You are on page 1of 10

Advances in Engineering Software 40 (2009) 202–211

Contents lists available at ScienceDirect

Advances in Engineering Software


journal homepage: www.elsevier.com/locate/advengsoft

Rough set algorithm for crack category determination of reinforced


concrete structures
Yeong Min Kim a, Chee Kyeong Kim b,*, Jae Cheol Lee c
a
MIDAS Information Technology Company Limited, SKn Technopark Tech Center, 15th floor 190-1 Sangdaewon1-dong, Jungwon-gu, Seongnam, Gyeonggi-do, 462-721,
Republic of Korea
b
Department of Architectural Engineering, Sunmoon University, Kalsan-ri, Tangjeong-myeon, Asan-si, Chungnam, 336-708, Republic of Korea
c
Department of Architectural Engineering, TongMyong University, 535 Yongdang-dong, Nam-gu, Busan, 608-711, Republic of Korea

a r t i c l e i n f o a b s t r a c t

Article history: Rough set theory is a new mathematical approach to imprecision, vagueness and uncertainty in data
Received 24 July 2007 analysis. This paper presents a new approach to rough set theory in determining the category of crack
Received in revised form 14 February 2008 causes for insufficient and imprecise crack characteristics observed in regular inspection of concrete
Accepted 4 April 2008
structures. The categories of crack causes are classified into four classes, that is, (1) concrete material,
Available online 27 May 2008
(2) construction work, (3) service and environmental factors, and (4) structure and applied loads. The
crack characteristics include time of formation, shape, regularity, cause of concrete deformation, and
Keywords:
range. The decision table was constructed considering crack characteristics as condition attributes and
Rough set
Decision algorithm
the categories of crack causes as decision attributes. A minimal decision algorithm for this decision table
Crack characteristics was generated on the basis of rough set theory; the algorithm is equivalent to the original decision table,
Concrete Structures but requires minimum subsets of condition attributes. It turned out in determining the category of crack
causes that, ‘‘time of formation” had the most important influence among crack characteristics, and
‘‘shape” could be omitted with the least influence on diagnosis.
 2008 Elsevier Ltd. All rights reserved.

1. Introduction concrete structures. The categories of crack causes are classified


into four classes, that is, concrete material, construction work,
The maintenance activities for concrete structures are required service and environmental factors, and structure and applied loads.
to maintain and enhance the usability and durability of structures The minimum crack characteristics are required in constructing
throughout their life cycle [1]. Concrete structures are damaged the decision algorithm, and hence secure flexibility against uncer-
and deteriorated by various causes; among them cracks act as a tainty and imperfection of the investigated data, especially, if some
primary cause of decreased structural capacity, durability, water- characteristics cannot be investigated.
proofing, etc. Therefore, early diagnosis and appropriate counter- Rough set theory is a new mathematical approach to impreci-
measures play an important role in the maintenance and sion, vagueness, and uncertainty in data analysis, which was first
management of concrete structures. introduced by Pawlak in 1982 [2,3]. The problem of determining
Routine inspection of concrete structures or the early stages of the category of crack causes from the given crack characteristics
crack diagnosis only make available insufficient and incomplete is a decision making problem characterized by insufficient and
crack characteristics which are not sufficient to diagnose the incomplete information, and hence belongs to the region where
causes of cracks. Therefore, in these stages it is necessary to classify rough set theory is highly applicable.
the category of crack causes of individual structural members Fuzzy set theory has long been successfully applied in various
approximately, and then integrate these data to obtain the overall engineering field. The features of concrete cracks gathered from vi-
crack patterns of the structure. Finally, in the detailed diagnosis sual inspection are generally uncertain and ambiguous, here fuzzy
stage, the causes of cracks are investigated in detail, focusing on based data manipulation has shown good applicability [4,5]. But in
the selected crack category. the application of fuzzy set theory, there had arisen difficulty that
This paper presents a decision algorithm based on rough set if some data are unable to obtain, these data should be manipu-
theory to determine the category of crack causes, using the approx- lated in advance before fuzzy related functions, such as fuzzy infer-
imate crack characteristics observed during regular inspection of ence are performed [6].
Since the introduction of rough set theory, it has demonstrated
* Corresponding author. Tel.: +82 41 530 2321; fax: +82 41 530 2839. substantial importance in expert systems, decision making sys-
E-mail address: ckkim@sunmoon.ac.kr (C.K. Kim). tems, machine learning, inductive reasoning, pattern recognition,

0965-9978/$ - see front matter  2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.advengsoft.2008.04.002
Y.M. Kim et al. / Advances in Engineering Software 40 (2009) 202–211 203

