3D SPATIAL DATA MINING ON DOCUMENT SETS FOR THE

DISCOVERY OF FAILURE CAUSES IN COMPLEX TECHNICAL

DEVICES

Timo Götzelmann,

Knut Hartmann,

Andreas Nürnberger

and Thomas Strothotte

Department for Simulation and Graphics

Department for Knowledge and Language Engineering

Keywords:

Spatial data mining, visualization, association rules, discovery of error causes, product improvement.

Abstract:

The retrospective fault analysis of complex technical devices based on documents emerging in the advanced

steps of the product life cycle can reveal error sources and problems, which have not been discovered by

simulations or other test methods in the early stages of the product life cycle. This paper presents a novel

approach to support the failure analysis through (i) a semi-automatic analysis of databases containing product-

related documents in natural language (e. g., problem and error descriptions, repair and maintenance protocols,

service bills) using information retrieval and text mining techniques and (ii) an interactive exploration of the

data mining results. Our system supports visual data mining by mapping the results of analyzing failure-related

documents onto corresponding 3D models. Thus, visualization of statistics about failure sources can reveal

problem sources resulting from problematic spatial conﬁgurations.

1 INTRODUCTION

Our approach to retrospectively analyze error sources

of complex technical devices has been inspired by

statistical analysis methods on patient data sheets

used in medical science and pharmaceutics in order

to detect correlations between potential causes and

diseases. Analogously, within the product life cycle

of complex technical devices, a multitude of different

documents emerge, which are often directly related

to errors or malfunctions. The statistical analysis

of comprehensive databases containing documents

in natural language which emerged throughout the

product life cycle can help to discover correlations

between conditions causing malfunctions and thus

help to ﬁnd knowledge which is already present, but

hidden in the data. Due to the large number of

correlations and hypotheses which could be derived

through statistical methods, powerful visualization

techniques can assist domain experts in extracting

relevant data from complex data sets.

If these statistical data can be linked to a spatial

context, visualizations which integrate abstract data

into maps may reveal correlations even more easily

to a human observer than it is possible with abstract

statistical diagrams. A famous example of the power

of that type of visualization is the work of Dr. John

SNOW (a British physician and one of the fathers

of epidemiology), who detected the cause of the

great cholera epidemic in the year 1854, when many

thousands citizens died in London (Tufte, 1997). By

associating deaths statistics with addresses, where the

victims lived in, and their visualization on a map,

his graphics were able to communicate the source

of this disease — centrally located pumps incubated

with germs. Based on this idea meanwhile several

approaches have been proposed to map statistical

information based on geographic metadata. This

includes a broad variety of data and applications

(Bernhardsen, 2002), from studies in social sciences

(Orfor et al., 1997) to analysis and design of telephone

networks (Flavin and Totton, 1996; Schmidt, 1998).

Different to these approaches, that are based

usually on simple mappings based on geographic

metadata, the aim of the approach proposed in

this paper is to analyze a set of documents which

are describing malfunctions and errors for complex

technical devices, followed by a projection of the

results on a corresponding 3D model. Domain experts

can evaluate the results gained by the automatic

analysis of documents in natural language with a set

of standard data mining methods by exploring an en-

137

Götzelmann T., Hartmann K., Nürnberger A. and Strothotte T. (2007).

3D SPATIAL DATA MINING ON DOCUMENT SETS FOR THE DISCOVERY OF FAILURE CAUSES IN COMPLEX TECHNICAL DEVICES.

In Proceedings of the Second International Conference on Computer Graphics Theory and Applications - AS/IE, pages 137-145

DOI: 10.5220/0002082101370145

 SciTePress

riched 3D model interactively in order to ﬁnd spatial

relationships between the different components of the

product.

This paper is organized as follows: After sketch-

ing an application scenario and considering some

basic questions and motivations of our approach, we

discuss the related work which inspired our work.

The next section deﬁnes the data basis and discusses

the individual steps of our visual data mining ap-

proach. Then we present some results which could

be achieved with our prototypical system. Finally, we

summarize the main assumptions and contributions of

our paper and provide ideas for future research.

2 CONTINUOUS FAILURE

ANALYSIS

The project which is presented in this paper was

driven by several questions:

1. Why are information retrieval and data mining

techniques useful in the proposed application

scenario?

