Semi-automated Ontology Population from Building Construction

Drawings

Polina H

afner, Victor H

afner, Hendro Wicaksono and Jivka Ovtcharova

Institute of Information Management in Engineering, Karlsruhe Institute of Technology, Zirkel 2, Karlsruhe, Germany

Keywords:

Ontology Population, CAD Construction Drawings, Pattern Recognition, Intelligent Building.

Abstract:

Ontologies have been applied as knowledge representation in different domains, including intelligent building

management. One of the challenges in using ontologies is the population with building speciﬁc information,

such as the building elements and the energy consuming devices. The population usually has to be done man-

ually by analysis and interpreting the building drawings, thus it requires extensive work. This is due to the

lack of semantic information in the existing building construction drawings, which only contain geometrical

information. However, it is possible to understand the semantics of the drawings, if the knowledge in interpret-

ing the semantics of the symbols, shapes and other geometric information is present. This paper introduces

a tool to extract the semantic information from CAD drawings and populate the ontology using the extracted

semantic information in a semi-automatic way. The drawing primitives from CAD ﬁles are used to perform

the pattern matching and classiﬁcation algorithms to extract the semantic information. The resulting semantic

information is then mapped to the corresponding ontology classes of a T-Box ontology. Finally individuals

of the corresponding classes are created to populate the ontology and their geometric properties like world

coordinate position and bounding box are set.

1 INTRODUCTION

Buildings are becoming more intelligent. There are

various reasons for this new trend, like energy efﬁ-

ciency, comfort, and rising complexity of multime-

dia devices and intelligent furnitures. Another view

on this is the web of objects (Atzori et al., 2010).

Whereas the web of objects describes an object ori-

ented web of intelligent agents, this work focuses on

the central intelligence of the building. The core com-

ponent of an artiﬁcial intelligence is a knowledge base

containing semantic description of the domain and al-

lowing the interaction with the intelligent agents. An

ontology is a common representation to describe the

semantics. It contains the information and concepts

needed for the intelligent building management.

The need for more semantic information has also

been addressed in the development of the Building In-

formation Model (BIM) (Howard and Bj

ork, 2008)

and the introduction of BIM standards like IFC and

gbXML. The way to create the CAD content goes

through a paradigm shift from drawing to conﬁguring

and this facilitates the collection of the semantic in-

formation. The reason one can not take advantage of

those new features now is that they will take effect at

least in the next several years because of their low use

(Laakso and Kiviniemi, 2012). Until then the present

data of recent buildings mostly consist of construction

drawings based on 2D geometric primitives.

Our work addresses the extraction of semantic in-

formations from 2D drawings with the goal to cre-

ate an ontology represented knowledge base for in-

telligent buildings. First one needs to deﬁne differ-

ent classes and relation deﬁnitions for building man-

agement domains, such as the classes representing

elements like room, door, window and the relations

like hasDoor, hasWindow, etc. This will result in a

T-Box ontology, which can be used as the common

structure and taxonomy for the building information

model. Then, one needs to populate the ontology with

the building speciﬁc instances like the list of rooms

with their position and dimensions, the positions of

doors and windows, and the building automation sys-

tems with all the sensors and actuators. The popula-

tion usually has to be done manually, which requires

extensive work. The building plans are usually drawn

in CAD as collections of primitives representing cer-

tain symbols. It needs pre-knowledge in order to in-

terpret the semantic of the drawings and populate the

corresponding ontology.

379

Häfner P., Häfner V., Wicaksono H. and Ovtcharova J..

Semi-automated Ontology Population from Building Construction Drawings.

DOI: 10.5220/0004626303790386

In Proceedings of the International Conference on Knowledge Engineering and Ontology Development (KEOD-2013), pages 379-386

ISBN: 978-989-8565-81-5

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

Figure 1: Possible representation of a CAD drawing (Wicaksono et al., 2012).

