You are on page 1of 12

Automation in Construction 141 (2022) 104460

Contents lists available at ScienceDirect

Automation in Construction
journal homepage: www.elsevier.com/locate/autcon

Bridge maintenance planning framework using machine learning,


multi-attribute utility theory and evolutionary optimization models
Hussaini Jaafaru a, Bismark Agbelie b, *
a
Department of Civil Engineering, The Catholic University of America, 620 Michigan, Washington, DC 20064, USA
b
The Catholic University of America, 620 Michigan, Washington, DC 20064, USA

A R T I C L E I N F O A B S T R A C T

Keywords: Although various tools and procedures have been developed to help transportation engineers objectively eval­
Bridges uate bridge maintenance needs, selecting and scoping projects still relies on engineering judgment. The present
Maintenance planning paper intends to help engineers evaluate and maintain bridges by developing a comprehensive bridge mainte­
Machine learning
nance planning framework (BMPF) within financial and performance constraints. The paper’s objectives are to
Multi-attribute criteria decision analysis
Multi-objectives optimization
maximize the performance condition level of bridges and to minimize the maintenance cost by optimally
Genetic algorithms planning the maintenance treatments. The framework includes bridge performance impact assessment, machine
learning models, multi-attribute utility theory ranking model and genetic algorithm optimization model. The
study analysed 95 bridges in a network with an 84% accuracy machine learning model prediction. Decision-
makers’ preferences were considered to rank all bridges using Multi-Attribute Utility Theory. 19 bridges were
then chosen for maintenance based on budget and performance using a genetic algorithm model. The BMPF was
observed to improve project productivity, reduce down time, and improve bridge inventory condition. Future
research can explore the use of other optimization approaches and also include traffic flow and construction cost
analysis for a better maintenance cost estimation. The machine learning model for performance prediction can be
enhanced by utilizing different techniques.

1. Introduction [2,43].
Bridge Management System (BMS) is a decision support tool based
Bridges play a crucial role in different aspects of our daily lives; on mathematical models to facilitate bridge condition predictions,
helping us to get to work or school, connect cities and contribute to the treatment recommendations, and a platform for which alternative pro­
economic growth of a country. In the United States, approximately grams can be efficiently investigated. According to the American Asso­
46,100 or 7.5% of all bridges were rated as structurally deficient, with ciation of State Highway and Transportation Officials (AASHTO), BMS is
an average of 178 million trips made across structurally deficient defined as a system designed to optimize the use of available resources
bridges each day. Furthermore, more than one-fourth of all bridges are for the inspection, maintenance, rehabilitation, and replacement of
over 50 years old, and many of the nation’s bridges are approaching the bridges [1,28,44].
end of their design life [51]. One of the key issues addressed by BMS is the question of whether
New and innovative approaches are needed to increase the lifespan and when a bridge should undergo maintenance and which repair steps
of bridges and decrease the lifetime cost of bridge structures. Therefore, should be taken. Meta-heuristic optimization models utilize objective
there is an urgent need to develop robust and reliable systems to monitor functions and play a key role in delivering an optimal maintenance plan
bridge infrastructure conditions and to apply sound maintenance stra­ for bridges. Various competing priorities have been applied in optimi­
tegies and planning. zation problems, such as minimizing total maintenance costs, maxi­
Maintenance planning is a bridge management function that in­ mizing surface quality or performance, minimizing health or hazard
volves a series of decisions on bridge type, location, and time of main­ risks, or optimum utilization of available resources and minimizing
tenance actions that should be taken over the bridge structure’s lifetime traffic flow disruption. Generally, any maintenance planned for a single
to minimize total maintenance costs and maximize bridge condition objective function may ignore or diminish the importance of other

* Corresponding author.
E-mail addresses: 70jaafaru@cua.edu (H. Jaafaru), agbelie@cua.edu (B. Agbelie).

https://doi.org/10.1016/j.autcon.2022.104460
Received 24 February 2022; Received in revised form 7 June 2022; Accepted 23 June 2022
Available online 8 July 2022
0926-5805/© 2022 Elsevier B.V. All rights reserved.
H. Jaafaru and B. Agbelie Automation in Construction 141 (2022) 104460

