESS (Science and Technology Publications, Lda.)

knowledge services that would not be possible with
descriptors generated by NLP or written by humans
as statements that are not formalized in a computable
logic. Systems have been reported in the literature
that attempt to test this hypothesis (Di Noia et al.,
2007; Li and Horrocks, 2003; Uschold et al., 2003;
Wang et al., 2004; Halaschek-Wiener and Kolovski,
2008). However, they do not support reasoning
against DL expressions that include assertions of
relationships between individuals described as in-
stances of ontology classes, so it is not possible to
create complex statements that involve, for example,
relationships between different instances of the same
class. Furthermore, most of these systems do not
have user interfaces that are designed to enable users
without information science backgrounds to create
logically based computer-understandable semantic
statements.
The EKOSS system has been designed to be in-
tegrated into the existing scientific knowledge publi-
cation process, so authoring the semantic statements
would become just one step in the process of creat-
ing a scientific publication. The idea of incorporating
creation of computer-interpretable descriptors in the
process of publishing research articles has been
raised before (Marcondes, 2004; Berners-Lee and
Hendler, 2001; Rzhetsky et al., 2008; Ceol et al.,
2008). However, previous attempts to design such a
system have had limited success. The EKOSS sys-
tem aims to overcome the limitations of these pre-
vious attempts by providing a solid foundation in
computer-interpretable semantics together with the
intuitive authoring tools (Kraines et al., 2006a).
Here, we describe the application of EKOSS to
the failure DB. We first give the motivation for con-
structing the failure DB and for applying EKOSS to
it. Next, we present a corpus of semantic statements
that were created for 212 of the cases using EKOSS,
and we show how the semantic richness of the
statements can improve the accuracy of searching
the failure DB for knowledge that could be reused in
new engineering applications. Finally, we discuss the
effectiveness of the ontology as a conceptualization
for expressing the mechanisms behind technology
failures and suggest some areas for improvement.
2 FAILURE KNOWLEDGE
In 2002, the Japan Science and Technology Agency
(JST) started a project to create a database, the fail-
ure DB, containing case studies of major failures in
science and technology (JST, 2010). One aim was to
make the knowledge from the case studies reusable
in new engineering applications to avoid repeat of
past failures (Tamura, 2003; Hatamura et al., 2003a).
Over five years, descriptions for over 1000 cases of
major failures from around the world, including
famous disasters such as the sinking of the Titanic,
were prepared under the supervision of five experts
on the analysis of failure mechanisms in chemical
engineering, mechanical engineering, material
science, and civil engineering. The descriptions were
made available in both English and Japanese on the
JST “shippai chishiki” website (JST, 2010).
The failure DB is structured via three “failure”
Mandala, which classify types of causes, actions and
results of failure processes (Hatamura et al., 2003b).
For each failure case, key terms from each Mandala
were chosen and arranged in a scenario that depicts
the unfolding of the events surrounding the particu-
lar knowledge failure.
3 SEMANTIC STATEMENTS
Semantic statements were created for 212 of the fail-
ure cases, including the 100 “hyaku-sen” cases form-
ing the core of the failure DB. The semantic state-
ments were authored using a DL ontology, called
SCINTENG, elements and applications of which
have been presented elsewhere (Kraines et al.,
2006b; Kraines et al., 2006c). By using a DL reason-
er, knowledge sharing services can be provided
based on inference at the level of semantic relation-
ships, e.g. to identify failure cases that have similar
causal mechanisms to new engineering failures.
Terms in the Mandala for “cause of failure” and
for “action” were added to the SCINTENG ontology
as subclasses of “class of human activity.” However,
most terms in the Mandala for “result” are actually
events or properties of physical objects. Therefore,
we have included them in the SCINTENG ontology
under the relevant parent classes of “disaster event”
and “abnormality property.”
Using the EKOSS authoring tools, each semantic
statement took one to four hours to create. The se-
mantic statements were created using information
given in the failure DB as follows:
1. Establish an event-activity chain that describes
the course of the incident based mainly on the
“sequence” and “cause” case descriptions.
2. Add physical objects, materials, locations, times,
etc. by referring mainly to information in the
“sequence” and “cause” case descriptions.
3. Refer to the “scenario” to add failure types to the
activities.
The semantic statements contain an average of
35 instances and 45 properties each, for a total of
7455 instances and 9603 properties. Of the 1000
SUPPORTING REUSE OF KNOWLEDGE OF FAILURES THROUGH ONTOLOGY-BASED SEMANTIC SEARCH
165

classes in the SCINTENG ontology, 507 were used
at least once, and 135 were used 10 times or more.