There are many challenges when extracting se-

mantic geometrical information directly from CAD

drawings and many software products for CAD on the

market used for drawing building layouts. These tools

have versions and variants resulting in many differ-

ences and incompatibilities between them, different

interfaces and import/export ﬁle formats. A further

challenge is the varying representation of the building

elements in such a drawing. The quality of the infor-

mation often fully depends on the person who inputs

the data (Vanlande et al., 2008). Figure 1 shows some

examples of different representation possibilities. It

differs in the degree of information density and level

of representation depth. Another problem of the au-

tomated recognition of building elements from CAD

drawings are the different languages used. Each au-

thor labels his drawing using a particular language,

depending on the country. In addition, CAD draw-

ings are often made for different perspectives of the

building so called viewports (top or front projection).

Then a recognition program has to distinguish all per-

spectives. Other kinds of CAD drawings are block

schemes for building domains such as ventilation,

heating, access controls, photovoltaic, electric circuit

etc. To populate the ontology correctly it is impor-

tant to be able to use all layouts of the building (all

ﬂoors and all perspectives) in order to avoid informa-

tion loss.

A fully automated pattern matching method to

extract semantic informations from the CAD draw-

ings is not practicable since every drawing differs in

its conventions. The proposed solution focuses on

a semi-automatic approach, enabling the user to se-

lect a building element. The object type is deﬁned

by the user and his information can then be populated

into the ontology. The user selection can be used to

ﬁlter all identical objects automatically using pattern

matching algorithms. This is especially useful for re-

curring objects like doors and furniture. Another type

of information are the zones and rooms which can

also be easily populated.

This paper presents the methodology of OntoCAD

and its potential in the ﬁeld of building ontologies.

This introduction is followed by the state of the art

in ontology population, building layouts, and pattern

matching on primitives. Then the methodology and

implementation of our solution is described. The last

part is dedicated to the conclusion and outlook.

2 STATE OF THE ART

The following section presents the state of the art and

related works in the three topics: ontology population,

building layouts and geometric pattern matching.

2.1 Ontology Population

Ontology population has been a major challenge in

constructing an ontology-based knowledge base. Lit-

erature often describes that a manual construction

of ontology individuals leads to costly and exten-

sive work. Some researchers have proposed methods

to populate ontology semi-automatically in different

domains. However, the population methods depend

strongly on the information sources.

In the semantic web domain, where the informa-

tion sources are semi-structured documents, such as

HTML and XML, the ontology is populated based on

the syntactic and semantic similarity between the on-

tology and the web tables containing terms that are

extracted from web documents. The web documents

are collected by a web crawler. The populated ontol-

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

380

ogy is used for automatic cataloging of IT products

(Song et al., 2009).

Another method for ontology population is in the

risk management domain. The method tries to popu-

late the ontology semi-automatically from fact sheet

documents using combined Natural Language Pro-

cessing (NLP) techniques. It extracts the verbs from

natural language text and matches them to the cor-

responding relations in the T-Box ontology. How-

ever, the human intervention is still needed for control

and validation (Makki et al., 2009). There are some

other works that propose ontology population meth-

ods from unstructured texts based on NLP approaches

(Vargas-Vera et al., 2007), (Maynard et al., 2009). Up

to now, there exists no method to populate ontologies

in the building management domain.

2.2 Building Layouts

CAD design tools, such as AutoCAD, ArchiCAD or

Revit are commonly used for the creation of building

layouts representing two or three dimensional draw-

ing (plans, sections, elevations). Further CAD-based

software is used to plan and model many domains

of a building, such as ventilation, heating, access

controls and photovoltaic (Krahtov et al., 2009). The

number of elements in a sketch and its complexity

may vary (Donath, 2009) (see Figure 1).

Building information modeling (BIM) is the pro-

cess of development and use of computer generated

models to simulate the lifecycle of a facility including

planning, design, construction and operation (Azhar

et al., 2008). The resulting model of BIM is a data-

rich, object-oriented, intelligent and parametric digi-

tal representation of the facility and serves as shared

knowledge resource which helps the decision mak-

ing at each stage of the facility lifecycle (Azhar et al.,

2008). With the BIM model the former 2D construc-

tion drawings are augmented with intelligent contex-

tual semantic, where objects are deﬁned in terms of

building elements and systems such as spaces, walls,

beams and columns (CRC, 2007). To achieve the BIM

concepts, an open standardized data model called IFC

(Industry Foundation Classes) for enabling interop-

erability between BIM software and containing the

semantic information of the facility has been devel-

oped. The reason one can not use the beneﬁts of IFC

for the population of building ontologies is that to-

day after almost 20 years of IFC development (since

1994) we witness the low usage in actual construction

drawings. Laakso assumes that this is caused by the

slow adoption of collaborative model-based construc-

tion processes and industry reluctance to switch over

to new IT tools (Laakso and Kiviniemi, 2012). Even

if IFC becomes more used, there will still be cases

with drawing-like data without semantic information.