objectives. For example, a plan may minimize total maintenance cost by deterioration modeling, ranking using comparisons of different main­
sacrificing the performance condition of the pavement or vice versa. tenance alternatives, and maintenance planning using optimization
Therefore, multi-objective optimization is a more suitable approach for techniques. Most of these approaches are focused on individual bridge
tackling such a problem by making a trade-off between different sections (single facility), while few consider maintenance planning for
objective functions [41]. an entire network level.
Many agencies use BMS to develop maintenance plans by allocating
the budget, which is still mainly driven by subjective ranking and
preferences of domain experts. Most of these systems typically employ 2.1. Bridge condition assessment and prediction using deterioration
single-objective optimization analysis to allocate budget, but do not modeling
consider other aspects. Therefore, the purpose of this project is to design
a comprehensive bridge maintenance planning framework (BMPF) Models for bridge deterioration have tremendous potential to
capable of predicting the optimal intervention schedule for a bridge enhance bridge repair decision-making processes. Several deterministic
network. The framework will contain an impact assessment and pre­ and stochastic models for deterioration modeling were developed on the
diction model for bridge performance, a multi-attribute utility theory based on of mathematical and statistical concepts. For example, Ferreira
ranking model, and a multi-objective genetic algorithm optimization et al. [16] used a probabilistic approach on sectional pavement con­
model to identify the optimal balance between bridge performance and nections to improve pavement management. The method used a non-
budget constraints. homogeneous discrete Markov chain to forecast the future conditions
of the pavement of a particular pavement network. Durango and
2. Literature review Madanat [14] presented two adaptive control formulations that
included uncertainty in characterizing a facility’s deterioration rate in
In general, maintenance planning and optimization literature studies the process of developing maintenance and repair policies. Agrawal
can be divided into three groups: reliability assessment based on et al. [5] introduced the development of Markov chains and Weibull
distribution-based methods for measuring the deterioration rates of

Fig. 1. Bridge maintenance planning framework.

2
H. Jaafaru and B. Agbelie Automation in Construction 141 (2022) 104460

different bridge components in the state of New York. Wang et al. [47] aspects of BMS which includes bridge condition assessment and pre­
proposed a methodology for developing a model for deterioration using diction using deterioration modeling, bridge ranking using comparisons
the Bayesian network which helps to update the model as new data of different maintenance alternatives, and bridge maintenance using
become available. Liu and Zhang [34] developed a data-driven approach optimization models for maintenance planning.
that enables prediction of future conditions of highway bridge compo­
nents by training a Convolutional Neural Network (CNN) model. Dete­ 3. Methodology of Bridge Maintenance Planning Framework
rioration modeling enables the forecast and assessment of bridge (BMPF)
conditions therefore making it essential for optimal planning of main­
tenance activities and estimating the cost. This section introduces the methodology of the Bridge Maintenance
Planning Framework (BMPF) which seeks to develop optimal mainte­
2.2. Bridge ranking using comparisons of different maintenance nance plans for a network of bridge infrastructure assets. The proposed
alternatives bridge maintenance framework incorporates four modules; i) an impact
assessment module to decide on the type of maintenance intervention
An effective BMS should rank treatments according to their urgency and its resulting impact ii) a Machine Learning performance prediction
and importance in order that the most critical objectives are met based module for bridge condition assessment and forecasting iii) Multi-
on funding limitations. Recent studies indicate an increasing trend in the Attribute Utility Theory (MAUT) module for ranking of bridges based
application of MCDA (Multi-Criteria Decision Analysis) methods for on preference uncertainty and risk attitude of decision-makers, and
optimization of resources, strategies, and interventions. A few methods finally iv) an Evolutionary Optimization Algorithm (EA) module to
of MCDA that have particularly gained attention in this regard include develop optimal bridge maintenance plans. The complete bridge main­
Pareto Front, MAUT (Multi -Attribute Utility Theory), AHP (Analytical tenance planning framework is presented in Fig. 1.
Hierarchy Process), MAVT (Multi Attribute Value Theory), Goal pro­
gramming, ELECTRE (Elimination and Choice Expressing Reality) and
3.1. Problem formulation (optimization equations)
TOPSIS (Technique for Order by Similarly to Ideal Solution). Dabous
and Alkass [11] proposed a rational decision-making technique for
This study’s principal objectives are to maximize the performance
bridge management using a modified AHP method to evaluate and rank
condition level of bridges and to minimize the maintenance cost by
bridge rehabilitation strategies. Bukhsh et al. [8] applied AHP, MAUT,
optimally planning the maintenance treatments. The multi-choice part
and ELECTRE III on the maintenance planning decision problem of 22
of the problem is that it is necessary to select exactly one candidate of
bridges from the Netherlands road network. Their findings show that the
maintenance activity for each bridge, where the do-nothing alternative
MCDA methods yielded similar bridge ranking results despite the
is included in the candidate set. The multi-dimensional part of the
inherent differences in their methodologies. For more on ranking
problem is that there are several constraints, such as budget and per­
methods see ([12,20]).
formance goals. This can be viewed as a knapsack problem, which is a
combinatorial optimization problem used to illustrate both problem and
2.3. Bridge maintenance using optimization models
solution. Given a set of items with specific weights and values, the aim is
to get as much value into the knapsack as possible given the weight
There is extensive literature on the topic of pavement and bridge
constraint of the knapsack. The multi-objective aspect of the problem
maintenance optimization. Maintenance planning has historically been
refers to the use of more than one criterion in the objective function
carried out using a single-objective optimization process. Conventional
[39]. First, some variables were described and then the formulation was
single-objective optimization techniques such as linear programming,
constructed to explain the process of multi-objective bridge network
dynamic programming and integer programming have been widely used
optimization.
in pavement literature [22,27]. Wang et al. [48] proposed a bi-objective
Let Ujk be equal to the utility associated with project j for bridge k. In
integer programming (IP) model to address the challenges relating to
general, this utility is a function of various performance indicators, such
maintenance and rehabilitation (M&R) scheduling. One of the objectives
as:
of this model is to maximize the effectiveness of pavement maintenance
( )
treatments while minimizing the maintenance costs incurred by the Ujk = f Cjk , GCRjk (1)
user. However, the exact solution methods for such a bi-objective IP
model only works well for simple and homogeneous network-level where
management problems. Amin et al. [49] proposed an optimized deci­ Cjk = Cost function to compute cost of maintenance activity j for
sion support system for pavement M&R operations guided by bridge k,
geo-physical risk and community vulnerabilities. The study applied GCRjk = General condition rating for bridge k at the end of the
Linear Programming (LP) to optimized M&R strategies that ensure good program period if maintenance activity j is implemented.
pavement condition for all roads at a minimum budget. In the LP Depending on the formulation of the problem, the utility can be
approach, pavement sections with similar characteristics were grouped defined as a function of only a subset of the performance measures,
together. In this respect, the computational effort to solve this type of which can be chosen by the decision maker.
problem is much lower than that of their IP counterparts. However, the Let:
weighting method of the LP approach does not guarantee a complete set n = Number of bridges in the network.
of optimal solutions. Researchers have made several modeling and so­ Ujk = utility function associated with maintenance activity j for
lution efforts to address the bi-objective M&R scheduling issue. For bridge k.
example, Durango and Madanat [15] presented a quasi-Bayesian opti­ Cjk = Cost function to compute cost of maintenance activity j for
mization model to find joint inspection and maintenance policies for bridge k,
infrastructure facilities under performance model uncertainty. Liu et al. GCRjk = General condition rating for bridge k at the end of the
[35] suggested using a genetic algorithm (GA) to solve a model of program period if maintenance activity j is implemented.
bi-objective optimization for bridge deck rehabilitation. For more on GCRmin = General condition rating threshold specified by the deci­
GA-based approaches see Goldberg and Holland [23], Morcous and sion maker.
Lounis [38], Lee and Kim [33]. The main difference of the current work Budget = Available Funds for maintenance activities.
when compared to previous works discussed in the literature review is Xjk = 1 if maintenance activity j is selected for bridge k, and
that this project is a comprehensive framework that combines all the 0 otherwise.