38.6% of the instances are physical objects, 17.1%
are activities, 17.2% are events, 8.8% are classes of
activities, 6.9% are substances, and 11.4% are other
classes such as quantities, units, and properties. Of
the 195 properties in the ontology, 112 were used at
least once, and 63 were used 10 or more times. Most
of the properties are inverses or subproperties of
“has activity participant” (15.6%), “has event partic-
ipant” (14.7%), “has end event” (10.2%), “has start
event” (8.8%), “has location” (8.1%), “has activity
class” (6.9%), “has part” (6.0%), “has substance”
(5.0%), or “physically contains” (5.2%).
4 QUERY MATCHING ANALYSIS
Direct measurement of the increase in precision and
recall of knowledge searches that can be obtained by
the EKOSS system requires enough semantic state-
ments describing search targets such as case reports
to make comparisons with conventional search en-
gines. Because we only have a limited number of
semantic statements, we have developed a different
approach for assessing the ability of EKOSS to in-
crease search precision and recall.
Figure 1: The general query for a search for knowledge
describing some activity or phenomenon with a specified
class of activity and an end event that is the start event of a
second activity, which ends in a second event that has a
participating physical object. Boxes show query variables
and are labeled “class name”, colon, “instance label”.
Physical objects and events are shown as grey and white
square boxes. Activities and classes of activities are shown
as white and grey rounded boxes. Arrows show properties
asserted between pairs of query variables.
First, we created a general search query with six
instance variables of top level classes and five prop-
erties from the SCINTENG ontology (Figure 1).
This query matches with 88 of the 212 semantic
statements, so we can consider it to be a commonly
occurring semantic pattern. Next, we chose three
semantic statements from the 88 matching the gener-
al query, and we used the mid-level classes of the
ontology that occurred in at least 50 statements,
shown in Table 1, to manually construct three search
queries based on the general query that resulted in
one-to-one matches (100% precision) with each of
the selected semantic statements. The resulting “base
queries” are shown in Tables 2 to 4. Instance va-
riables are labeled by the name of the class followed
by a capital letter to differentiate variables having
the same class.
Table 1: Commonly occurring mid-level classes, the total
count of instances of the class and all of its subclasses, and
the major superclass from the knowledge model.
Class Count Top Class
artificial physical object 1108 physical object
single body event 475 event
urban region 413 spatial location
person 354 actor
disaster event 313 event
artifact actor 288 actor
machine artifact 288 physical object
termination event 267 event
multibody event 251 event
fluid object 248 physical object
operation 232 activity
reaction phenomenon 225 activity
industrial plant 135 physical object
chemical reaction 121 activity
mass transport 116 activity
Next, we removed the properties from the base
queries to create queries containing just lists of mid-
level classes, and we matched those queries against
the 212 semantic statements. The number of addi-
tional matches that occur can be considered “false
positives”, which indicate the increase in precision
Table 2: The base query for the case “Collapse of the
Seongsu Bridge in Seoul, Korea,” hereafter “Seongsu
Bridge.” Domain is the domain instance variable, range is
the range instance range, and relation is the property con-
necting the two variables.
Domain Relation Range Instance
operationA has class human failureA
physical objectA participant of termination eventA
termination eventA end of mass transportA
termination eventB start of mass transportA
termination eventB end of operationA
that was achieved by using the properties as a filter.
Finally, we changed selected mid-level classes in the
KMIS 2010 - International Conference on Knowledge Management and Information Sharing
166

base queries back to the corresponding top-level
classes, keeping all of the properties this time. The
new matches resulting from these operations are
potential “true positives” that were missed by the
base queries, which indicate the increase in recall
achieved by using inference based on class subsump-
tion. These operations to “relax” the query restric-
tions are shown in Table 5 together with the number
of statements matching the query for each case.
Table 3: The base query for the case “Fire disaster in the
Tokyo-Nagoya Nihon-zaka tunnel,” hereafter “Nihon-zaka
tunnel.” See Table 2 for explanation of table columns.
Domain Relation Range Instance
artificial physical
objectA
participant of destruction eventA
destruction eventA end of chemical reactionA
eventA start of chemical reactionA
eventA end of operationA
operationA has actor artifact actorA
operationA has class human failureA
Table 4: The base query for the case “Leakage of primary
coolant at Mihama Unit 2,” hereafter “Leakage at Miha-
ma.” See Table 2 for explanation of table columns.