2. Which visualization techniques are appropriate

for the spatial mapping of the results of informa-

tion retrieval and data mining techniques?

3. How to evaluate visual data mining techniques

based on large product-related document sets?

Our approach aims at a continuous evaluation of

failure causes over the full life cycle of a complex

technical device, i. e., in design, manufacturing, qual-

ity assurance, marketing, maintenance, and repair. In

all these stages a multitude of documents emerge,

some of them describing geometric aspects such as

3D models. One important aspect of the design

stage is to detect failure causes in order to minimize

the overall manufacturing, maintenance, and support

costs. Frequently, extensive simulations are run to test

important features and properties of a product, while

Figure 1: FEM crash-simulation of a virtual car (Source:

(Zienkiewicz and Taylor, 2000)).

powerful visualizations are needed to extract impor-

tant features from vast simulation datasets (Fig. 1

presents a visualization of a simulation of arising

deformations on a car using ﬁnite-element methods

FEM).

In order to discover a broad variety of failure

causes, our approach aims at extending failure tests

from the design stage to the entire product life cycle

by analyzing failure-related documents. When com-

plex technical devices are in use, they are typically

maintained and repaired in a continuous fashion. The

dealer’s workshops store documents about service

protocols and bills of the actions done with the

individual components of the products. While the

automatic analysis of information stored in structured

databases is well studied and broad amount of meth-

ods are available (Hand et al., 2001; Hipp et al., 2002;

Berthold and Hand, 2003), the automatic extraction of

a formal representation for the knowledge contained

in unstructured text documents is still beyond the cur-

rent state-of-the-art in natural language processing;

a manual analysis would be impossible due to the

pure amount of documents. Therefore, we employ

a combination of information retrieval and statistical

data analysis (text mining) techniques in order to

gather useful and statistically relevant information

that can then be used for visualization.

Semantically segmented 3D models, i. e., geo-

metric models where the individual geometric com-

ponents are associated with their denotation and

descriptive texts, enable a ﬂexible spatial mapping

of results of statistical analysis. The spatial data

mining approach, i. e., the visualization of statistical

data on their spatial reference object by modifying

visual properties to encode data, can reveal a-priori

unknown facts, which were hidden in the database.

By interactively exploring the enriched 3D model,

unknown sources and correlations of failures can be

discovered that rely on the spatial conﬁguration of

Figure 2: Overlapping development with former product

versions’ life cycles.

GRAPP 2007 - International Conference on Computer Graphics Theory and Applications

138

several components and the shape of complex formed

geometric objects.

As the product life cycles of the products in a

product family commonly overlap. The life cycle

starts with a developing stage (see Fig. 2-A and -

B). Often similar products are in the stage of selling

and servicing them while new products are already

in the developing stage (see Fig. 2-B). Hence, a

retrospective analysis of weak spots of on-market

products can help to improve new ones.

In the next section we brieﬂy review the ter-

minology and the (visual and spatial) data mining

techniques that are closely related to our work.

The description of the architecture of our prototype

includes some preliminary answers to the second

question raised at the beginning of this section. The

discussion of the third challenge — an evaluation of

our approach — is given at the end of the paper.

3 RELATED WORK

TUFTE (Tufte, 1997, chap. 2) coined the term In-

formation Graphics for visualizations aiming at an

improved understanding of statistics and to detect

correlations in the statistical data. From TUFTE’s

point of view, SNOW’s cholera map and other hand-

made illustrations are ideal examples of information

graphics. But due to their static nature, neither

their style or other parameters of the presentation

nor their content can be adjusted to explore further

aspects of the underlying data in an interactive modus.

Hence, several research directions combine automatic

analysis techniques of large data sets with dynamic,

interactive visualizations:

The research in Data Mining (see for instance

(Witten and Frank, 1999; Hand et al., 2001; Berthold

and Hand, 2003; Tan et al., 2005)) is focused on

the non-trivial extraction of potentially signiﬁcant

relationships and regularities which are implicit but

hidden in large databases (this aspect is focused

by another term Knowledge-Discovery in Databases

(Frawley et al., 1992)).

Humans have a great ability to recognize patterns.