3
H. Jaafaru and B. Agbelie Automation in Construction 141 (2022) 104460

Table 1 define the respective maintenance activities, their impact, and associ­
Maintenance treatments based on bridge conditions. ated condition ranges presented in Table 1. Maintenance treatment ac­
Bridge Treatment Treatment details Condition impact tivities are divided into four categories including monitoring/ do
condition intervention nothing alternative, preventive maintenance, medium maintenance, and
rating (BCR) major maintenance. Bridges in an excellent condition do not need any
7–9 Monitoring/ Do Monitoring and 0 maintenance treatment except periodic monitoring and inspection to
Nothing Inspection determine the performance of the bridge. Preventive maintenance is
6–7 Preventive Recoating, lubricating 2 defined as distress maintenance activities such as crack sealing,
Maintenance bearings, sealing deck
joints and cracks
recoating rails and patching performed with the primary objective of
5–6 Medium Deck patch and repairs, 3 slowing the rate of deterioration and maintaining good bridge condition.
Maintenance bearing restoration, Medium maintenance is defined as activities such as deck patch and
retaining wall repair, repairs, surface treatments applied to the entire bridge pavement sec­
guardrail repair
tions. Medium maintenance includes activities carried out to correct the
0–4 Replacement / Concrete deck, Excellent condition
Major substructure or after major defects and integrity of a bridge, for example, corrective maintenance.
Maintenance superstructure repair, rehabilitation or While major maintenance actions include reconstruction and structural
complete replacement replacement overlays of the entire bridge section. It is assumed that the GCR score
becomes 9 after replacement.
and
3.2.2. Performance indicators
Lk = set of candidate maintenance activity considered for bridge k.
This section provides a procedure to quantify the impact of mainte­
∑n ∑
Maximize Ujk Xjk (2) nance activity on different aspects by quantifying performance in­
dicators in monetary terms such as maintenance cost and user delay cost.
k=1
j∈Lκ

Subject to.
a) Maintenance cost: Maintenance costs are a monetary value that an
∑n ∑
Cjk Xjk ≤ Budget (3) agency bears because of maintenance activities. To obtain the unit
maintenance cost per m2 of a bridge, the maintenance cost per type of
k=1
j∈Lκ