information retrieval, and so on. Using rough set theory, Attoh- POSR ðXÞ ¼ RX ð4Þ
Okine proposed decision support systems for pavement rehabilita- NEGR ðXÞ ¼ U  RX ð5Þ
tion and maintenance [7], Kaname et al. presented a rough set
BNDR ðXÞ ¼ RX  RX ð6Þ
approach to information retrieval [8], and Furuta et al. extracted
experiential knowledge of experts from diagnostic results of the Fig. 1 illustrates the approximation of a set X by knowledge R and
danger levels of slope-failure as decision tables of a minimal deci- the corresponding positive, negative and boundary regions. Each
sion algorithm [9]. These studies showed that rough set theory small rectangle represents an equivalence class. From this figure
could be a useful tool for information analysis of data having the one can note with certainty that any element x 2 POSR(X) belongs
characteristics of imprecision, incompleteness and unobtainable. to X with respect to R, and that any element x 2 NEGR(X) does not
belong to X with respect to R. The upper approximation of a set X
is the union of the positive and boundary region, that is,
2. Rough set theory and decision making problems RX ¼POSR(X) [ BNDR(X). One cannot decide with certainty whether
or not an element x 2 BNDR(X) belongs to X. For arbitrary elements
In this section, rough set theory is reviewed and a decision table x 2 RX, one can only conclude that x possibly belongs to X [10].
– an important type of knowledge representation system in apply-
ing rough set theory – is briefly described. 2.1.3. Characterization of the imprecision of a rough set
The imprecision of a rough set is due to the existence of the
2.1. Notion of rough set theory boundary region, and the imprecision can be characterized either
numerically or topologically. Numerical characterization is defined
Rough set theory is an extension of classical set theory and be- by an accuracy measure or roughness, and topological character-
gins by relating certain information to all the objects under consid- ization is defined by the concepts of upper and lower approxima-
eration. Objects with the same characteristics cannot be tion. ‘‘Accuracy measure” indicates the degree of completeness of
distinguished from each other, and this feature of indiscernibility knowledge. The accuracy measure of X with respect to R is defined
forms the mathematical basis for rough set theory. as Eq. (7), here card means number of elements of a set.
cardRX
2.1.1. Knowledge base and the indiscernibility relation aR ðXÞ ¼ ; ð0 6 aR ðXÞ 6 1Þ ð7Þ
Let U denote a finite non-empty set, called the universe; the set cardRX
R1 = {X1, X2, . . . ,Xn} which satisfies Eq. (1) is called the classification Roughness is the opposite of accuracy measure and represents the
or equivalence relation of U, and Xi is called an equivalence class of incompleteness of the knowledge. The roughness of X with respect
the equivalence relation R1. to R can be defined using the concept of accuracy measure as
qR ðXÞ ¼ 1  aR ðXÞ; ð0 6 qR ðXÞ 6 1Þ ð8Þ
X i # U; X i –U; X i \ X j ¼ U; [X i ¼ U; ði–j; and i; j ¼ 1; . . . ; nÞ ð1Þ
A rough set can be topologically classified into the following four
The set R = {R1, R2, . . . ,Rm}, comprised of equivalence relations of U, classes:
is called the knowledge base of U, and the knowledge base K is rep-
resented as K = (U,R).  Roughly R-definable: RX–U and RX–U
If P # R and P–U; then \P (the intersection of all equivalence  Internally R-undefinable: RX ¼ U and RX–U
relations belonging to P) is also an equivalence relation, called  Externally R-undefinable: RX–U and RX ¼ U
the indiscernibility relation over P, and the notation for this rela-  Totally R-undefinable: RX ¼ U and RX ¼ U
tion is IND(P). The classification or partition of U by IND(P) is ex-
pressed as U/IND(P) or U/P, and an equivalence class of IND(P) is With the accuracy measure we can know the range of the
called a P-basic category. boundary, but not its structure, on the other hand, the topological
classification of the rough set gives us insight into the structure of
2.1.2. Approximation of a rough set the boundary, but nothing about its range. Therefore, in applying a
Let X # U, and R the equivalence relation. If X is the union set of rough set, both numerical and topological characterization need to
arbitrary R-basic categories, then X is R-definable, otherwise X is R- be considered.
undefinable. An R-definable set is called an R-exact set and an R-
undefinable set is called an R-rough set. A rough set is approxi-
mately defined by the two exact sets of lower approximation and
upper approximation.
Now consider knowledge base K = (U,R). The lower approxima-
tion of X # U by R is the union of all equivalence classes E of equiv-
alence relation R 2 IND(K) which are included in X, i.e.,

RX ¼ [fE 2 U=RjE # Xg ð2Þ

The upper approximation of X # U by R is the union of all equiva-


lence classes E of equivalence relation R 2 IND(K) which have a non-
empty intersection with X, i.e.,

RX ¼ [fE 2 U=RjE \ X–Ug ð3Þ

Based on the lower and upper approximation of X # U by R, the


universe U can be divided into three disjoint regions, that is, a posi-
tive region POSR(X), a negative region NEGR(X), and a boundary re-
gion BNDR(X), as follows:
Fig. 1. Positive, boundary and negative regions of a set X.
204 Y.M. Kim et al. / Advances in Engineering Software 40 (2009) 202–211

2.1.4. Knowledge reduction and dependency the decision table. The restriction of decision rule dx over P and Q
When we define some category of knowledge, it is not always is expressed as dxjP and dxjQ, and called the condition and decision
necessary to use all of it. The elimination of non-necessary knowl- of decision rule dx, respectively.
edge is called knowledge reduction, and here, the two concepts of If the following Eq. (14) is satisfied for every attribute x, y(x – y)
‘‘reduct” and ‘‘core” are used. Reduct is the minimal necessary part of universe U, then decision rule dx is called consistent, otherwise,
of the knowledge which is sufficient to define all basic categories inconsistent.
appearing in current knowledge, and core is the peculiar part
if dx jP ¼ dy jP then dx jQ ¼ dy jQ ð14Þ
which cannot be eliminated in knowledge reduction.
Let P and Q be equivalence relations of universe U, then the The simplification of the decision table is the process of eliminating
P-positive region of Q is defined by the union of all the objects of attributes which are not necessary in decision making. With the re-
U which can be classified as the equivalence class of U/Q by the duced decision table one can derive the same decision with fewer
knowledge U/P, i.e., condition attributes, which means that one need not consider
POSP ðQ Þ ¼ [ PX ð9Þ non-necessary conditions, which thus reduces the effort necessary
X2U=Q for decision making. Furthermore, with the reduced condition attri-
Let P and Q be equivalence relations of U. If Eq. (10) is satisfied, then butes we can organize an expert’s knowledge in more compact
C 2 P is said to be Q-dispensible in P, otherwise, C 2 P is Q-indispen- form. The simplification of the decision table is processed, first, to
sible in P. If all C are Q-indispensible in P, P is said to be independent obtain the reducts of the condition attributes by eliminating col-
with respect to Q. umns of non-necessary condition attributes, and second, to elimi-
nate the values of non-necessary condition attributes.
POSINDðPÞ ðINDðQ ÞÞ ¼ POSINDðPfCgÞ ðINDðQ ÞÞ ð10Þ