Therefore, Visual Data Mining systems enable do-

main experts to adjust and control the data mining

process. The preprocessed data sets are visualized in

abstracted 2D or 3D graphics and can be interactively

explored. These visualizations can also be considered

as information graphics. Visual data mining uses

information visualization techniques to generate ab-

stract views of the preprocessed data (Keim, 2002).

However, external spatial information is neither taken

into account for determining association rules nor for

generating the visualizations.

Spatial Data Mining (see for instance (Ester et al.,

2000)) is a special type of visual data mining which

exploits spatial data contained in graphical informa-

tion systems (GIS). These systems generate maps that

contain additional data. The integration of statistic

data into a spatial context can help viewers to ﬁnd new

relationships and rules. Note, that the cholera map

is a perfect example for these kind of visualizations.

Like in visual data mining, the visual capabilities

of domain experts to detect pattern are exploited by

combining the graphical and computational power of

interactive computer systems. The term Geographic

Data Mining (Miller and Han, 2001) is used for

systems, where the spatial context is restricted to GIS

data, i. e., two-dimensional spatial data.

The approach we propose in this paper naturally

extends the idea of spatial data mining: We project

abstract data related to real-world objects onto cor-

responding three-dimensional models. Subsequently,

the resulting visualization can be interactively ex-

plored by domain experts. Due to its characteristics

we call this approach 3D Spatial Data Mining.

4 A SPATIAL DATA MINING

ARCHITECTURE

The architecture of our spatial data mining approach

is presented in Fig. 3. Starting from a corpus of

documents, which are analyzed automatically and a

3D model serving as a corresponding spatial context,

our architecture comprises three sequential steps.

First, a domain expert restricts the analysis to a set

of relevant documents (data selection). The data

analysis comprises several text mining techniques

and an analysis of spatial relations between error

sources in the 3D model. Finally, the results

are visualized by adapting the visual properties of

corresponding geometric component in the 3D model

(data visualization). An interactive exploration of this

enriched 3D model enables domain experts to modify

all parameters of the data mining pipeline in order to

detect unknown failure sources. The next subsections

describe the individual components of the spatial data

mining pipeline in more detail.

4.1 Preprocessing the Corpus of

Documents

The data mining process is based on a set of

documents D (e. g., service reports, bills, problem

3D SPATIAL DATA MINING ON DOCUMENT SETS FOR THE DISCOVERY OF FAILURE CAUSES IN COMPLEX

TECHNICAL DEVICES

139

Figure 3: The process of 3D spatial data mining.

descriptions, . . . ). Each of the documents contain text

in natural language. We assume that the documents’

terms referring to individual components of the

technical device correspond to those terms used in the

semantic annotation of the individual component of

the 3D model (see Fig. 4). All ﬁll-words and words

which are not used in the spatial context (i. e., in the

annotation of the 3D model M) are not considered

in the approach presented in this paper. Moreover,

our approach does not employ a syntactic or semantic

analysis — a restriction which is common in text data

mining (Hotho et al., 2005).

4.2 Data Selection

Domain experts should be able to control the scope

of the data analysis. Therefore, optionally they can

select a subset of documents, reﬂecting their point

of interest. In spatial data mining, it is usual to

utilize relational databases to perform SQL queries

on it. But, since our analysis is based on an

unstructured corpus of documents in natural language

we use information retrieval techniques to select only

relevant documents for this task based on a given

query.

For an efﬁcient representation of a large corpus

of documents, we employ the standard vector space

Figure 4: Terms used in documents correspond to annotated

components of 3D model.

model (Salton et al., 1975), where both the query

q and the documents d ∈ D are transformed into

a vector representation

d and ~q respectively. Our

initial index T contains terms used in the document

set D which are also contained in the semantic

annotation of the associated spatial context M, i. e.,

{

T | t ∈ D ∧ t ∈ M

}

. Based on these index terms we

compute weighted word vectors for each document

and the query as described in (Salton et al., 1994): For

each term t a weight w

is computed that describes its

importance for the description of the document d.