( )∑ ∑ treatment activity is multiplied by the amount of material used,


1 n
GCRjk Xjk ≥ GCRmin (4) which is then divided by the bridge’s deck area size (m2). The
n k=1
j∈Lκ maintenance cost for the selected maintenance intervention is pro­
vided at element level, which is then summed up to represent the cost

Xjk = 1, k = 1, 2, …, n (5) for a bridge as a whole. Therefore, the maintenance cost of each
j∈Lk bridge is computed as a sum of the unit costs of all the maintenance
treatments multiplied by the treated area of a bridge. The formula to
Xjk ∈ {0, 1} k = 1, 2, ⋯, n, j ∈ Lk (6) calculate the maintenance cost is provided as follows [8]:
The objective function Eq. (2) is to maximize the bridge network ∑i=n
MC = UCAi × A (7)
benefit, which is a function of multiple performance indicators. The first i=1

constraint Eq. (3) is the budget constraint, which simply states that the
where MC is total maintenance cost of the bridge, UCAi is the unit cost
total cost of all the selected maintenance activity candidates must be less
maintenance activity per element i, n is the number of selected bridge
than the budget available. The second constraint Eq. (4) is the non-
elements i and A is the deck area of a bridge (m2).
budget constraint which states that the performance condition of a
bridge in the network should be above a certain threshold. The non-
b) User delay cost: The user delay cost estimates the impact of a
budget constraint in this study is the General condition rating of a
maintenance treatment on the availability of the bridge and un­
bridge in the network to be at least GCRmin. The third constraint Eq. (5)
derlines its importance on the network. It represents the value of
specifies that exactly one candidate must be selected for any bridge,
extended travel time of the road users due to work-zones in a mon­
where do-nothing is included as one possible alternative while Eq. (6)
etary form. The formula to calculate the user delay cost (UDC) are
states 1 if maintenance activity j is selected for bridge k, and 0 otherwise.
expressed as follows [45]:
UDC = ETT × ADTh × Voh × Dh (8)
3.2. Bridge performance and quantification of maintenance activities
L L
ETT = − (9)
This section introduces a method to link bridge conditions with vr vs
various maintenance treatments and provides a procedure to quantify
the socio-economic impacts of maintenance activities. where ETT is extra travel time due to speed restrictions, ADTh is average
daily traffic per hour passing over the bridge on a working day, Voh is
3.2.1. Condition state and maintenance treatments value of an hour of the user time, Dh is duration of the maintenance
The National Bridge Inventory (NBI) General Condition Ratings activity, L is length of working zones (m), vs is standard velocity (m/h)
(GCR) are used to describe the bridge in relation to the condition as and vr is reduced velocity due to maintenance.
constructed. The materials used in the bridge, as well as the physical
condition of the deck, superstructure, and substructure elements, are 3.3. Performance prediction using machine learning techniques
considered. This information is used to determine GCR on a numerical
scale ranging from 0 (failed condition) to 9 (excellent condition) as It is essential to forecast asset performance in order to plan the
described in the FHWA Coding Guide [19]. The observed damages and maintenance activities optimally and to estimate the costs. The deteri­
their severity define the condition state of a bridge and consequently the oration models can be used to examine the impact of certain parameters
required maintenance treatments. Table 1 provides a simplified rule- on the deterioration of bridges, such as material properties, the bridge
based choice of maintenance treatments related to the General Condi­ environment, and daily traffic ([4,21];). In this study, Machine Learning
tion Rating (GCR). The following studies [9,36,38] were considered to methods are applied for the prediction of bridge performance

4
H. Jaafaru and B. Agbelie Automation in Construction 141 (2022) 104460

stakeholders and/or decision-makers value differently is adopted in this


study. AHP, MAVT, MAUT, TOPSIS are all examples of MCDA methods
that provide the relative ranking of all the alternatives under consider­
ation based on the preference structure of the decision-maker. The
application of MAUT for multi-objective maintenance planning in
transportation infrastructure is not fully explored [8]. Therefore, this
study aims to explore and incorporate the potential of MAUT for
network level bridge maintenance planning.

3.4.1. Uncertainty and preference assessment using MAUT


This module uses Multi-Attribute Utility Theory (MAUT) principles
to incorporate experts’ preference for uncertainty and risk attitude for
bridge maintenance planning. The concept of the utility function is
based on gambling, where a gambler or decision maker must take
certain risks with the same probability of obtaining the best value or the
worst value [29,30]. The MAUT ranking model first accommodates
multiple performance indicators, then allows decision-makers to state
their maintenance preferences in uncertainty. The model then assigns a
rank to each bridge by performing a trade-off of multiple performance
objectives. The resulting prioritization filters-out a wide range of bridges
in the bridge network that do not meet the objectives and, therefore,
should not be scheduled for maintenance in the near future. In the
following, a brief explanation and application steps of the MAUT process
is provided:

Fig. 2. Machine learning model framework. a) This study uses the Exponential Utility Function (EUF) to obtain the
utility scores of each attribute, as it captures a decision maker ‘s
conditions. The ensemble machine Learning technique utilized in the preferential uncertainty and risk tolerance [30]. The formula for
present study falls under stacked generalization of combining multiple calculating the utility scores Ua of each attribute a is given below for
models together. This method is a loss-based learning method which the alternatives x:
calculates the optimal combination of a pool of prediction algorithms. ( )
The optimal combination minimizes the cross-validated risk (error)
− xa

(10)
RT
Ua (xa ) = A − B × e
during the training phase of the multiple models. The ensemble machine
learning technique minimizes errors on the predictions by correcting Where:
mistakes on the predictions made by the individual base models. If in­ ( )
dividual base models produce similar mistakes, then combining base − min(xa )
RT

models is worthless. There are two techniques such as bagging and A= (


e
) ( ) (11)
boosting which use various resampling methods to achieve diverse base [ − min(xa ) − max(xa )
]
models [46]. This method of machine learning is important to predictive
RT RT
e − e
models as it decreases variance, bias, or improves predictions. Fig. 2
shows a simple representation of the stacked generalization framework. 1
The stacked machine learning model combines multiple classifica­ B= ( ) ( ) (12)
tion models through a meta-classifier (fusion model). The individual [ ]
− min(xa ) − max(xa )
RT RT
e − e
classification models are trained based on the complete training dataset
then the meta-classifier is fitted based on the outputs of the individual
classification models in the ensemble. The meta-classifier can either be RT = [
− CEi
] (13)
trained on the predicted class labels or probabilities from the ensemble. ln − 0.5Ua (max(xa ) )− 0.5Ua (min(xa ) )+A

The main idea behind the structure of a stacked model is to use one or
B

more first level models, make predictions using these models and then
where Ua is a utility score, A and B are constant scaling, RT is the risk
use these predictions as features to fit one or more second level models
tolerance and e is an exponential constant.
on top. To avoid overfitting, cross-validation is usually used to predict
When computing EUF, the scaling constants and RT are cyclically
the OOF (out-of-fold) part of the training set. The final prediction in this
dependent. Eq. (14) can be used with the goal seeker function of MS
variant is based on the mean or mode of all the predictions. The
Excel [37] to solve this.
ensemble machine learning models built in this research were Random
Forest, Support Vector Machines, Artificial Neural Network and − max(xi ) − min(xi )
(14)
CE
eRT = 0.5e RT + 0.5e RT
Xgboost.
The EUF requires decision-makers to respond to the lottery question
3.4. Bridge ranking using Multi-Criteria Decision-Analysis (MCDA) of the maximum and minimum value of the attribute, where the indif­
ference between the best and the worst possible outcome must be
Effective decision-making in maintenance treatment requires a pri­ reached. The indifference point for the decision-maker is referred to as
oritization ranking framework. Prioritization is carried out by deter­ the certainty equivalent (CE). Risk tolerance is calculated on the basis of
mining the safety risk, cost effectiveness, budget availability, and the the expected value (EV) and CE, where EV is the mean of the worst and
order in which various pavement or bridge sections are scheduled to be best value of the attribute values. The CE value is chosen by experts on
maintained [3]. Therefore, a Multi Criteria Decision Analysis (MCDA) the basis of the principle [31]:
which is a general framework for supporting complex decision-making Risk Neutral, if EV = CE (Linear shaped)
situations with multiple and often conflicting objectives that Risk Avoiding, if EV ≥ CE (Concave shaped)

5
H. Jaafaru and B. Agbelie Automation in Construction 141 (2022) 104460

Algorithm (GA) optimization model is used in this study. Developed by


John Holland in 1970, GA is a combinatorial optimization search tech­
nique inspired by Darwinian evolutionary theory of natural selection,
genetics and survival-of-the-fittest [23].
GAs is attractive to be applied in engineering optimization tasks due
to two reasons. First, a GA evaluates many points in the search space
simultaneously as opposed to a single point, thus reducing the chance of
converging to the local optimum. Second, a GA uses only values of
objective functions; therefore, they do not require the search space to be
differentiable or continuous.
The idea behind GA is to initially generate and then iteratively
enhance a population of solutions with the help of variation (crossover,
mutation) and selection (parents, survivor) operators. Eventually, a
near-optimal solution will be obtained by examining a relatively small
number of possible solutions. The objective functions must evaluate
each individual (single maintenance treatment) of the population (all
maintenance treatments) in order to find optimum solutions. Further
stochastic operators i.e., cross-over, the mutation is applied on selected
populations based on their relative fitness score assigned by objective
function to generate the next population.
Non-dominated sorting genetic II (NSGA-II) was utilized in this
study. NSGA-II has three special characteristics, fast non-dominated
sorting approach, fast crowded distance estimation procedure and sim­
ple crowded comparison operator [13]. In NSGA-II, the population is
ordinarily initialized and once the population is initialized, it is sorted
based on the non-domination into each front. A non-dominant member
means that there is no other member in the population which has better
value than this member on both objectives. The advantage of NSGA-II is
that the diversity among non-dominated solutions is maintained using
the crowding procedure (no extra diversity control is needed), and
elitism protects an already found optimal solution from being deleted.
Where elitism means that the fittest of individuals are guaranteed a
place in the next generation without undergoing mutation. Fig. 3 pro­
vides the general overview of GA.