Domain Relation Range Instance
artificial physical
objectA
has substance elemental
materialA
artificial physical
objectA
participant of reactionA
eventA end of human activityA
eventA start of reactionA
fluid objectA participant of single body
eventA
human activityA has class human failureA
machine artifactA has part artificial physical
objectA
5 DISCUSSION
Eliminating properties from the base queries added 6
to 13 matches (Table 5). This confirms that the one-
to-one matches with the selected semantic statements
did not result just from choosing rare combinations
of classes. It also shows that the high precision of the
base queries results at least in part from the assertion
of properties between the instances.
Table 5: Results of the query relaxation analysis. “Relaxa-
tion Change” is the specific change made to the base
query: “>>” indicates a class change, “HA” is “human
activity” and “PO” is physical object. Labels “Rx” indicate
that the change for the entry having the same label in the
“Label” column was made, e.g. “R2+R3” means that “op-
eration” was changed to “human activity” and “mass
transport” was changed to “activity”. “Num” is the number
of matching cases, which includes the selected semantic
statement, so there is always at least one match.
Case Relaxation Change Num Label
Seongsu
Bridge
remove all properties 13 -
operation >> HA 2 R2
mass transport >>
activity
2 R3
termination eventB >> event 5 R4
R2+R3 4 -
R2+R4 7 -
R3+R4 12 -
R2+R3+R4 18 -
Nihon-
zaka tunnel
remove all properties 7 -
artifact actor >> actor 2 R1
operation >> HA; chemical
reaction >> activity
1 R2
artificial PO >> PO; destruc-
tion event >> event
2 R3
R1+R2 3 -
R1+R2+R3 21 -
Leakage at
Mihama
remove all properties 14 -
delete “elemental
material”
1 R1
fluid object >> PO 2 R2
reaction phenomenon >>
activity
1 R3
R1+R2 3 -
R1+R2+R3 7 -
R1+R2+R3+
machine artifact >> PO
11 -
No clear pattern emerged from replacing single
classes in the base queries with corresponding upper
classes. In the “Seongsu Bridge” and the “Nihon-
zaka tunnel” queries, replacing the termination or
destruction event with “event” had the greatest effect,
while in the “Leakage at Mihama” query, only re-
placing specific physical objects with “physical ob-
ject” resulted in increases of matches.
In almost all cases, replacing combinations of specif-
ic classes with their corresponding upper classes
resulted in far more matches than the sum of the
additional matches when individual classes were
replaced. For example, in the “Nihon-zaka tunnel”
SUPPORTING REUSE OF KNOWLEDGE OF FAILURES THROUGH ONTOLOGY-BASED SEMANTIC SEARCH
167

query, replacing three classes simultaneously versus
replacing them separately increased additional
matches ten-fold (1+0+1 versus 20). In other words,
there were 18 statements that matched when all three
mid-level classes were replaced but do not match
when any one of the mid-level classes was replaced.
This non-linearity of query selectivity in the EKOSS
system contrasts with the linearity of Boolean
searches where queries with multiple terms con-
nected by “OR” simply give the union of the
matches for each individual term.
EKOSS allows a user to express the semantics of
a search condition both by asserting properties and
by selecting the class specificity. As examples of
increased search recall, we can consider the six addi-
tional matches to the relaxed query for “Seongsu
Bridge” labeled “R24” to be likely candidates for
cases that did not match the original query exactly
but that have similar failure mechanisms. The query
“Seongsu Bridge – R24” can be expressed as:
Some activity or phenomenon with a
specified human failure type has an end
event that is the start of a mass transport
phenomenon, which ends in a termination
event with a participating physical object.
One additional match is the case “Steam Eruption
from Nuclear Power Plant Cooling System.” The
matching part can be expressed as:
a measuring activity with disregard of
procedure ended in an artifact destruction
event causing a leaking material transport,
which ended in a death event of “5 workers”.
Another match is the case “Brittle Fracture of Hy-
drodesulfurization Reactor during Pressure Test,”
whose matching description can be expressed as:
an operation with insufficient
understanding caused a multibody
movement event that caused a mass body
explosive movement, ending in an artifact
destruction event of the “facilities of the
factory”.
Neither of these cases matched with the “Seong-
su Bridge” query when all properties were removed,
so keyword matching would not have produced these
matches. However, it is clear that they share com-
mon semantics in the mechanism by which the fail-
ure occurred. This demonstrates the benefits of the
semantic search in discovering cases from other do-
mains where the knowledge of the “Seongsu Bridge”
failure could have been useful.