= tf

· log(N/n

), (1)

where N is the size of the document collection D, n

the number of documents in D that contain term t and

deﬁnes how often the term t occurs in document

d. Based on these weights a vector is deﬁned for

each document:

d = (w

.. .,w

) , where n is the

number of terms in T . The similarity of the query

and a document vector is ﬁnally computed based on

the inner product of both vectors (commonly called

cosine similarity):

sim(

d,~q) =

d ·~q

|d| ×

|q|

∑

t∈T

× w

∑

t∈T

)

∑

t∈T

)

The resulting subset are all documents with a deﬁned

minimum relevance rel

min

result

| sim(

d,~q) > rel

min

∧ d ∈ D

The subset of documents chosen by this optional

selection step are subsequently considered in the data

analysis steps.

4.3 Data Analysis

The analysis of the documents follows SHNEIDER-

MAN’s Information Seeking Mantra: Overview ﬁrst,

zoom and ﬁlter, then details on demand (Shneider-

man, 1996). In this paradigm domain experts can

GRAPP 2007 - International Conference on Computer Graphics Theory and Applications

140

switch between different perspectives: (i) an overview

about the results of analyzing the selected subset of

documents, (ii) the restriction of the presentation to

selected aspects in order to ﬁnd correlations, and (iii)

and in-depth inspection of the spatial conﬁguration in

the 3D model.

To decouple the analysis from the visualization

system, we ﬁrst deﬁne a relevance vector

~r which represents the relevance of each

component c of the 3D model M, where

{

~r | c ∈ M ∧ relevance(c) ∈ [0..1]

}

Overview. To provide an overview of the subset

of documents onto the 3D model we sum up the

overall term frequencies and normalize them as

follows:

overview

∑

d∈D

∑

d∈D

Zoom and Filter. The underlying data are

analyzed by employing standard data mining methods

in order to ﬁnd association rules — a common

technique in data mining. There are several different

approaches for rule-ﬁnding. In this step a set of

association rules can be selected and inspected while

interacting with the enriched 3D model.

Detail on Demand. Each individual~r found with

the exchangeable standard text mining approach or

determined by spatial mining, can be selected by the

user during the exploration of the 3D model.

4.3.1 Text Mining

Besides the visualization of simple frequency sta-

tistics of names or annotations of model parts that

are mentioned in the text documents, especially

information about frequently co-occuring names or

annotations might provide strong indications about

reasons for system faults that are caused by a com-

bination of faults on speciﬁc parts. In order to detect

these frequently occuring names or tags we decided

to use association rule learning methods.

There are several approaches for mining rules in

sets of text documents (transactions). The two best-

known basic algorithms for mining association rules

are Apriori (Agrawal et al., 1993) and Eclat (Zaki

et al., 1997). In our approach we apply the Apriori

implementation of (Borgelt, 2003) for determination

of rules for our text collection~r

text-rule

The input for the association rule learner are

lists containing the terms used in each document

d ∈ D (the so-called item sets). The ﬁrst step of

the association rule learning algorithm determines

frequent itemsets, i.e. it extracts sets of terms (items)

that frequently occur together in the documents.

The required minimal frequency with which the

items must occur together in order to be selected as

’frequent’ is deﬁned by the support value. In a second

step association rules are generated for which a pre-

deﬁned conﬁdence, i.e. the frequency with which the

rules are supported by the documents, holds. Thus,

we ﬁnally obtain rules of the form:

← t

[,t

,. . .](Conﬁdence x%,Support y%)).

The list of derived association rules is ﬁnally pre-

sented to domain experts in an interactive 3D browser

which allows to select relevant rules and adapts

visualization accordingly.

For example, if in fault protocols of machineries

three parts are frequently mentioned together with a

speciﬁc fault, the association rule learner will propose

– among other rules – an association rule that depicts

these parts together with the cause (ideally, with the

cause as consequent of the rule). Depending on the

3D browser conﬁguration these three parts might be

automatically highlighted and thus providing visual

information about this detected dependency to the

user.

4.3.2 Spatial Mining

3D models are geometric approximations of objects

in the real world. We assume that we can use those

3D models that have been created in the development

stage of the product’s life cycle. Therefore, the spatial

relations between components in the real product

can be analyzed in these 3D models. Although the

discovery task is done primarily by domain experts

with the enriched 3D model, our approach determines

suggestions of the failure causes based on the asso-

ciation rules found by text mining. The following

steps are applied on all association rules~r

text-rule

and on

the overview vector ~r

overview

. The geometric analysis

emphasizes objects, which are close to the center of

spatial accumulations.