If S # P is P-independent and POSS(Q)=POSP(Q) is satisfied, then S is


said to be the Q-reduct of P, that is, REDQ(P), and the union of all the Table 1
Q-indispensible attributes is said to be the Q-core of P, that is, COR- Category of crack causes and list of crack causes corresponding to each category
EQ(P). The relation of these two notions is expressed as
Category of crack cause Crack Causes of cracks
COREQ ðPÞ ¼ \REDQ ðPÞ ð11Þ No.
Concrete material A1 Abnormal setting of cement paste
Let K = (U,R) be the knowledge base and P, Q  R. If we can induce A2 Heat of hydration of cement
knowledge Q from knowledge P then Q is said to be dependent on A3 Abnormal expansion of cement
P, and the degree of dependency k is defined as A4 Clay in aggregate
A5 Low quality aggregate
cardPOSP ðQ Þ A6 Reactive aggregate
k ¼ cP ðQ Þ ¼ ; ð0 6 k 6 1Þ ð12Þ
cardU A7 Chloride in concrete
A8 Settlement and bleeding of concrete
If k = 1 then Q is said to be totally dependent on P, if 0 < k < 1 then Q A9 Shrinkage of concrete
is said to be roughly dependent on P, and if k = 0 then Q is said to be
Construction work B1 Unequal dispersing of admixture
totally independent of P. B2 Overtime mixing
B3 Segregation during pumping
2.2. Knowledge representation system and the decision table B4 Improper placing sequence
B5 Rapid placing
B6 Insufficient compaction
A decision table is an important class of knowledge representa- B7 Vibration or loading before hardening
tion system in the application of rough set theory [7]. A decision B8 Rapid shrinkage during initial curing
table specifies what decision should be taken when certain condi- B9 Initial frost damage
tions are satisfied. Most decision making problems can be formal- B10 Improper treatment of work joint
B11 Movement of reinforcing bar
ized by a decision table, which explains the usefulness of a decision
B12 Insufficient protective covering
table in the process of decision making. Knowledge reasoning is a B13 Deformation of forms
process of finding new facts from existing facts, and here, the logic B14 Leakage (from form or ground)
for knowledge reasoning is called decision logic, the rules for B15 Early removal of forms
knowledge reasoning are called decision rules, and the sets of deci- B16 Settlement of form support

sion rules are called a decision algorithm. If all the decision rules of Service and environmental C1 Change of external temperature and
the decision algorithm are true, then the decision algorithm is said factors humidity
C2 Temperature/humidity gap between
to be consistent.
member surface
In knowledge representation system K = (U, A), for every primi- C3 Repetition of freezing and thawing
tive attribute a 2 A there exists a set of its values known as the C4 Fire
domain of a, denoted as Va. For every a 2 A, there exists a function C5 Surface heating
C6 Chemical reaction of acid and chloride
a:U ? Va, which assigns a unique attribute value from Va to every
C7 Corrosion of reinforcing bar
object x 2 U. For subsets of attributes B  A, the binary relations ex- (neutralization)
pressed as Eq. (13) are called the indiscernibility relation over B, C8 Corrosion of reinforcing bar (penetrated
chloride)
INDðBÞ ¼ fðx; yÞ 2 U 2 jaðxÞ ¼ aðyÞ; 8a 2 Bg ð13Þ
Structure and applied loads D1 Eternal long time loading within design
In knowledge representation system K = (U,A), if P and Q (P \ Q = U) load
D2 Eternal long time loading over design load
are subsets of A, P represents the condition attributes, and Q the
D3 Dynamic short time loading within design
decision attributes, then this knowledge representation system load
can be called a decision table. The equivalence classes of IND(P) D4 Dynamic short time loading over design
and IND(Q) are called the condition and decision equivalence clas- load
ses, respectively. D5 Lack of section size and reinforcing bars
D6 Differential settlement of structure
For every x 2 U, a function dx:A ? V is defined to satisfy
D7 Freezing of ground
dx(a) = a(x) ("a 2 P [ Q), and here, dx is called a decision rule of
Table 2
Checklist for causes of cracks

Category of crack Crack Time of formation Shape Regularity Cause of concrete deformation Range
cause No.
Several hours Several More than several Turtle Linear Regular Irregular Contraction Expansion Settlement  bending  shear Material Member Structure
1 day days tens of days
Surface Penetrate
Concrete material A1     
A2       
A3      
A4       
A5      
A6       

Y.M. Kim et al. / Advances in Engineering Software 40 (2009) 202–211


A7      
A8     
A9         
Construction work B1       
B2         
B3         
B4       
B5      
B6      
B7      
B8       
B9       
B10       
B11     
B12     
B13     
B14     
B15      
B16       
Service and C1          
environmental C2       
factors C3        
C4         
C5         
C6      
C7     
C8     
Structure and applied D1      
loads D2      
D3      
D4      
D5       
D6     
D7     

205
206 Y.M. Kim et al. / Advances in Engineering Software 40 (2009) 202–211

3. Decision algorithm for determining the category of crack and 58 are inconsistent. The inconsistent decision rules are repre-
causes sented as shadowed.