Several information systems for reusing know-
ledge about failures have been reported in the litera-
ture (Goble and Bukowski, 2007; Jacobo et al.,
2007; James, 2005; Moon et al., 1998; Stone et al.,
2005; Warren Liao et al., 1999). However, none use
semantic inference to match failure cases with in-
formation needs. For example, although the ES-
FAME system uses case based and rule based rea-
soning, it does not support logical reasoning to infer
implied relationships (Jacobo et al., 2007). As dis-
cussed in the introduction, even existing systems for
semantic matching do not support reasoning about
chains of instance relationships or instances of the
same class having different attributes, which we be-
lieve is necessary for effective reuse of knowledge
about complex issues such as failure mechanisms.
The SCINTENG ontology was able to express
the failure mechanisms in the 212 cases we used.
The activity event model of the ECM gave us a way
to logically extend the original case scenarios. Be-
cause failure mechanisms are classified under “class
of individual”, which is disjoint with classes such as
“event” and “activity”, the knowledge of the activi-
ty-event chain is orthogonal from the knowledge of
the failure mechanisms. This allows us to search for
specific failure mechanisms, specific activity-event
chains, or combinations of the two. Finally, the
SCINTENG ontology lets us describe in detail the
physical objects in a failure case, including their
relationships to events and activities in the activity-
event chain, compositional relationships with each
other, and their material composition.
The SCINTENG ontology is unable to express
some important aspects of a failure case. Several
types of the concepts commonly used in case reports,
e.g. job types and event scales, are not available in
the ontology. Some of the sub-trees in the ontology
are not rigorously defined, particularly the physical
process tree and the substance tree. Adding more
complex class descriptions would reduce ambiguity
and enable more powerful logic inference. Also, the
granularity of the ontology could be increased to
give greater matching precision. However, we must
consider the tradeoff between these advantages and
the increased complexity arising from the additional
assumptions that are incorporated in the ontology
(Buckingham Shum et al., 2007). The ability to con-
struct semantic relationships helps to overcome the
relatively coarse granularity of the vocabulary. For
example, although there is only one term for “disre-
gard of procedures” in the ontology, the particular
procedure that was disregarded can be specified in
detail using semantic relationships.
This paper has attempted to demonstrate the fea-
sibility and the effectiveness of applying EKOSS to
the reuse of knowledge about failures in situations
where the applicability of the relevant knowledge is
not obvious without the use of semantic inference.
The interested reader can explore the matches to the
KMIS 2010 - International Conference on Knowledge Management and Information Sharing
168

queries presented in this paper by accessing the pub-
lic EKOSS server at http://www.ekoss.org.
ACKNOWLEDGEMENTS
We are grateful to Y Nakamura who created most of
the semantic statements used in this paper and to B
Kemper who implemented the original version of
EKOSS. We also thank Y Hatamura, H Kobayashi,
M Kunishima, M Nakao, and M Tamura for advice
concerning the analysis of the cases in the failure DB.
Funding for this research was provided by the
Knowledge Failure Database project at the Japan
Science and Technology Agency and the Office of
the President of the University of Tokyo.
REFERENCES
Berners-Lee, T., Hendler, J., 2001. Publishing on the Se-
mantic Web. Nature 410: 1023-1024.
Blake, C., Rendall, M., 2006. Scientific discovery: a view
from the trenches. In Proc.9th Intl Conf on Discovery
Science, Barcelona, Spain.
Brown, S., 2007. Forensic engineering: Reduction of risk
and improving technology (for all things great and
small). Eng Failure Anal, 14(6): 1019-1037.
Buckingham Shum, S., Uren, V., Li, G., Sereno, B., Man-
cini, C., 2007. Modeling naturalistic argument- ation
in research literatures: Representation and interaction
design issues. Intl J Intel Sys 22(1): 17-47.
Ceol, A., Chatr-Aryamontri, A., Licata, L., Cesareni, G.,
2008. Linking entries in protein interaction database to
structured text: The FEBS Letters experiment. FEBS
letters 582(8): 1171-1177.
Darlington, J. F., Booker, J. D., 2006. Development of a
design technique for the identification of fatigue in-
itiating features. Eng Failure Anal, 13(7): 1134-1152.