The cholera map mentioned in the introduction

motivates a heuristic to detect unknown error sources

by revealing clusters and agglomerations of spatial

related errors. These potential failure sources might

not be reported in the maintenance documents. Thus,

they might not be discovered purely by text mining

techniques. We use that insight and transfer it to

our approach by determining the weighted centroid

cand of the faulty components C. To do so, initially

the bounding box centers

b of the components of the

3D model are determined (see Fig. 5). Finally, the

component’s relevance is considered in a weighted

centroid:

cand =

∑

i∈C

3D SPATIAL DATA MINING ON DOCUMENT SETS FOR THE DISCOVERY OF FAILURE CAUSES IN COMPLEX

TECHNICAL DEVICES

141

Figure 5: Determining the object nearest to the weighted

centroid.

Next, the component which is nearest to the weighted

centroid is determined:

min

i∈M



∑

j=1

(cand

− b

i, j

)



Finally, a relevance vector ~r

spatial-rule

is constructed

where the component determined is emphasized.

Of course, this heuristic is only applicable for

models where a spatial neighborhood might be re-

sponsible for a fault. However, in complex machiner-

ies a fault might also be frequently caused by parts

that are spatially far away but technically connected,

e.g., by pipes, wires or transmission systems.

4.4 Data Visualization

There are several constraints for an ideal visualization

of relevance values associated with graphical objects:

First, the most relevant objects should be detected

easily, i. e., they have to be visible and should have

a minimal contrast to the background. Therefore, we

employ attentive or pre-attentive mechanisms to focus

the attention of the viewer onto the most relevant

objects.

Secondly, we should provide indications for the

relevance values for all objects, as (i) an identiﬁcation

of salient objects relies on the identiﬁcation of contex-

tual objects, (ii) as the relevance values are based on

heuristics, and (iii) as our visualizations should assist

domain experts in detecting failures due to spatial

conﬁgurations. Graphical emphasis techniques must

not alter spatial conﬁgurations or shapes of objects.

Thus, graphical abstraction techniques are applied in

an importance driven presentation of the object.

There are a couple of graphical emphasis tech-

niques which can convey the relevance or salience of

geometric components in illustrations (the broad va-

riety of illustration techniques in technical documen-

tations and scientiﬁc textbooks already inspired the

research on non-photorealistic rendering (Gooch and

Gooch, 2001; Strothotte and Schlechtweg, 2002)).

We evaluated a number of graphical emphasis tech-

niques with respect to our requirements (see Tab. 1).

Table 1: A qualitative evaluation of graphical emphasis

techniques for 3D models.

Method Conﬁg. Shape Visible

Color Attributes original original no

View Adaption original original no

Transparency original original yes

Cutaway altered altered yes

Simpliﬁcation original altered (yes)

Size altered altered (yes)

The ﬁrst dimension color attributes comprises

changes applied to material attributes like color,

brightness, saturation, and reﬂectance. Unfortunately,

this emphasis technique as well as the view adaption

on the model does not ensure, that unimportant

components do not hide important ones. Carefully

adjusted transparency values for irrelevant objects, or

the application of abstraction techniques that alter the

geometry like cutaway views, polygon simpliﬁcations

and deformations of the 3D models can commonly

avoid that problem. But as the latter techniques

either alter the spatial conﬁguration or the shape

of geometric entities they are not adequate for our

approach. Thus, modifying the transparency of each

individual component, according the their importance

seems to be optimal.

To control the transparency, we

use the relevance vector~r which is determined by text

or spatial mining techniques.

4.5 Data Exploration

The detected association rules are presented to the

user, who explores the enriched 3D model interac-

tively. The user can decide if the model has to be

enriched according ~r

overview

or any ~r

rule

. Additionally,

the interactive 3D browser should offer the user to

select several transfer functions H(~r) (e. g., linear,

logarithmic) of the importance value to the compo-

nents’ transparency. For the sake of modularity the

transfer functions have to guarantee that the resulting

values are normalized, e. g.,

H(~r) =











~r, linear

log(10(~r + 1)), logarithmic

... where H(~r) ∈ [0..1].

However, (Viola et al., 2004) and (Diepstraten et al.,

2002) more extensively studied transparency techniques.