In this section, we discuss the generation of a rough set based 3.2. Composition of rough sets and reduction of knowledge
decision algorithm for determining the category of crack causes
to which an inspected crack belongs, and evaluate the applicability From Table 4, the equivalence classes of condition attributes P
of this decision algorithm using actual crack cases. and decision attributes Q of universe U are shown in Eq. (15) and
Eq. (16), respectively:
3.1. Construction of the decision table
U=P ¼ fE1 ; E2 ; E3 ; E4 ; E5 ; E6 ; . . . ; E65 g
E1 ¼ fx1 g; E2 ¼ fx2 g; E3 ¼ fx3 g; E4 ¼ fx4 ; x58 g ð15Þ
According to the Japanese Concrete Institute (JCI), the categories
of crack causes can be divided into four classes, that is, (1) concrete E5 ¼ fx5 g; E6 ¼ fx6 ; x28 ; x82 g; . . . ; E65 ¼ fx91 g
material, (2) construction work, (3) service and environmental fac- U=Q ¼ fX 1 ; X 2 ; X 3 ; X 4 g
tors, and (4) structure and applied loads. Each category has its own X 1 ¼ fx1 ; x2 ;    ; x25 g; X 2 ¼ fx26 ; x27 ; . . . ; x61 g ð16Þ
detailed list of crack causes numbering 9, 16, 8 and 7, respectively, X 3 ¼ fx62 ; x63 ;    ; x91 g; X 4 ¼ fx92 ; x93 ; . . . ; x98 g
with overall crack causes numbering 40 as shown in Table 1. The
detailed crack causes are estimated using the checklist of crack fea- The dependency or consistency of the decision algorithm of decision
tures (represented in Table 2), which is widely used in Japan and attributes Q to condition attributes P is calculated as in Eq. (17). The
Korea [11,12]. degree of dependency k is calculated to be 0.408, which is between
Table 2 can be regarded as a decision table comprised of condi- 0 and 1, which means that Q is roughly dependent on P. The rela-
tion attributes of crack symptoms and decision attributes of crack tively low value of 0.408 means that there are many inconsistent
causes. This decision table is inconsistent because it maps more cases where, for the same set of condition attributes, there exist
than two crack causes for the same crack symptoms; for example, non-identical decision attributes.
the cracks numbered B2 and B3 have identical crack symptoms. It cardPOSP ðQ Þ 40
may also be hard to determine from this decision table if certain k ¼ cP ðQ Þ ¼ ¼ ¼ 0:408
cardU  98 
symptoms are insufficient or even missing. For this case we can
here; cardPOSP ðQ Þ ¼ card [ PX ð17Þ
counter cracking more successfully if we can extract the minimal X2U=Q
characteristics necessary for diagnosis. According to the impor- ¼ cardðPX 1 [ PX 2 [ PX 3 [ PX 4 Þ ¼ 9 þ 19 þ 12 þ 0 ¼ 40
tance of each crack feature for crack diagnosis, the absence of some
feature(s) could make the diagnosis difficult or completely ineffec- The accuracy measure of equivalence classes of each decision attri-
tive. Therefore, it is useful to know the importance of each crack bute with respect to equivalence classes of condition attributes is
feature so as to determine what features should be examined, shown in Eqs. (18)–(21), respectively.
and to determine the reliability of the decision if some features cardPX 1 9
are missing. aP ðX 1 Þ ¼ ¼ ¼ 0:170 ð18Þ
cardPX 1 53
Table 3 shows attributes and their values in constructing the cardPX 2 19
decision table from Table 2. Table 4 is the decision table taken both aP ðX 2 Þ ¼ ¼ ¼ 0:352 ð19Þ
cardPX 2 54
from the relation of crack features and crack categories of Table 2
cardPX 3 12
and the crack attributes defined in Table 3. We extracted a total of aP ðX 3 Þ ¼ ¼ ¼ 0:235 ð20Þ
98 decision rules from Table 2, out of which 40 rules are consistent cardPX 3 51
cardPX 4 0
aP ðX 4 Þ ¼ ¼ ¼ 0:000 ð21Þ
cardPX 4 16
Table 3 The roughness of equivalence classes of each decision attribute with
Attributes and attribute values of decision table for causes of cracks respect to those of the condition attributes are calculated from the
Classification of Attribute name Attribute Attribute Attribute results of the accuracy measure as in Eqs. (22)–(25), respectively.
attributes symbol condition value
qP ðX 1 Þ ¼ 1  aP ðX 1 Þ ¼ 0:830 ð22Þ
Condition Time of formation a Several hours 1 0
attributes day
qP ðX 2 Þ ¼ 1  aP ðX 2 Þ ¼ 0:648 ð23Þ
(P) Several days 1 qP ðX 3 Þ ¼ 1  aP ðX 3 Þ ¼ 0:765 ð24Þ
More than several 2 qP ðX 4 Þ ¼ 1  aP ðX 4 Þ ¼ 1:000 ð25Þ
tens of days
Shape b Turtle 0
From the calculated results of the accuracy measure and roughness,
Linear Surface 1
Penetrate 2
it can be concluded that if we classify the category of crack causes
Regularity c Regular 0 from the diagnosed crack features using the checklist in Table 2, we
Irregular 1 can classify cracks due to construction work (X2) best, and service
Cause of concrete d Contraction 0 and environmental factors (X3) second, concrete material (X1) third,
deformation Expansion 1
and we can not classify cracks due to structure and applied loads
Settlement 2
bending shear (X4) at all.
Range e Material 0 From the topological characterization of equivalence classes of
Member 1 decision attributes to those of condition attributes, we can say that
Structure 2 concrete material (X1), construction work (X2) and service and
Decision Category of crack f Concrete material A
attributes cause Construction work B
environmental factors (X3) are roughly P-definable for they satisfy
(Q) Service and C R X – U and RX–U, and structure and applied loads (X4) are inter-
environmental nally P-undefinable for they satisfy R X = U and RX–U. This means
factors that in the category of crack causes, X1, X2 and X3 have rules for
Structural and D
determining whether cracks with specific features belong to each
applied loads 
or are complementary with each of them X c1 ; X c2 ; X c3 . On the other
Y.M. Kim et al. / Advances in Engineering Software 40 (2009) 202–211 207