Di Noia, T., Di Sciascio, E., Donini, F. M., 2007. Semantic
matchmaking as non-monotonic reasoning: a descrip-
tion logic approach. J Artificial Intelligence Research,
29: 269-307.
Gerstein, M., Seringhaus, M., Fields, S., 2007. Structured
digital abstract makes text mining easy. Nature 447:
142.
Goble, W. M., Bukowski, J. V., 2007. Development of a
mechanical component failure database. Reliability
and Maintainability Symposium, pp: 451-455.
Halaschek-Wiener, C., Kolovski, V., 2008. Syndication on
the Web using a description logic approach. J Web
Semantics, 6: 171-190.
Hatamura, Y., Nakao, M., Iino, K., 2003a. Prototyping a
failure knowledge database. Joho Shori 44(7): 733-739
(Japanese).
Hatamura, Y., Iino, K., Tsuchiya, K., Hamaguchi, T.,
2003b. Structure of failure knowledge database and
case expression. CIRP Annals – Manufacturing Tech
53(1): 97-100.
Jacobo, V. H., Ortiz, A., Cerrud, Y., Schouwenaars, R.,
2007. Hybrid expert system for the failure analysis of
mechanical elements. Eng Failure Anal 14(8): 1435-
1443.
James, M. N., 2005. Design, manufacture and materials;
their interaction and role in engineering failures. Engi-
neering Failure Analysis 12(5): 662-678.
JST, 2010. “Failure knowledge database” Available at:
http://shippai.jst.go.jp/en (accessed 1 July 2010).
Kraines, S. B., Guo, W., Kemper, B. E., Nakamura, Y.,
2006a. EKOSS: a knowledge user centered approach
to knowledge sharing, discovery, and integration on
the Semantic Web. In Proc. of the 5th Intl Semantic
Web Conf, Athens, GA, USA, November 5-9, 2006.
Kraines, S. B., Batres, R., Koyama, M., Wallace, D. R.,
Komiyama, H., 2006b. Internet-based integrated envi-
ronmental assessment: using ontologies to share com-
putational models. J. Industrial Ecology 9: 31-50.
Kraines, S.B., Koyama, M., Batres, R., 2006c. An emer-
gent simulation modeling approach for discovery of
knowledge on phenomena in chemical systems. J.
Chemical Engineering of Japan, 39: 1010-1027.
Li, L. Horrocks, I., 2003. A software framework for mat-
chmaking based on semantic web technology. In Proc.
of the 12th Intl Conf on World Wide Web, Budapest,
Hungary, pp: 331-339.
Marcondes, C. H., 2004. From scientific communication to
public knowledge: the scientific article Web published
as a knowledge base. In M. Dobreva, J. Engelen (Eds.),
Proc 8th ICC/IFIP Intl Conf on Electronic Publishing,
Leuven, Belgium, pp:119-126.
Moon, Y. B., Divers, C. K., Kim, H.-J., 1998. AEWS: an
integrated knowledge-based system with neural net-
works for reliability prediction. Computers in Industry,
35(2): 101-108
Rzhetsky, A., Seringhaus, M., Gerstein, M., 2008. Seeking
a new biology through text mining. Cell 134: 9-13.
Stone, R. B., Tumer, L. Y., Wie, M. V.,2005. The func-
tion-failure design method. J Mech Design 127(3):
397-407.
Tamura, M., 2003. Learn from failure! failure knowledge
in chemical substances and plants and its use. Chemi-
stry 58(8): 24-29 (Japanese).
Uschold, M., Clark, P., Dickey, F., Fung, C., Smith, S.,
Uczekaj, S., Wilke, M., Bechhofer, S., Horrocks, I.,
2003. A semantic infosphere. Intl Semantic Web Conf
2003, pp:882-896.
Wang, J., Jin, B., Li, J., 2004. An ontology-based pub-
lish/subscribe system. In H.-A. Jacobsen (Ed.): Mid-
dleware 2004, LNCS 3231, pp: 232-253.
Warren Liao, T., Zhan, Z-H., Mount, C. R., 1999. An inte-
grated database and expert system for failure mechan-
ism identification: Part I: automated know- ledge ac-
quisition. Eng Failure Anal 6(6): 387-406.
Wintle, J. B., Pargeter, R. J., 2005. Technical failure
investigation of welded structures (or how to get the
most out of failures). Eng Failure Anal 12(6): 1027-
1037.
SUPPORTING REUSE OF KNOWLEDGE OF FAILURES THROUGH ONTOLOGY-BASED SEMANTIC SEARCH
169