4. Case study

Fig. 3. Genetic algorithm framework. The proposed BMPF is validated on a case study of concrete and steel
bridges located in Washington DC. Historic bridge inventory data
(1992–2019) was obtained from the NBI database. The inspection re­
Risk Taking, if EV < CE (Convex shaped)
cords in the NBI database are annually provided by the Federal Highway
Administration (FHWA) and are freely available on the FHWA home­
b) Trade-offs are made between attributes by assigning the relative
page. The formatting of the files is defined in the Recording and Coding
importance based on the preferences of the decision-maker.
Guide for the Structure Inventory and Appraisal of the Nation’s Bridge.
rate(xa ) The key features included in this study are Structure Name, Structure
ka = ∑A (15)
a=1 (xa )
Type, Latitude, Longitude, Year Built, Structure Length, Design Load,
Roadway Width, Annual Daily Traffic (ADT), Percent Daily Trucks,
Where ka represents attribute’s relative importance, rate (xa) is a Degrees Skew, and Horizontal Clearance. Bridge inspection interval data

weight assigned by a decision maker to attribute a, and A
a=1(xa) is the is important because the condition of a certain bridge may deteriorate or
sum of all the weights given to attributes. improve depending on whether maintenance activity has been done or
not during the inspection interval. Since bridge inspection is usually
c) The utility score of each attribute and its assigned weight factors are performed every 2 years, all bridge IDs which show an average inspec­
aggregated to generate the final ranking of the alternatives by tion interval >2 years were removed from the data. This will ensure that
computing additive aggregation as follows: bridge conditions are accurately represented in the dataset. The data
∑n
U(x) = ka Ua (xa ) (16) was screened several times to remove missing or incomplete and ques­
i=1
tionable data entries. The distribution of bridge types and material of the
where ka is the relative importance of attribute, Ua(xa) is a single main structural elements can be visualized in Fig. 4. Only concrete and
attribute utility for each attribute a for an alternative x. steel structures were applied to this study and the dataset for Wash­
ington DC have 60 Steel, 19 Continuous Steel, 7 Prestressed Steel, 2
Concrete, 7 Continuous Concrete, and 7 Prestressed Concrete bridge
3.5. Maintenance planning using genetic algorithms structural type making a total of 95 bridges in the dataset considered.
For feature descriptions and statistics see Table A.1 in the appendix.
The development of the maintenance plan is a multi-objective opti­ A short-term maintenance plan (5 years period) is programmed with
mization problem, with the aim to reduce the life cycle costs of assets the aim to keep the network-level bridge condition rating score GCR of
and fulfill the performance requirements over a specified planning ho­ at least 6 (GCRmin = 6) with an estimated budget of $50 million. The
rizon. To get the optimal maintenance planning strategy a Genetic estimated budget was chosen based on domain knowledge for the

6
H. Jaafaru and B. Agbelie Automation in Construction 141 (2022) 104460

Fig. 4. Distribution of bridges by structural type.

Table 2
Calculated maintenance and user delay cost.
structure ID number AGE GCR Maintenance cost User delay cost

078 36 5 771,168 14.75


1108 59 6 3,272,279 155.30
0250 15 6 3,786,504 193.06
0550 10 6 4,801,806 385.39
1140(MASS AVE) 52 6 3,018,088 284.44
1140 52 6 2,707,569 271.73
251 15 6 2,113,224 290.59
1401 50 6 5,236,114 147.70
1200(LITTLE R.) 56 5 8,330,749 240.03
1026 60 6 3,282,024 179.83

purpose of this study. Although budgets are usually carried out annually,
in this study the budget is used freely throughout the planning period.
The goal is to develop a rational bridge maintenance planning frame­
work to predict the most appropriate intervention schedule for bridges
where the system-level structural reliability is maintained with the
Fig. 5. Bridge GRC condition distribution.
minimum possible cost during the bridge lifecycle.