2.3 Geometric Pattern Matching

To populate the building ontology in an efﬁcient way,

pattern matching algorithms are used to ﬁnd all enti-

ties like doors, windows and furniture. There are a lot

of applications where pattern matching plays an im-

portant role. These include pose determination, com-

puter aided design, robot vision and many more. This

work considers a very small subset of pattern match-

ing methods. Spatial pattern matching is the process

of ﬁnding a geometric transformation to match two

given images. Only the special case where the im-

age is a 2D vector graphic is called geometric pat-

tern matching. Moreover only exact and total pattern

matching is considered. This is the case for whole

matches with an optimal transformation. This means

that matching patterns are identical. As described by

(Hagedoorn, 2000) there are different methods for ge-

ometric pattern matching, for instance graph match-

ing and geometric hashing. Graph matching means

that the structure of a pattern is described as a graph

and matching is performed between the graphs. Ge-

ometric hashing means that the pattern as a whole is

described by a normalised description. The drawback

of geometric hashing is that it works for known pat-

terns, as a data structure (the hash table) has to be

constructed for the whole data.

This work uses a correspondence method where

patterns are matched by ﬁtting pairs of geometric

primitives. This method combines the geometric

primitives that make up the input pattern. An example

is pairing line segments, where each combination of

two line segments in the patterns must ﬁt to make a

match.

3 SOLUTION

Our methodology uses a semi-automated approach

for the extraction of semantic information from build-

ing CAD drawings using user input at different stages.

The drawings are exported from CAD design software

using the exchange format DXF. OntoCAD imports

and draws the primitives, layers and view-ports. A T-

Box ontology (see Section 3.2) is used as input, which

allows the user to choose building elements from a

building taxonomy. The user can see and use the vec-

tor based primitive representations to add semantic

information. The population process is accelerated

through the OntoCAD user interface and the pattern

Semi-automatedOntologyPopulationfromBuildingConstructionDrawings

381

Figure 2: Generic and building speciﬁc ontology.

matching algorithms to ﬁnd the similar objects. The

user has the possibility to directly validate the results

and apply necessary corrections. Each step is con-

tinuously and automatically saved to a building spe-

ciﬁc ontology containing A-Box elements. The ad-

vantage of a user-centred method is that the user can

provide meta information about the data like drawing

type (top or side view) or additional semantic infor-

mation like room name or number, supported by an

intuitive graphical user interface. The target user of

our solution can be a facility manager, building owner,

or building management system specialist.

Figure 3: Main classes of the building management ontol-

ogy that will be populated.

3.1 The Ontology Source

In our work, the building management ontology is

represented in OWL (Web Ontology Language), a

W3C speciﬁed knowledge representation language

(Smith et al., 2004). Basically there are two types

of ontologies. A generic ontology (T-Box ontology)

represents a common information model for building

energy management, which is then populated and ex-

tended with building speciﬁc information resulting in

more building speciﬁc ontologies (A-Box ontology).

It is illustrated in Figure 2.

The ontological classes as well as their attributes

and relations representing the resources needed for

the building management are created manually by ex-

perts. The ontology containing these hand-crafted el-

ements is called generic ontology. The generic on-

tology only contains the ontological classes or T-

Box components that describe the knowledge struc-

ture, deﬁnitions and terminology. It does not contain

any ontological individuals or A-Box components and

contains no building speciﬁc information. Figure 3

depicts the ontology main classes representing the dif-

ferent resources needed for the building management.

The class BuildingElement models the building struc-

tures that are observed, examined and analyzed in en-

ergy management activities. The building elements

are passive entities which have states, but do not have

capabilities to measure or to observe their own states.