Table 4
Unreduced decision rules and core of condition attributes

Rule Condition attributes (P) Decision attributes (Q) Q-core of P COREQ(P) Rule Condition attributes (P) Decision attributes (Q) Q-core of P COREQ(P)
U abcde f abcde U abcde f abcde
1 0 1 100 A 0–––0 50 1 1121 B 1 – 1 – –
2 1 1 000 A 1–––0 51 0 0101 B 0 – – – –
3 1 2 000 A 1–––– 52 0 0102 B 0 – – – –
4 1 1 001 A 53 0 1101 B 0 – – – 1
5 1 2 001 A 12–0– 54 0 1102 B 0 – – – 2
6 2 0 110 A 55 1 0121 B – – – – –
7 2 0 110 A 56 2 0121 B – 0 – 2 1
8 1 0 100 A 1–––– 57 2 1021 B
9 2 0 100 A 58 1 1001 B
10 1 1 100 A 1–––– 59 1 1002 B 1–––2
11 2 1 100 A 60 1 1021 B
12 2 1 120 A ––120 61 1 2021 B
13 2 0 120 A –––20 62 2 1000 C
14 2 1 011 A 63 2 1001 C
15 2 1 012 A 64 2 1002 C
16 0 1 021 A 65 2 1010 C ––010
17 2 0 000 A –00–0 66 2 1011 C
18 2 0 001 A 67 2 1012 C
19 2 0 002 A 68 2 1020 C ––020
20 2 1 000 A 69 2 1021 C
21 2 1 001 A 70 2 1022 C
22 2 1 002 A 71 2 2000 C
23 2 2 000 A 72 2 2001 C
24 2 2 001 A 73 2 2002 C
25 2 2 002 A 74 2 2010 C – – – 1 –
26 2 0 100 B 75 2 2011 C – 2 – 1 –
27 2 0 101 B 76 2 2012 C – 2 – 1 –
28 2 0 110 B 77 2 2020 C – – – 2 0
29 2 0 111 B 0 – – – – 78 2 2021 C
30 0 0 001 B 0 – – – – 79 2 2022 C
31 0 0 002 B 0 – – 0 – 80 2 0100 C
32 0 1 001 B 0 – – – – 81 2 0101 C
33 0 1 002 B 0 – – – – 82 2 0110 C
34 0 2 001 B 0 – – – – 83 2 0111 C
35 0 2 002 B 84 2 1100 C
36 2 0 001 B 85 2 1101 C 21101
37 2 0 002 B 86 2 1110 C
38 2 1 001 B 87 2 1111 C – 1 1 1 1
39 2 1 002 B 88 2 0102 C 2 – 1 – 2
40 2 2 001 B 89 2 0112 C – – – – 2
41 2 2 002 B 02––– 90 2 1102 C 2 – 1 0 2
42 0 2 021 B ––––– 91 2 1112 C – – 1 1 2
43 0 2 121 B 92 2 1021 D
44 2 2 021 B –21–– 93 2 1022 D
45 2 2 121 B 94 2 2021 D
46 0 1 021 B 0–1–– 95 2 2022 D
47 0 1 121 B 96 1 1021 D
48 2 1 121 B ––122 97 1 2021 D
49 2 1 122 B 98 2 1121 D

hand, X4 does not have any rule for its belongingness, but merely
functions for the belongingness of the complementary of it (X c4 Þ.
The importance of the condition attribute is defined by the con-
cept of the degree of alternation of the positive region of the deci-
sion rules if the corresponding attribute is deleted. The importance
of attribute h is defined using the notion of dependency defined in
Eq. (12), i.e.,
Ih ¼ cP ðQ Þ  cPfhg ðQ Þ ð26Þ

Fig. 2 shows the importance of each condition attribute calculated


using Eq. (26). From this we can conclude, if we determine to which
category of crack causes the given crack belongs, that ‘‘time of for-
mation” has the most influence, as it alters the positive region of
equivalence classes of the condition attributes the most, and
‘‘shape” has the least influence as it alters the positive region the
least. This result can be used as a guide for crack diagnosis and indi-
cates that the time of crack formation should be investigated with
the most importance to increase the reliability of the diagnosis. Fig. 2. Importance of each condition attribute in decision making.
208 Y.M. Kim et al. / Advances in Engineering Software 40 (2009) 202–211