5. Discussion and results bridge was determined as the minimum GCR score of the deck, sub­
structure, and superstructure condition score. Table 2. Presents a sample
5.1. Computing performance indicators of the calculated maintenance and user delay cost for some of the
bridges in the case study.
To account for the socio-economic impact of maintenance activities Bridges in poor conditions require considerable higher maintenance
on agencies and bridge users, maintenance costs and user delay costs cost compared to bridges in relatively better condition. Therefore, it is
were calculated for all bridges in the case study. The approximate cheaper to execute maintenance action on time rather than wait until a
maintenance costs for each bridge were calculated using Eq. (7) while bridge reaches a higher level of damage. Fig. 5 shows the condition
the user delay cost has been calculated using Eq. (8). The GCR of each distribution of the bridges found in the dataset of the case study.
Most of the bridge’s GCR score ranges from 5 to 7, with the exception

7
H. Jaafaru and B. Agbelie Automation in Construction 141 (2022) 104460

Table 3 Support vector machines (SVM): The idea behind SVM is to find a
Machine learning model parameters. separation line that splits the data between two different clusters in a
Model Parameters training dataset. This line separates the clusters of data in such a way
that the individual clusters are farthest away from the line. A new test
Random Forest bootstrap = True, criterion=’gini’, min_samples_leaf = 1,
min_samples_split = 2, n_estimators = 300 data point is predicted depending on where it will be placed on either
Support Vector C = 1, decision_function=’ovr’, degree = 3, side of the line. The determination of this line is done through quadratic
Machine (SVM) gamma=’auto’, kernel=’rbf’, programming and resembles an optimization problem [50].
MLP-NN Activation = relu, hidden_layer = 2, random_state = 1, Artificial Neural Net: Artificial Neural net is a network of simple units
learning_rate_init = 0.001, solver=’lbfgs’
Xgboost Gamma = 0.49, learning_rate = 0.58, max_depth = 8.82,
called neurons which are subjected to raw inputs. Based on the inputs,
n_estimators =109.98 the neurons then vary their internal state and produce an output. A
network, in the form of a directed weighted graph, is formed by con­
necting the outputs to the inputs. The internal aspects of the learning
of a few bridges with scores greater than 7 and lower than 5. This limited process of neural nets are usually optimized through a process called
dispersion of the GCR score in the data set indicates the overall learning where researchers have applied methods such as fuzzy logic,
acceptable condition of the bridges in the network. As a result of Bayesian method, and genetic algorithms [40].
computing performance indicators, the bridges having GCR score higher Extreme gradient boosting (Xgboost): Xgboost is an optimized distrib­
than 7 were eliminated from the analysis, meaning they are in excellent uted gradient boosting model designed to be highly efficient, flexible,
condition and do not need maintenance in the near future. and portable. The general concept of gradient boosted trees is that
initially the model predictions are made with simple tree structures and
later on new tree structures that are created to predict the errors of prior
5.2. Performance prediction using machine learning techniques
models. These are subsequently added together to make the final pre­
dictions. The term gradient boosting comes from the fact that the loss
In developing the bridge assessment and performance condition
minimization during addition of the new models is achieved using a
prediction model, four different machine learning algorithms were
gradient descent algorithm [10,24,25].
considered. Meanwhile, to find the optimal parameters for the base
The machine learning model consists of stacking the output of indi­
models of the machine learning model a grid search parameter optimi­
vidual base models and using another classifier model to compute the
zation method was utilized. A brief description for each base model is
final prediction. Stacking allows us to use the strength of each individual
provided below:
classifier by using their output as input of a final classification model.
Random Forest: Random Forest is a supervised machine learning
The fine-tuned models selected for the base models are Random Forest,
model that is based on multiple decision trees. It is an ensemble learning
SVM, and Artificial neural network. The Xgboost model was used as the
algorithm that generates a set of decision trees from a randomly selected
meta-classifier (fusion model) to combine the models as discussed
subset of the training set. It then aggregates the votes from different
above. Table 3 presents the fine-tuned parameters of the machine
decision trees to decide the final class of the test object [7,26,32].

Table 4
Machine learning model results and classification report.
Models Training ROC-AUC Validation ROC-AUC

Accuracy Precision Recall F-1 Score Accuracy Precision Recall F1-Score

Random Forest 0.9185 0.93 0.91 0.90 0.8699 0.8241 0.81 0.82 0.82 0.7902
SVM (rbf) 0.7846 085 0.80 0.81 0.7799 0.7241 0.80 0.80 0.79 0.7307
MLP-NN 0.8308 0.85 0.83 0.82 0.7301 0.7062 0.69 0.69 0.69 0.6524
Stacked Model 0.8901 0.91 0.90 0.90 0.8839 0.8421 0.84 0.84 0.83 0.7864

Fig. 6. Feature importance.

8
H. Jaafaru and B. Agbelie Automation in Construction 141 (2022) 104460

Fig. 7. Multi-attribute utility theory bridge ranking.