The class BuildingElement and its subclasses repre-

sent the fundamentals of the BIM. It is aligned with

the domain layer in IFC2x4.

The aim of the ontology population in our work is

to create individuals for the classes and subclasses of

BuildingElement and to enrich them with geometric

properties. For this, the generic ontology must satisfy

requirements regarding the deﬁnition of data and ob-

ject properties. Those properties have to be linked to

each class. This means that a property holds informa-

tion to which classes it belongs.

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

382

Figure 4: Workﬂow of OntoCAD.

3.2 OntoCAD

OntoCAD is a tool to populate the generic ontology

with geometric information implementing the pro-

posed methodology. It relies on open libraries like

Python, OpenGL and PyGTK, which make it highly

portable. The architecture is modular and contains

the following ﬁve modules: Input interface (for the

CAD drawings and the generic ontology), data model,

graphical user interface (GUI), pattern matching and

population modules. The whole workﬂow is depicted

on Figure 4.

3.2.1 Input Interfaces

OntoCAD allows the import of 2D Layouts via the

DXF exchange format from Autodesk

. DXF export

is nowadays supported in most CAD modelling soft-

ware. OntoCAD imports only the geometric prim-

itives line, arc, circle and ellipse, other geomet-

ric primitives are reduced to supported primitives or

discarded. Other information like metadata is also

discarded. Our DXF parser is partially based on

the FreeCAD implementation, dxfReader.py

by Ed

Blake. It is essentially a state machine that constructs

the initial data hierarchy following the open DXF

speciﬁcations.

In our work, OWL is used as the language to rep-

resent the ontologies. OntoCAD supports the import

of OWL ﬁles with the RDF/XML format that repre-

sents the generic ontology. The classes represent the

taxonomy and are therefore organized in a hierarchi-

cal tree. To populate individuals with OntoCAD, it is

important to know the available classes and the data

properties for each class.

3.2.2 User Interaction

The graphical user interface consists of four main

www.autodesk.com

free-cad.sourceforge.net/SrcDocu/d7/dd1/

dxfReader 8py source.html

parts (see Figure 5). The toolbar allows to load layout

(.dxf) and ontology (.owl) ﬁles and toggle the visu-

alization of the already extracted objects and zones.

The left panel lists all layers contained in the drawing

ﬁle, each visualized with a different color. Selecting a

layer from the list will highlight it in the viewer. One

can toggle the visibility of the layer. This will also

affect the performance of the pattern matching algo-

rithm as it will ignore invisible layers. The central

viewer is quite important as the whole methodology

is based on the users experience and visual recogni-

tion of building elements. The viewer uses OpenGL

and a custom shader to render curved primitives like

arcs and circles. The user does not change the primi-

tives except when loading a CAD ﬁle, thus the vertex

buffer objects do not need to update during the pop-

ulation process. This makes the viewer very perfor-

mant and guarantees a ﬂuent navigation and interac-

tion. The navigation consists of panning and zooming

into the data and selecting groups of primitives with a

polygon selection tool.

The right side panel allows the user to manipulate

the ontology. The ﬁrst tab, “Properties”, is a map-

ping of the data properties to a function of OntoCAD.

This is necessary because the data properties are ar-

bitrary strings with no information on how to com-

pute them and cannot be automatically set. OntoCAD

functions calculate for instance geometrical data like

position, area, length, etc. A data property from the

ontology that describes a position could be ‘hasPosi-

tion’, ‘hasLocation’, ‘hasXCoordinate’, etc. Thus it

is important that the user maps those before starting

to populate individuals. This conﬁguration is done

only once, then the user can start populating the ontol-

ogy. Another tab, “Ontology”, visualizes the ontology

class tree to show all the available classes of individu-

als and to help the user to understand the structure of

the OWL ﬁle. The “Entity” tab, as seen in Figure 5,

contains a dropdown menu with all available classes,

a button to add individuals and a list of the data prop-

erties of the currently selected class. If a data property

has been mapped to “User input”, then a text entry

Semi-automatedOntologyPopulationfromBuildingConstructionDrawings

383

Figure 5: OntoCAD GUI.

will appear next to it, this allows to add additional se-

mantic information, for instance the room type, room

number and room label. When populating two cases

are distinguished:

• Zones or rooms where their perimeter is deﬁned

by the polygon selection tool. There OntoCAD

can provide the selection area, the polygon ver-

tices, or the bounding box of the selection.