Table 5 decision rules. Explaining the method of obtaining a core condition


Reduced decision rules attribute in detail goes as follows: If we eliminate attribute ‘‘a”
Reduct Q-reduct of P Decision Reduct Q-reduct of P Decision from rule 1 (01100), then rule 1 and 84 (21100) cannot be dis-
REDQ(P) attributes (Q) REDQ(P) attributes (Q) cerned with the remaining attributes (1100). Therefore, each
U abcde f U abcde f attribute value of attribute ‘‘a” of rules 1 and 84, that is, 0 and 2,
1 0 – – – 0 A 47 0 – 1 – 1 B respectively, are core condition attributes of each rule.
2 1 – – – 0 A 47(1) 0 – 1 2 – B In Table 5, the reducts of the condition attributes for the deci-
3 1 – – – 0 A 49 – – 1 2 2 B
3(1) 1 2 – 0 – A 50 1 – 1 2 – B
sion rules of Table 4 are represented. The reducts of condition attri-
5 1 2 – 0 – A 50(1) 1 – 1 – 1 B butes are those sets of minimal attributes which make the
8 1 – – – 0 A 51 0 0 – – – B consistent decision rule true, and can be induced by adding inde-
8(1) 1 0 – 0 – A 51(1) 0 – – 0 1 B pendent attribute values C which do not satisfy Eq. (10) at the core
8(2) 1 – 1 0 – A 51(2) 0 – 1 – 1 B
condition attribute, COREQ(P). In rule 1 the core becomes reduct di-
10 1 – – – 0 A 52 0 0 – – – B
10(1) 1 – 1 0 – A 52(1) 0 – – – 2 B rectly, this means that rule a0b1c1d0e0 ? fA can be reduced to the
12 – – 1 2 0 A 53 0 – 1 – 1 B equivalent decision rule a0e0 ? fA. Rule 3 a1b2c0d0e0 ? fA has two
13 – 0 – 2 0 A 53(1) 0 – – 0 1 B reducts, a1e0 ? fA and a1b2d0 ? fA, which means the original rule
13(1) – – 1 2 0 A 54 0 – – – 2 B can be reduced to these two equivalent decision rules.
17 – 0 0 – 0 A 55 1 0 – 2 – B
30 0 0 – – – B 55(1) 1 0 – – 1 B
30(1) 0 – 0 0 – B 55(2) – 0 – 2 1 B 3.3. Generation of the minimal decision algorithm
30(2) 0 – – 0 1 B 55(3) 1 – 1 2 – B
31 0 – – – 2 B 56 – 0 – 2 1 B Table 6 is the minimal decision algorithm generated from the
31(1) 0 0 – – – B 59 1 – – – 2 B
reduced decision rules of Table 5, and is equivalent to the decision
31(2) 0 – 0 0 – B 65 – – 0 1 0 C
32 0 – 0 0 – B 68 – – 0 2 0 C rules of Table 4. This minimal decision algorithm consists of 23
32(1) 0 – – 0 1 B 74 – 2 – 1 – C rules overall, and satisfies all the consistent decision rules of Table
33 0 – – – 2 B 74(1) – – 0 1 0 C 4.
33(1) 0 – 0 0 – B 75 – 2 – 1 – C If we represent some selected minimal decision algorithm from
34 0 2 – – – B 76 – 2 – 1 – C
Table 6 in an ‘‘if {condition}, then {decision}” form, the result can
34(1) 0 – 0 0 – B 77 – 2 – 2 0 C
34(2) 0 – – 0 1 B 77(1) – – 0 2 0 C be stated as follows:
35 0 2 – – – B 85 2 1 1 0 1 C
35(1) 0 – – – 2 B 87 – 1 1 1 1 C  Rules for concrete material
35(2) 0 – 0 0 – B 88 2 0 1 – 2 C
Rule 1: a0e0 ? fA
42 0 2 – – – B 88(1) 2 – 1 0 2 C
43 0 2 – – – B 89 – 0 – 1 2 C
if (time of formation; several hours 1 day) and (range;
43(1) – 2 1 – – B 89(1) – – 1 1 2 C material)
43(2) 0 – 1 2 – B 89(2) 2 0 1 – 2 C then (category of crack cause; concrete material)
43(3) 0 – 1 – 1 B 90 2 – 1 0 2 C :
45 – 2 1 – – B 91 – – 1 1 2 C
Rule 5: b0c0e0 ? fA
if (shape; turtle) and (regularity; regular) and (range;
material)
Table 6 then (category of crack cause; concrete material)
Minimal decision algorithm
 Rules for construction work
Rule Related reducts (duplicate Q-reduct of Decision Decision rules Rule 6: a0b0 ? fB
reducts in parentheses) P REDQ(P) attributes if (time of formation; several hours 1 day) and (shape;
(Q)
U abcde f P?Q
turtle)
then (category of crack cause; construction work)
1 1 0 – – –0 A a0e0 ? fA
:
2 2,3,8,10 1 – – –0 A a1e0 ? fA
3 5(3(1)) 1 2 – 0– A a1b2d0 ? fA Rule 15: a1e2 ? fB
4 12,13(1) – – 1 20 A c1d2e0 ? fA if (time of formation; several days) and (range; structure)
5 17 – 0 0 –0 A b0c0e0 ? fA then (category of crack cause; construction work)
6 30,31(1),51,52 0 0 – –– B a0b0 ? fB
 Rules for service and environmental factors
7 32,33(1),34(1),35(2) 0 – 0 0– B a0c0d0 ? fB
(30(1),31(2)) Rule 16: c0d1e0 ? fC
8 42,43(34,35) 0 2 – – – B a0b2 ? fB if (regularity; regular) and (cause of concrete deformation;
9 45 – 2 1 – – B b2c1 ? fB expansion) and (range; material)
10 47,53(43(3),51(2)) 0 – 1 – 1 B a0c1e1 ? fB then (category of crack cause; service and environmental
11 49 – – 1 2 2 B c1d2e2 ? fB
factors)
12 50,55(3) 1 – 1 2 – B a1c1d2 ? fB
13 54(31,33,35(1)) 0 – – – 2 B a0e2 ? fB :
14 56(55(2)) – 0 – 2 1 B b0d2e1 ? fB Rule 23: c1d1e2 ? fC
15 59 1 – – – 2 B a1e2 ? fB if (regularity; irregular) and (cause of concrete deformation;
16 65,74(1) – – 0 1 0 C c0d1e0 ? fC
expansion) and (range; structure)
17 68,77(1) – – 0 2 0 C c0d2e0 ? fC
18 75,76(74) – 2 – 1 – C b2d1 ? fC
then (category of crack cause; service and environmental
19 85 2 1 1 0 1 C a2b1c1d0e1 ? fC factors)
20 87 – 1 1 1 1 C b1c1d1e1 ? fC
21 88,89(2) 2 0 1 – 2 C a2b0c1e2 ? fC 3.4. Case study and evaluation of applicability
22 90(88(1)) 2 – 1 0 2 C a2c0d1e2 ? fC
23 91(89(1)) – – 1 1 2 C c1d1e2 ? fC
In this section, we show the application of the rough set based
minimal decision algorithm in determining the category of crack
In Table 4, we also represented the core of condition attribute P causes for the cracks actually diagnosed by the experts and evalu-
with respect to decision attribute Q (COREQ(P)) for the consistent ate the applicability of this decision algorithm. The application
Y.M. Kim et al. / Advances in Engineering Software 40 (2009) 202–211 209

