Académique Documents
Professionnel Documents
Culture Documents
Ying-Hua Huang1
Abstract: Accurate prediction of bridge condition is essential for the planning of maintenance, repair, and rehabilitation. An examination
of the assumptions 共for example, maintenance independency兲 of the existing Markovian model reveals possible limitations in its ability to
adequately model the procession of deterioration for these purposes. This study uses statistical analysis to identify significant factors
Downloaded from ascelibrary.org by GADJAH MADA UNIVERSITY on 11/26/17. Copyright ASCE. For personal use only; all rights reserved.
influencing the deterioration and develops an application model for estimating the future condition of bridges. Based on data derived from
historical maintenance and inspection of concrete decks in Wisconsin, this study identifies 11 significant factors and develops an artificial
neural network 共ANN兲 model to predict associated deterioration. An analysis of the application of ANN finds that it performs well when
modeling deck deterioration in terms of pattern classification. The developed model has the capacity to accurately predict the condition of
bridge decks and therefore provide pertinent information for maintenance planning and decision making at both the project level and the
network level.
DOI: 10.1061/共ASCE兲CF.1943-5509.0000124
CE Database subject headings: Bridges; Deterioration; Predictions; Neural networks.
Author keywords: Bridges; Deterioration; Predictions; Neural networks.
Introduction and Weyers 1992; Bettigole and Robison 1997; Kirkpatrick et al.
2002; Scott et al. 2003兲. This paper ignores all other elements and
Transportation agencies run bridge management systems 共BMS兲 identifies significant factors influencing bridge deck deterioration
for the maintenance of bridges in order to ensure public safety. by using the relevant historical inspection and inventory data and
Inspection of the bridges yields data that they use to develop records of maintenance. An artificial neural network 共ANN兲
deterioration models in order to estimate the future bridge condi- model for predicting bridge deck deterioration is developed to
tion for timely planning of maintenance, repair, and rehabilitation. model the possible nonlinear relationships between the observed
The success of BMS depends on its ability to estimate future condition and the identified factors. Bridge managers using this
conditions accurately 共Kleywegt and Sinha 1994; Madanat et al. ANN model can, potentially, predict the condition of individual
1995; Thompson et al. 1998兲. Several BMSs 共for example, Pontis decks and timely specify information pertinent to maintenance
and BRIDGIT兲 have used Markovian transition probabilities in needs.
deterioration models to estimate the future conditions of bridge
elements. The underlying assumption of the Markov chain is that
at any given state of the process the future states are independent Bridge Deterioration Modeling
of the past 共Kleywegt and Sinha 1994兲. Although a number of
researchers use Markovian approach to help decisions in mainte- Along with the development of BMS, researchers contrived vari-
nance 共Tao et al. 1994; Jha and Abdullah 2006; Morcous 2006兲, ous deterioration models to improve the bridge management de-
many researchers have questioned the realism of this assumption cision processes. Madanat et al. 共1997兲 introduced an ordered
and by implication the ability of the Markovian approach to probit analysis to develop an incremental and discrete deteriora-
model bridge deterioration for purposes of the allocation of re- tion model where the difference in observed ratings is an indicator
sources 共Busa et al. 1985; Madanat et al. 1997; Mauch and of the underlying latent deterioration. Mauch and Madanat 共2001兲
Madanat 2001; Frangopol et al. 2001; Robelin and Madanat developed a semiparametric hazard rate model to predict times
2007兲. between changes in the condition of concrete bridge decks. They
In order to expose the weaknesses of the Markov chain ap-
concluded that age and other explanatory variables, such as aver-
proach this study uses statistical analysis to identify significant
age daily traffic 共ADT兲, contribute to bridge deck deterioration.
factors influencing the deterioration and develops an application
They also concluded that because the Indiana Bridge Inventory
model for estimating the future condition of bridges. Most bridges
database did not contain all relevant data, certain important cova-
become unserviceable because of deterioration of the deck 共Cady
riants, for example, records of maintenance of bridge decks, were
not included in the model. Robelin and Madanat 共2007兲 formu-
1
Assistant Professor, Dept. of Construction Engineering, National lated a realistic historical model of bridge deck deterioration by
Yunlin Univ. of Science and Technology, Douliou, Yunlin 64002, Taiwan. using a Markov chain while retaining aspects of the history,
E-mail: huangyh@yuntech.edu.tw
namely, deterioration and maintenance, as part of the model to
Note. This manuscript was submitted on August 26, 2009; approved
on January 30, 2010; published online on November 15, 2010. Discussion
overcome the limitations of the existing Markovian models.
period open until May 1, 2011; separate discussions must be submitted Some researchers modeled bridge deterioration at project
for individual papers. This paper is part of the Journal of Performance level. Hyman et al. 共1983兲 showed that the condition is related to
of Constructed Facilities, Vol. 24, No. 6, December 1, 2010. ©ASCE, the age of the bridge through a piecewise linear regression model.