• Objects or other small sets of primitives that make

up logical entities. In such a case it is often in-

teresting to search for all duplicates in the draw-

ing to populate all instances at once, for example

lamps or doors. In that case OntoCAD can pro-

vide the bounding boxes, the position, the width

or the length for every instance.

3.2.3 Data Model

The internal data model of OntoCAD consists of the

CAD layout and the ontology. The layout uses ﬁve

CAD primitives: points, line segments, circles, el-

lipses and arcs, organized by layers. When loading

a layout, the primitives are copied to a vertex buffer

object for the OpenGL viewer. The ontology consists

of classes in a hierarchical tree, two lists of data prop-

erties and object properties and a list of individuals.

3.2.4 Module for Pattern Matching

Our goal is to get as much semantic information as

possible from the building’s 2D layout, and as auto-

mated as possible. Our approach consists of group-

ing primitives into objects based on a user deﬁned se-

Figure 6: Example of accessing the similar primitives using

the similarity criterion.

lection of primitives. The user has the possibility to

use a polygon selection tool to select all the primi-

tives that belong to a logical building element. The

targeted building elements are recurring ones. This

applies mostly to furniture like tables, chairs, or other

objects that have been inserted multiple times by the

author. Our approach uses primitives to ﬁnd similar

building elements with the following steps:

• gather a set with primitives similar to the selected

ones

• compute the set of relations between the selected

primitives

• group the similar primitives into objects based on

the above set of relations

The initial step is to take a subset of all the primitives

using criteria that deﬁne two primitives as similar. Im-

portant aspects of such a criterion is the rotation and

scale invariance. For our purpose it is important to

have rotation invariance but not scale invariance. This

is based on the assumption that rotated shapes are ro-

tated identical objects but scaled shapes are likely to

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

384

be different objects. Our criteria for lines is the length,

for circles and arcs it is the radius and for ellipses the

minimum and maximum radius. To speed up the pro-

cess a map is kept in every layer where the primitives

are sorted by their similarity criterion. This allows

fast access to the primitives, a subset example is de-

picted on Figure 6. The picture shows how the data is

organized, the numbers represent the amount of prim-

itives in the layer.

To compare patterns of primitives it is important

to deﬁne how to describe such a pattern. For this pur-

pose relations between the different types of primi-

tives are introduced. The ﬁrst and most important re-

lation is the line segment to line segment relation. It

is important to have rotation invariance but no scale

invariance as for the similarity criterion. The line seg-

ment to line segment relation is the set of the four dis-

tances between the line segment endings as illustrated

on Figure 7. This deﬁnition of the relation allows

to ﬁnd matches with transformations corresponding

to euclidean isometries as it preserves the length be-

tween two points.

The last step is to group the primitives into ob-

jects. The constraint is to match the selected set of

primitives. One needs to deﬁne what is considered

being a match. First the sets need to have the same

number of primitives, and second all relations have to

exist that are also in the reference set. An example of

matching similar chairs, even rotated, can be seen on

Figure 8.

The algorithm has been highly optimized, also by

using kd-trees and nearest neighbour search. In most

cases the pattern matching performs in under 5 sec-

onds with a building layout having 10

line segments.

3.2.5 Output

The result of the population process is the building

speciﬁc ontology (A-Box ontology). OntoCAD com-

putes the values for the individuals, their data and ob-

ject properties. This includes geometric information

such as size, area or world position of objects and the

perimeter for zones. Other information are individual

semantic data like type, name, globally unique identi-

ﬁers (GUIDs) and the building elements afﬁliation to

zones. All changes are saved continuously and auto-

matically to disk (.owl ﬁle). The results can be opened

Figure 7: The line to line relation is deﬁned by the square

values of d1, d2, d3 and d4.

Figure 8: Geometric pattern matching of furniture.

and further edited with other ontology tools (for in-

stance Prot

e).