examples and the corresponding diagnosis results were extracted of crack formation is estimated not just after concrete placement
from the technical manuals published by Japanese Concrete Insti- but at least several days after concrete placement. The width of
tute [12]. Five types of cracks, that is, four wall cracks and one slab the cracks is in the range of 0.2–1 mm, and most of the cracks ex-
crack, generated in four structures were chosen as application tended both to the upper and lower work joints. It is uncertain
examples. whether the cracks penetrate but it seems to be somewhat on
the penetrate side. It is estimated that the cause of concrete defor-
3.4.1. Case 1: cracks on a concrete wall mation was contraction and the range of the cracks are at member
Features of cracks. After 4 weeks of concrete placement, cracks level.
were found on the wall as shown in Fig. 3a. The cracks in the mid-
dle region are 0.3 mm width and penetrated the wall. The range of Diagnosis by the rough set based minimal decision algorithm. From
the cracks is estimated to be member level and the cause of con- the above crack features, it is evident that regularity (c) is regular,
crete deformation seems to be neither contraction nor expansion. cause of concrete deformation (d) is contraction, and range (e) is
The regularity of the cracks is not certain, but it seems to be some- member, so the resulting condition attributes can be expressed
what on the irregular side. as c0d0e1. The decision rule which satisfies these features does
not exist in the decision algorithm in Table 6. Among the unused
Diagnosis by the rough set based minimal decision algorithm. From crack features, time of crack formation (a) and shape of crack (b)
the above crack features, it is evident that the time of crack forma- are second evident features and it can be regarded as several days
tion (a) is several tens of days, the shape (b) is penetrate, cause of and penetrate, respectively, with the high possibility, so the result-
concrete deformation (d) is settlement  bending  shear, and the ing condition attributes become a1b2c0d0e1. The decision rule
range (e) is member, and the resulting condition attributes can which satisfies these features is Rule 3, that is, a1b2d0 ? fA. There-
be expressed as a2b2d2e1. The decision rule which satisfies these fore we can judge that the category of given cracks is concrete
features does not exist in the decision algorithm in Table 6. Among material (A).
the unused crack features, the regularity (c) is second evident fea- It can be seen from the above diagnosis that, in diagnosing the
ture and it can be regarded as irregular with the high possibility, so above cracks, time of formation (a), shape (b) and cause of concrete
the resulting condition attributes become a2b2c1d2e1. The decision deformation (d) are essential features, but regularity (c) and range
rule which satisfies these features is Rule 9; that is, b2c1 ? fB. (e) are not essential features.
Therefore we can judge that the category of given cracks is con-
struction work (B). Diagnosis by an expert. An expert diagnosed the given cracks as due
It can be seen from the above diagnosis that, in diagnosing the to the ‘‘heat of hydration of cement (A2)” and the category of the
above cracks, shape (b) and regularity (c) are used as essential fea- cause of this crack is concrete material (A) which is identical to that
tures, but time of formation (a) cause of concrete deformation (d) derived from the rough set based decision algorithm.
and range (e) are not essential features.
3.4.3. Case 3: cracks on exterior wall of 2 stories building
Diagnosis by an expert. An expert diagnosed the given cracks as the Features of cracks. Cracks were found on two stories RC building on
‘‘improper treatment of work joint (B10)” and the category of this August after 1 years of building completion as shown in Fig. 3c. The
crack cause is construction work (B), which is identical to that de- 1st floor was used as office area and air cooling was performed, and
rived from the rough set based decision algorithm. 2nd floor was used as warehouse and there was no air cooling.
Cracks of 1st floor were generated vertically both on the middle
3.4.2. Case 2: cracks on the side wall of a concrete box structure of each wall and corners of opening on inner side of exterior wall
Features of cracks. Regular cracks of 2–5 m intervals were found on and not penetrated the wall. Cracks of 2nd floor were on the whole
the side wall of concrete box structures as shown in Fig. 3b. The triangular shaped and 80% of them penetrated the wall. It is esti-
cracks were found during the removing of forms about 2 weeks mated that major cause of concrete deformation is expansion.
after concrete placement, and from the shape of cracks the time The cracks of 1st floor show regularity on each wall, so the range

Fig. 3. Example of crack cases.