GRAPP 2007 - International Conference on Computer Graphics Theory and Applications

142

5 RESULTS AND DISCUSSION

We developed an experimental application of the pre-

sented framework, using Coin3D for the interaction

with the enriched 3D model and Qt for the graphical

user interface. For our approach, we generated test

data sets of documents, which contained unequal

portions of the terms used in the assigned 3D models.

In our application the user was able (i) to make a

selection on the data set of documents via deﬁning

an IR query. The resulting subset was (ii) analyzed

by the Apriori (Agrawal et al., 1993) algorithm,

utilizing the user deﬁned parameters #association

terms, minimum support and minimum conﬁdence.

The analysis of ~r

overview

was (iii) visualized in the

corresponding components of the 3D model using the

chosen transfer function. Finally, the user was able to

(iv) interactively explore the enriched 3D model and

select each relevance vector ~r

text−rule

and ~r

spatial −rule

of the previously determined association rules, which

adapted the appearance of the 3D model (see Fig. 6).

As all construction and failure-related data of

commercial products are highly conﬁdential, it is

unlikely to get real data sets (e. g., 3D-models, failure

reports, and service bills) or to get a permission to

publish possible results. But discussions with several

industrial manufactures supported the basic assump-

tion of our approach — that there is a need of visual

data mining techniques for failure-related documents.

Thus, we developed a tool, which generates artiﬁcial

test documents sets, containing predeﬁned terms with

user chosen frequencies with an unequal distribution

over the resulting documents, to approximate real

documents. These controlled document sets were

used for the experimental application to evaluate our

user studies.

The ﬁrst test was performed in order to evaluate

Figure 6: Interactive explorer with relevance vectors~r

rule

how suitable transparency is to encode object-related

statistical data and how many levels users can identify

while interactively exploring an enriched 3D model.

We used 5 document sets with different term distrib-

utions. The geometric test conﬁguration consisted of

27 spheres, which were spatially arranged to a cube

with the edge length of 3 spheres. The corresponding

terms used in the document sets were ’sphere1’ to

’sphere27’. The documents were analyzed accord-

ing to term frequencies, which were considered as

relevance values; a linear transfer function was used

to map them onto the transparency values of the

corresponding spheres. In the test application were

sliders for each of the spheres on the right side of

the screen. The users had to explore the test object

interactively and assign the individual values for the

transparencies of the spheres onto the corresponding

sliders.

In this test, there were 5 levels of transparency

to distinguish. For each of the test runs the term

frequencies (transparency values) were chosen ran-

domly and written to a ﬁle. Accordingly, the user

selected values were logged as well, so that it was

possible to compare the actual and assumed values.

It consisted of 5 test runs, which the 21 participants

had to solve. Student’s t-test signiﬁcantly showed

(F=374.368, p<0.001) that most of the users correctly

recognized the correct importance of the objects (see

Fig. 7). Although transparency seems to be an

adequate abstraction technique, the statistic reveals

that some users may have difﬁculties differentiating

some of the values (min/max). In that case, offering

several transfer functions could reduce that problem.

We performed a second test in order to test the

user’s ability to recognize spatial relationships in a

3D context. The test was oriented on the paradigm

of the cholera map. The users were given a set of

27 spheres, which were arranged in the same fashion

like in the ﬁrst test. A certain number of these spheres

were opaque, while the remaining spheres were 90%

Figure 7: Results of ﬁrst user test on transparency.

3D SPATIAL DATA MINING ON DOCUMENT SETS FOR THE DISCOVERY OF FAILURE CAUSES IN COMPLEX

TECHNICAL DEVICES

143

transparent. The users were asked to ﬁnd out, which

of the transparent spheres was spatially most related

with all of the opaque ones.

Table 2: Times for solving tests.

Group 1 2 3 4

A 31.7s 25.5s 8.9s 11.7s

B 50.0s 48.0s 16.1s 73.7s

Both the fault rate and the time used were

measured. Additionally, we distinguished between

participants with experience (Group A: 15 persons)

in 3D applications and those without (Group B:

6 persons). Most of the users solved the 4 test

runs correctly (∼95%), but there were signiﬁcant

(F=11.916, p<0.001) time differences between group

A and B (see Tab. 2). Since the target group of this

approach are construction engineers, this should not

be a problem.

Figure 8: Exploring faults projected onto a car model.