ISSN 0887-3828/2010/6-597–602/$25.00. Busa et al. 共1985兲 used linear regression to model bridge deterio-
6. New bare deck+ AC overlay and membrane+ concrete null hypothesis 共p ⬎ 5%兲. Therefore, it indicates that the mainte-
overlay+ concrete overlay; nance history of a deck is not a significant factor for influencing
7. New bare deck with coated bars+ concrete overlay; the rate of deterioration from Condition State 1 to 2, but it sig-
Downloaded from ascelibrary.org by GADJAH MADA UNIVERSITY on 11/26/17. Copyright ASCE. For personal use only; all rights reserved.
8. New bare deck with coated bars+ AC overlay and nificantly influences the rate of deterioration from Condition State
membrane+ concrete overlay; and 2 to 3. In other words, the influence of maintenance history is
9. New bare deck with coated bars+ concrete overlay apparent when decks have reached Condition State 2. The records
+ concrete overlay. demonstrate that the maintenance-independent assumption of
For the 942 decks maintained, there are 2,300 inspection Markovian deterioration model does not hold with respect to
records indicating the condition of the decks. For analysis, this decks; maintenance history does influence the performance of
study defined A12 and A23 as the age of transition from Condition decks.
State 1 to 2 and from Condition State 2 to 3, respectively. The
mean age of A12 of the 942 decks studied is 11.35 years. Of the Condition and Age
nine maintenance history types, decks with Maintenance History The distribution of deck samples recorded as having deteriorated
Type 4 showed the lowest average age of 8.67 years, and the from one condition state to another was studied to examine
decks with Maintenance History Type 2 showed a highest average whether the distribution is the same among decks of different age
age of 20 years. The mean age of A23 of the decks studied is 16.02 categories. If the age-independent assumption of Markovian de-
years: of the nine maintenance history types, the decks with
terioration models holds, the distributions should be independent
Maintenance History Type 4 showed the lowest average age of
of age categories.
10.33 years, and the decks with Maintenance History Type 2
Data from the inspection of 1,241 decks with no history of
showed a highest average age of 23 years 共Table 1兲. Although the
maintenance were used to eradicate the influence of maintenance
decks are with the various maintenance histories, they are even-
records on the condition of the decks. For these 1,241 decks, there
tually maintained with the concrete overlays. But, in the mean-
are 2,488 relevant inspection records. Inspection records of decks
time, the deteriorations of the decks from the Condition State 1 to
with a biannual inspection period were used to form the transition
2 are observed at different average ages, and likewise from 2 to 3.
matrix. Taking initial Condition State 1 as example, as shown in
Therefore, ANOVA was conducted to evaluate the possibility
Table 2, the number of decks in five different age categories that
which the maintenance history has the statistically significant ef-
fect on the mean age of transition. had deteriorated from Condition State 1 to another condition state
Two null hypotheses were tested. The first null hypothesis is was counted and summarized. For Age Category 1 共1–10-year-old
that the mean age of transition from Condition State 1 to 2, i.e., decks兲, there were 73 decks in Condition State 1. After 2 years, 69
A12, for decks with different maintenance history types is the decks remained at Condition State 1 while 2 decks deteriorated to
each of Condition States 2 and 3. Therefore, the mean value of the
same, i.e., Ā12共1兲 = Ā12共2兲 = ¯ = Ā12共9兲, with Ā12共n兲 representing next condition state is 1.08. For Age Categories 2–5, the mean
the mean value of A12 for decks maintained with maintenance values of the next condition state are 1.14, 1.31, 1.53, and 1.52,
history type n. The second null hypothesis is that decks with
respectively. ANOVA was conducted to examine the null hypoth-
different maintenance history types have the same A23, i.e., mean
esis that for decks in Condition State 1 the mean values of the
age of transition from Condition State 2 to 3, i.e., Ā23共1兲 next condition state are all equal for decks among age categories.
= Ā23共2兲 = ¯ = Ā23共9兲, with Ā23共n兲 representing the mean value of The result of ANOVA shown in Table 2 indicates age dependency
A23 for decks maintained with maintenance history type n. The between the current condition 共Condition State 1兲 and the future
results of ANOVA are shown in Table 1 with corresponding sig- condition state. Older decks in Condition State 1 have a higher
nificance level 共p-value兲. These results show that there is not chance to deteriorate than younger decks in the same condition
enough evidence at a 95% level of confidence to reject the first state, as shown in Table 2.
Table 2. Distribution of Numbers of Decks Deteriorating from Initial Condition State 1 共Decks with No Maintenance兲
Next condition state
Age of deck at
initial condition state 1 Standard
Category 共years兲 1 2 3 4 5 Mean state deviation p-value
1 关1, 10兴 69 2 2 0 0 1.0822 0.3633 0.0001a
2 共10, 20兴 301 44 3 0 0 1.1437 0.3751
3 共20, 30兴 71 15 4 2 0 1.3152 0.5366
4 共30, 40兴 31 7 4 3 0 1.5333 0.9195
5 共40, 50兴 23 13 4 0 0 1.5250 0.6789
a
Significant at 95% level of confidence.