4 SUMMARY AND OUTLOOK

Our methodology enables the user to extract and con-

solidate semantic information in a consistent and ef-

ﬁcient manner from 2D building layouts to populate

an ontology. We addressed the challenge of handling

the highly unpredictable data in a generic way by in-

cluding the user to parse small geometric formations

into semantic information. The second challenge re-

lated to process big amounts of data from buildings in

an efﬁcient way. We overcame this by combining the

user input with pattern matching algorithms.

Our implementation, the OntoCAD tool, allows

the user to extract small recurring patterns from a

CAD layout with very little effort and time. Our ﬁrst

use case is to extract all building elements from a

building layout and use the resulting speciﬁc ontology

combined with SWRL rules to run ontology reason-

ing algorithms in order to infer the energy efﬁciency

states of the building elements.

Our future work will further explore the possibili-

ties and use cases of OntoCAD in different domains.

An interesting extension would be the construction of

a 3D building model from the 2D layout. The primi-

tives forming the walls have to be extruded, moreover

the position of doors, windows and furniture can be

used to augment the building with detailed 3D build-

ing models. We see a high potential of OntoCAD in

the ﬁelds of building construction, but also in factory

planning and production management.

ACKNOWLEDGEMENTS

The work introduced in this paper is partially funded

by European Union as part of the project KnoholEM.

Semi-automatedOntologyPopulationfromBuildingConstructionDrawings

385

REFERENCES

Atzori, L., Iera, A., and Morabito, G. (2010). The internet of

things: A survey. Computer Networks, 54(15):2787–

2805.

Azhar, S., Hein, M., and Sketo, B. (2008). Building infor-

mation modeling: Beneﬁts, risks andchallenges. Pro-

ceedings of the 44th ASC National Conference.

CRC (2007). Adopting BIM for facilities management :

Solutions for managing the sydney opera house. ISBN

: 978-0-9775282-2-6 Industry Publication.

Donath, D. (2009). Bauaufnahme und Planung im Bestand :

Grundlagen — Verfahren — Darstellung — Beispiele.

Vieweg+Teubner.

Hagedoorn, M. (2000). Pattern Matching Using Similarity

Measures. PhD thesis.

Howard, R. and Bj

ork, B.-C. (2008). Building information

modelling - experts’ views on standardisation and in-

dustry deployment. Adv. Eng. Inform., 22(2):271–280.

Krahtov, K., Rogalski, S., Wacker, D., and Ovtcharova,

J. (2009). A generic framework for life-cycle-

management of heterogenic building automation sys-

tems. In Proceedings to Flexible Automation and In-

telligent Manufacturing: 19th International Confer-

ence.

Laakso, M. and Kiviniemi, A. (2012). The IFC standard -

a review of history, development, and standardization.

Journal of Information Technology in Construction.

Makki, J., Alquier, A.-M., and Prince, V. (2009). Ontology

population via NLP techniques in risk management.

International Journal of Humanities and Social Sci-

ences 3.

Maynard, D., Funk, A., and Peters, W. (2009). Sprat: a tool

for automatic semantic pattern-based ontology pop-

ulation. In International Conference for Digital Li-

braries and the Semantic Web.

Smith, M., Welly, C., and McGuinness, D. (2004). OWL

web ontology language guide. W3C Recommendation.

Song, H., Park, S.-B., and Park, S.-Y. (2009). An auto-

matic ontology population with a machine learning

technique from semi-structured documents. In Inter-

national Conference on Information and Automation

ICIA, pages 534–539.

Vanlande, R., Nicolle, C., and Cruz, C. (2008). IFC and

building lifecycle management. Automation in Con-

struction, 18(1):70–78.

Vargas-Vera, M., Moreale, E., Stutt, A., Motta, E., and

Ciravegna, F. (2007). Semi-automatic ontology pop-

ulation from text. Sharman, R., Kishore, R., Ramesh,

R. Ontologies, 14:373–402.

Wicaksono, H., Aleksandrov, K., and Rogalski, S. (2012).

An intelligent system for improving energy efﬁciency

in building using ontology and building automation

systems. In Automation, pages 531– 548.

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

386