6 CONCLUSION

This paper introduced (i) a novel approach which

maps analysis results from document sets on corre-

spondingly annotated 3D models. We (ii) suggested

a framework, which separates the analysis and visu-

alization to be ﬂexible enough to be interchangeable

with other analysis methods. We (iii) developed an

experimental application of this approach, generated

test data sets, and (iv) applied a user study, to evaluate

our approach and the methods used (see Fig. 8).

The most challenging problem of our — and of

all other approaches based on text analysis — is to

cope with the ambiguity of natural language and the

huge amount of domain speciﬁc and common-sense

knowledge required to analyze texts. Maintenance

documents normally contain only the correct desig-

nations of components, hence if the 3D model is

annotated correspondingly, there will be no problem

to establish the links between the terms contained

in documents an the geometric components of the

3D model. But HARTMANN’s text illustration system

(Hartmann et al., 2002; Hartmann and Strothotte,

2002) has shown, that even shallow morphological,

syntactical, and semantical analysis can improve the

robustness of the text analysis. Further on, thesauri

and word taxonomies (e. g., WordNet (Fellbaum,

1998)) could be used to get more general results.

Additional meta-data can be exploited to analyze

temporal changes. Bills and service reports com-

monly include dates and product numbers. This

information has to be extracted by specialized pars-

ing algorithms and should be associated with the

document vector. Chronological ordered data sets

would be able to reveal temporal association rules. In

combination with additional geometric information,

especially the connectivity between the individual

components, the propagation of faults can be traced

(e. g., one component perishes and the neighboring

components are negatively affected by it).

Another challenge is the validation of simulation

data with failure reports extracted from many prod-

ucts over a long period of time and the application of

our approach to other geometric representations (e. g.,

voxel models) and domains.

(i) Our visual data mining pipeline is designed in

a modular fashion and we plan to integrate a volume

renderer as a visualization component. By comparing

FEM simulations from the design phase with the

retrospectively enriched voxel models, a target/actual

comparison with approach could help to adjust the

parameters for future FEM studies.

(ii) Another interesting application domain is

medicine and pharmacy, where unique (Latin) terms

are used as denotations of for all domain entities

(organs, muscles, and bones). There also exist

polygonal (e. g., viewpoint catalog) and voxel models

(e. g., visible human data set) which approximate the

human body with its single parts and are annotated

GRAPP 2007 - International Conference on Computer Graphics Theory and Applications

144

with their denotation. As these terms are almost

unambiguous, data mining techniques can be applied

on medical or pathological ﬁndings (e. g., (McDonald

et al., 1998; Moore and Berman, 2000)) to ground

spatial related correlations between the records in

the database. As an example, it could be revealed

that an artiﬁcial bone often negatively affected other

organs, muscles or bones; or e. g., if there is a

spatial relation between 2 organs, which often suffer

damages together.

REFERENCES

Agrawal, R., Imielienski, T., and Swami, A. (1993). Mining

Association Rules between Sets of Items in Large

Databases. In Conf. on Management of Data, pages

207–216.

Bernhardsen, T. (2002). Geographic Information Systems.

John Wiley and Sons.

Berthold, M. R. and Hand, D. J., editors (2003). Intelligent

Data Analysis, An Introduction. Springer, 2nd edition.

Borgelt, C. (2003). Efﬁcient Implementations of Apriori

and Eclat. In 1st WS of Frequent Item Set Mining

Implementations (FIMI).

Diepstraten, J., Weiskopf, D., and Ertl, T. (2002). Trans-

parency in interactive technical illustrations.

Ester, M., Frommelt, A., Kriegel, H.-P., and Sander, J.

(2000). Spatial Data Mining: Database Primitives,

Algorithms and Efﬁcient DBMS Support. Data

Mining Knowledge Discovery, 4(2/3):193–216.

Fellbaum, C., editor (1998). WordNet: An Electronic

Lexical Database. MIT Press, Cambridge, MA.

Flavin, P. G. and Totton, K. A. (1996 ). Computer

Aided Decision Support in Telecommunications. BT

Telecomm. Series. Chapman Hall Publishers, London.

Frawley, W. J., Piatetsky-Shapiro, G., and Matheus, C. J.

(1992). Knowledge Discovery in Databases: An

Overview. AI Magazine, 13(3):57–70.