B, C/A 86.78/73.96 neurons in each hidden layer. For the 1,112 sets of input data, the
10 A, B/C 73.28/71.27 68.14/66.73 classifier yields average classification rates of 84.66 and 75.39%
A, C/B 67.22/64.09 for training and testing, respectively. The latter is considered to be
B, C/A 63.91/64.82 the prediction rate. To evaluate feasibility of the developed
15 A, B/C 55.19/54.74 55.24/54.58 model, concrete deck data obtained from one district in Wisconsin
A, C/B 55.30/54.69 were used. The data set contains an 836⫻ 16 matrix. The result
B, C/A 55.24/54.30 has an accuracy classification rate of 74.83% that approximates to
the test result of 75.39% of the model.
Number of hidden neurons
5 A, B/C 84.85/75.69 84.66/75.39
A, C/B 82.37/76.52 Conclusions
B, C/A 86.78/73.96
10 A, B/C 92.56/73.48 89.99/75.21 This study develops an ANN prediction model for deck deterio-
A, C/B 88.71/75.97 ration by adopting BP-MLP classifier. Using historical data of
B, C/A 88.71/76.17 decks in Wisconsin, it develops an 11-5-5–5-5-5-5 ANN, i.e., 11
15 A, B/C 93.11/71.27 93.02/72.90 inputs in the input layer, five hidden layers with five hidden neu-
A, C/B 93.66/74.03 rons in each layer, and five output labels in the output layer. In the
B, C/A 92.29/73.41 three-way cross validation, the ANN prediction model reaches
classification rates of 84.66 and 75.39% for the training sets and
the testing sets, respectively. These results indicate the potential
ratio of each feature to verify whether all values for a feature of the ANN model with BP-MLP classifier as a prediction tool.
remain constant or nearly constant, and thus the feature has to be The 11 factors identified as statistically significant for deck
removed. The redundant feature reduction examines the determi- deterioration in Wisconsin are the inputs of the ANN model.
nate rank of the feature matrix to verify the independence among These significant factors are maintenance history, age, previous
the features. After examining all the features, no feature needs to condition, and further eight factors including district, design load,
be removed and the total number of features remains 11. deck length, deck area, ADT, environment, degree of skew, and
number of spans. The results found that both age and the history
MLP Configuration Development and Validation of maintenance are statistically relevant to deck deterioration
The features listed in Table 5 were used to examine the BP-MLP 共NB: for decks maintained by concrete overlay the age refers to
classifier. The target labels are vectors 关1 0 0 0 0兴, 关0 1 0 0 0兴, the age of the concrete overlay兲, indicating that the two assump-
关0 0 1 0 0兴, 关0 0 0 1 0兴, and 关0 0 0 0 1兴, which present future tions of the Markovian deterioration model, namely, age indepen-
Condition States 1, 2, 3, 4, and 5, respectively. The training and dency and history independency, are invalid in this context.
testing data set contains a 3 , 336⫻ 16 matrix. The data set is The exploration of the potential of ANN for modeling deck
randomly divided into three equal subsets: A, B, and C on the deterioration found that ANN performs well when modeling deck
basis of three-way cross validation. Each subset contains a deterioration as a pattern classification problem. The developed
1 , 112⫻ 16 matrix. All neurons in the hidden layers use the hy- model has the capacity to accurately predict the condition of
perbolic tangent activation function; all output layers use the sig-
bridge decks and therefore provide pertinent information for
moidal activation function. The learning rate is 0.1, and the
maintenance planning and decision making. This is true at both
momentum constant is set as 0.8. The optimal configuration for
the project level and the network level. Application of the ANN
the BP-MLP classifier was determined by trial and error. The
classification rate is used as an indicator with a higher rate indi- prediction model to other elements of bridges or bridges in other
cating a higher classification completion. states or countries might require reidentification of significant pa-
The classification rates for various numbers of hidden layers rameters influencing deterioration and reassessment of the ANN
are shown in Table 6. The results show that five-layer MLP yields prediction model.
the best average classification rates in both training and testing. The recording of maintenance work carried out on bridge ele-
The average classification rate decreases when the number of hid- ments is a vitally important issue for effective BMS. Most agen-
den layers increases from 5 to 15. This result indicates that, for cies do not keep the records adequately. This study shows that the
this data set, any number of hidden layers greater than 5 does not record of maintenance directly influences the deterioration of a
yield better results. A different number of hidden neurons were bridge’s deck: meaning that a reliable bridge has a reliable record
tested with a five-layer MLP as a second step to determine the of maintenance.