210 Y.M. Kim et al. / Advances in Engineering Software 40 (2009) 202–211

of cracks can be regarded as member level, and the cracks of 2nd and ‘‘segregating during pumping (B3)” are dominant causes of
floor show regularity on the whole building, so the range of cracks promoting shrinkage of concrete.
can be regarded as structure level.
3.5. Evaluation of applicability
Diagnosis by the rough set based minimal decision algorithm. The
cracks of 1st and 2nd floor need to be diagnosed separately for they We generated the decision algorithm for consistent decision
showed peculiar features of their own. From the above crack fea- making with the minimal set of attributes. In determining the cat-
tures of 1st floor, it is evident that time of crack formation (a) is egory of crack causes using the generated decision algorithm, we
more than several tens of days, shape (b) is surface, regularity (c) do not have to consider non-essential attributes, whether their
is regular, cause of concrete deformation (d) is expansion, and attribute values are evident or not, but for essential attributes
range (e) is member, so the resulting condition attributes can be the diagnosis is impossible without knowing their attribute values.
expressed as a2b1c0d1e1. The decision rule which satisfies these fea- By minimizing the essential attributes, however, the diagnosis can
tures does not exist in the decision algorithm in Table 6. Therefore be processed more easily.
we cannot judge the category of crack causes for the cracks with For the five types of cracks generated in four structures, it was
the given features. possible to find the category of crack causes for three of them, but
From the above crack features of 2nd floor, it is evident that impossible for two of them. Rough set based decision algorithm
time of crack formation (a) is more than several tens of days, shape does not always ensure to reach the conclusion with the given fea-
(b) is penetrate, regularity (c) is regular, cause of concrete deforma- tures, but accommodates the decision making by using minimal
tion (d) is expansion, and range (e) is structure, so the resulting essential attributes. And if attributes and their attribute values
condition attributes can be expressed as a2b2c0d1e2. The decision are identical to those of decision rules derived by rough set based
rules which satisfies these features are Rule 18 and Rule 22, that decision algorithm, it is sure to produce the consistent decision
is, b2d1 ? fC and a2c0d1e2 ? fC, respectively. Therefore we can judge making.
that the category of cracks of 2nd floor is service and environmen-
tal factors (C).
It can be seen from the above diagnosis that, in diagnosing 4. Summary and conclusions
cracks of 1st floor there is no essential features, but cracks of 2nd
floor have two sets of essential features, one is time of formation This study presents a new approach to rough set theory in
(a), regularity (c), cause of concrete deformation (d) and range determining the category of crack causes with insufficient and
(e), and the other is shape (b) and cause of concrete deformation imprecise crack characteristics of concrete structures. A rough set
(d). based minimal decision algorithm was generated and the applica-
bility of this decision algorithm was evaluated by case studies. The
Diagnosis by an expert. An expert diagnosed that the cracks of 1st results of our study and related discussion can be summarized as
floor were generated mainly due to the ‘‘differences in temperature follows:
and humidity between member surfaces (C2)” and the category of In the realm of crack diagnosis, an expert’s experiential knowl-
the cause of this crack is service and environmental factors (C), and edge plays an important role, though the knowledge, in itself, is not
the cracks of 2nd floor were generated by the expansion of both the precise and moreover sometimes inconsistent. Along with the
roof slab and the whole structure due to high temperature of sum- problems of the expert’s knowledge, the crack features gathered
mer weather, and the cause of this crack can be regarded as from inspection also have the problems of imprecision, ambiguity
‘‘change of external temperature and humidity (C1)”, the category and insufficiency. By applying rough set theory, we were able to
of which is service and environmental factors (C) which is identical generate a decision algorithm which requires minimal crack char-
to that derived from the rough set based decision algorithm. acteristics and hence helped to overcome the problems of the ex-
pert’s knowledge by using the core part of the knowledge, and
3.4.4. Case 4: cracks on the floor slab accordingly secures flexibility against the uncertainty and imper-
Features of cracks. After 3 weeks of concrete placement, cracks fection of investigated data.
were found on the upper side of floor slab as shown in Fig. 3d. In determining the category of crack causes, the importance of
Some cracks were generated along the short edge of the beam each crack feature could be acquired by calculating the degree of
and the others were generated diagonally on four corners, and half each feature’s influence on consistent decision making. Among
of the cracks penetrated the slab. Similar crack patterns were the five crack features, that is, time of formation, shape, regularity,
found on neighboring floor slabs. It is estimated that the cause of cause of concrete deformation, and range, it turned out that ‘‘time
concrete deformation is contraction and the range of the cracks of formation” had the most important influence on decision mak-
is at member level. ing, and from this result we can conclude that in investigating
crack features, time of crack formation should be checked as accu-
Diagnosis by the rough set based minimal decision algorithm. From rately as possible to increase the reliability of the decision.
the above crack features it is evident that time of crack formation This study extracted a decision algorithm from existing knowl-
(a) is more than several tens of days, shape (b) is penetrate, regu- edge and generated decision rules for 3 of 4 crack categories but
larity (c) is regular, cause of concrete deformation (d) is contrac- could not generate decision rules for the remaining category of
tion, and range (e) is member, so the resulting condition ‘‘structure and applied loads”, for there does not exist consistent
attributes can be expressed as a2b2c0d0e1. The decision rule which decision rules corresponding to this category. This is because the
satisfies these features does not exist in the decision algorithm in crack features of ‘‘structure and applied loads” were roughly de-
Table 6. Therefore we cannot judge the category of crack causes fined and had duplicated characteristics with other categories.
for the cracks with the given features. Therefore, to generate decision rules which cover all the categories
and hence to enhance the exactness and reliability of diagnosis, it
Diagnosis by an expert. An expert diagnosed that if there does not is necessary to refine the existing knowledge and reinforce rational
prominent construction faults, it is estimated that ‘‘shrinkage of new knowledge to the existing knowledge.
concrete (A9)” is the major cause of cracks, and that it is necessary The proposed approach has its own limits in application in
to scrutinize construction records, because ‘‘overtime mixing (B2)” views of flexibility, for it treats crack symptoms on a binary basis.
Y.M. Kim et al. / Advances in Engineering Software 40 (2009) 202–211 211

If fuzzy concept is used in this area, it is expected that the flexibil- [3] Pawlak Z. Rough sets. Int J Comput Inform Sci 1982;11:341–56.
[4] Chao CJ, Cheng FP. Fuzzy pattern recognition model for diagnosing cracks in RC
ity of application to be more improved.
Structures. J Comput Civil Eng ASCE 1998;12(2):111–9.
[5] Cheng HD, Chen JR, Clazier C, Hu YG. Novel approach to pavement cracking
Acknowledgements detection based on fuzzy set theory. J Comput Civil Eng ASCE
1999;13(4):270–80.
[6] Kim YM, Kim CK, Hong GH. Fuzzy set based crack diagnosis system for
This research (05HaksimC03) was financially supported by the reinforced concrete structures. Comput Struct 2007;85:1828–44.
Ministry of Construction & Transportation of South Korea, Korea [7] Attoh-Okine N. Rough set application to data mining principles in pavement
Institute of Construction and Transportation Technology Evalua- management database. J Comput Civil Eng ASCE 1997;11(4):231–7.
[8] Kaname F, Tu BH. A rough set approach to information retrieval. Rough sets in
tion and Planning, and the Regional Research Centers Program knowledge discovery, vol. 2. Heidelberg: Physica-Verlag; 1998. p. 166–77.
(Bio-housing Research Institute) granted by the Korean Ministry [9] Furuta H, Hirokane M, Mikumo Y. Extraction method based on rough set
of Education & Human Resources Development, and the authors theory of rule-type knowledge from diagnostic cases of slope-failure danger
levels. Rough sets in knowledge discovery, vol. 2. Heidelberg: Physica-Verlag;
are grateful to the authorities for their support. 1998. p. 178–92.
[10] Yao YY, Wong SKM, Lin TY. A review of rough set models. Rough sets and data
References mining. Kluwer Academic Publishers; 1997. p. 47–75.
[11] Architectural Institute of Korea. Repair and rehabilitation of concrete
structures. AIK, 1997.
[1] Kim YM, Kim CK, Hong SG. Fuzzy based state assessment for reinforced
[12] Japanese Concrete Institute. Investigation of concrete cracks and guide to
concrete building structures. J Eng Struct 2006;28(9):1286–97.
repair and rehabilitation. JCI, 1980.
[2] Pawlak Z. Rough sets: theoretical aspects of reasoning about data. Kluwer
Academic Publishers; 1991.

You might also like