Gooch, B. and Gooch, A. (2001). Non-Photorealistic

Rendering. A. K. Pteres, Natick.

Hand, D., Mannila, H., and Smyth, P. (2001). Principles of

Data Mining. MIT Press, Cambridge.

Hartmann, K., Schlechtweg, S., Helbing, R., and Strothotte,

T. (2002). Knowledge-Supported Graphical Illustra-

tion of Texts. In Int. Working Conf. on Advanced

Visual Interfaces, pages 300–307.

Hartmann, K. and Strothotte, T. (2002). A Spreading

Activation Approach to Text Illustration. In 2nd Int.

Symp. on Smart Graphics, pages 39–46.

Hipp, J., Güntzer, U., and Nakhaeizadeh, G. (2002). Data

Mining of Association Rules and the Process of

Knowledge Discovery in Databases. In Industrial

Conference on Data Mining, pages 15–36.

Hotho, A., Nürnberger, A., and Paaß, G. (2005). A Brief

Survey of Text Mining. GLDV-Journal for Comp.

Linguistics and Language Technology, 20(1):19–62.

Keim, D. A. (2002). Information Visualization and Visual

Data Mining. IEEE Transactions on Visualization and

Compututer Graphics, 7(1):100–107.

McDonald, J. M., Brossette, S., and Moser, S. A.

(1998). Pathology Information Systems: Data Mining

Leads to Knowledge Discovery. Archives of Pathol-

ogy & Laboratory Medicine in Health & Fitness,

122(5):409–411.

Miller, H. J. and Han, J., editors (2001). Geographic Data

Mining and Knowledge Discovery. Taylor & Francis,

London.

Moore, G. W. and Berman, J. J. (2000). Anatomic

Pathology Data Mining. In Cios, K. J., editor, Medical

Data Mining and Knowledge Discovery, pages 61–

107. Springer Verlag, Berlin.

Orfor, S., Dorling, D., and Harris, R. (1997). Review of

Visualization in the Social Sciences: A State of the

Art Survey and Report. Technical report, School of

Geographical Sciences, University of Bristol, UK.

Salton, G., Allan, J., Buckley, C., and Singhal, A.

(1994). Automatic Analysis, Theme Generation, and

Summarization of Machine-Readable Texts. Science,

264:1421–1426.

Salton, G., Wong, A., and Yang, C. S. (1975). A Vector

Space Model for Automatic Indexing. Communica-

tions of the ACM, 18(11):613–620.

Schmidt, V. S. (1998). Turning Telecommunication Data

into Accessible Information with SDE and Data

Visualization Tools. In ESRI Europ. User Group Conf.

Shneiderman, B. (1996). The Eyes Have It: A Task by

Data Type Taxonomy for Information Visualization.

In IEEE Symp. on Visual Languages, pages 336–343.

Strothotte, T. and Schlechtweg, S. (2002). Non-

Photorealistic Computer Graphics: Modeling, Ren-

dering, and Animation. Morgan Kaufman, Los Altos.

Tan, P.-N., Steinbach, M., and Kumar, V. (2005). Introduc-

tion to Data Mining. Addison-Wesley.

Tufte, E. R. (1997). Visual Explanations: Images and

Quantitatives, Evidence and Narrative. Graphics

Press, Cheshire, Connecticut.

Viola, I., Kanitsar, A., and Gröller, M. E. (2004).

Importance-driven volume rendering. In Proceedings

of IEEE Visualization’04, pages 139–145.

Witten, I. H. and Frank, E. (1999). Data Mining: Practical

Machine Learning Tools and Techniques with Java

Implementations. Morgan Kaufmann.

Zaki, M., Parthasarathy, S., Ogihara, M., and Li, W. (1997).

New Algorithms for Fast Discovery of Association

Rules. In 3rd Int. Conf. on Knowledge Discovery and

Data Mining (KDD), pages 283–296.

Zienkiewicz, O. C. and Taylor, R. L. (2000). The

Finite Element Method. Vol 2: Solid Mechanics.

Butterworth-Heinemann, Oxford, 5th edition.

3D SPATIAL DATA MINING ON DOCUMENT SETS FOR THE DISCOVERY OF FAILURE CAUSES IN COMPLEX

TECHNICAL DEVICES

145