learning base models while Table 4 presents the results and classification in the network in ascending order where multiple alternatives were
reports of the machine learning bridge performance model. considered. Bridges are ranked on the basis of a minimized aggregated
The classification report evaluates the performance of the machine utility function where the bridge is ranked higher with a minimum
learning model. Following the classification report, precision is the condition rating, maintenance cost, and user delay cost.
measure of accuracy provided that a specific class has been predicted. It is important to note that with the application of MAUT, the ulti­
The recall is the percentage of positive labeled instances that were mate goal is not just the ranking of alternatives but also to turn the
predicted as positive while the f1- score is the combination of precision subjective decision-making process into a more rational way in which
and recall. The higher the figure of precision and recall the better the the principles of utility reflect the interests of decision-makers. Fig. 7
model performs. The confusion matrix compares the actual target values presents the MAUT ranked bridges where the MAUT scores are depicted
with those predicted by the machine learning model. Receiver operating in the y-axis while the Bridge ID are depicted along the x-axis.
curve area under the curve (ROC-AUC) is a performance measurement Bridge ID 078 with the GCR score of 5 and MAUT score of 95 is
for the classification at various thresholds settings (ROC-AUC score ranked first according to the MAUT ranking. The computed MAUT score
above 0.5 is considered good). The machine learning models were able establishes the ranking of each alternative. Based on the minimization
to perform condition prediction on the bridge network in the dataset objective function, the alternatives with lower aggregated MAUT scores
with accuracies ranging from 70.62% to 84.21%. The precision and will obtain a higher ranking. The rated attributes are relative to each
recall values of the Stacked model are the highest with accuracies of 0.84 decision-maker preference, therefore the importance of these attributes
for both while the F-1 score is 0.83 when compared to the rest of the in maintenance planning can vary from one decision-maker to another
classification models. [8].
Another objective of the machine learning predictions model is to
assess important information on those features influencing bridge per­ 5.4. Implementation of genetic algorithms optimization
formance conditions. To achieve this goal, the model provides a
weighted score that indicates how useful or valuable each feature was in A Genetic Algorithm (GA) optimization model was developed to seek
the construction of the model. The more an attribute or feature is used to an optimal maintenance plan for the bridges in Washington DC ac­
make key decisions with the model, the higher its relative importance. cording to MAUT ranking. Distributed Evolutionary Algorithm in Py­
Fig. 6 presents the feature importance where the influencing variables thon (DEAP), which is a novel evolutionary computation framework for
are shown in the y-axis while x-axis shows the numerical importance/ rapid prototyping and testing of engineering ideas was utilized in this
weights. It appears that Age and Design load are the two most important research for the GA application of bridge maintenance planning.
features affecting bridge performance conditions. The formulation of the optimization model is carried out in such a
way that a cost-effective maintenance strategy is reached by preserving
5.3. Multi-attribute utility theory model the performance level of the bridge network at a desirable level (GCR
score of 6 or higher). So, the objective functions are bridge performance
The first step of MAUT is to identify the single utility function (SUF) condition maximization and maintenance cost minimization.
of each attribute across all bridges. The normalized importance weights Based on the objective functions, NSGA-II is applied to the MAUT
are then calculated by using Eq. (15) while a global aggregated score ranked bridges in the dataset. The model was set-up with a population
was computed for each alternative by using Eq. (16). Calculating the size of 50 individuals with 100 generations (iterations). A partially
SUF value for each attribute is a lengthy operation. As a result, the matched crossover method with a 0.8% probability was used, where two
computation method is removed here for simplicity. For a thorough individual chromosomes are randomly swapped between the parents to
description of the process, see [8] see Figs. A.2 and A.3 in the appendix generate two unique children (i.e., maintenance plans). Finally, the
for the utility plots. The global aggregated score ranks all the 95 bridges mutation probability of 0.5 was applied in which adaptive mutation was

9
H. Jaafaru and B. Agbelie Automation in Construction 141 (2022) 104460

Table 5 adopted so that the probability can be adjusted during optimization. The
Optimal bridge maintenance plan. optimization model parameter was tuned several times by adjusting the
Bridge ID Before Maintenance Maintenance After population size, crossover, and mutation probabilities to obtain the best
condition treatment cost condition possible configuration for the given problem.
1140 (Mass 6 Medium 3,018,088 9 The optimal maintenance plan as described in Table 5 was chosen
Ave) Maintenance out of the 50 maintenance plans (individuals) generated in 100 itera­
1203 6 Medium 1,288,812 9 tions (generations). The plan includes a set of bridges along with the
Maintenance name of the treatment, maintenance costs and improved condition.
1108 6 Medium 3,272,279 9
Maintenance
1304 5 Medium 1,802,899 8 6. Conclusions
Maintenance
1139 7 Preventive 330,789 9
The effective management of bridges covers the processes of data
Maintenance
1200 5 Medium 8,330,749 8
collection, condition assessment and prediction, planning, optimization,
(LITTLE Maintenance and application of maintenance treatments. This paper describes an
R.) extensive framework for designing effective maintenance plans for a
0070 6 Medium 2,157,737 9 bridge network while taking performance and financial constraints into
Maintenance
account. The proposed Bridge Maintenance Planning Framework
1009 5 Medium 1,515,402 8
Maintenance (BMPF) uses a Machine learning model for bridge condition assessment
0503(EB) 7 Preventive 355,959 9 and prediction. Multi-attribute utility theory which is a method of
Maintenance MCDA (Multi-Criteria Decision Analysis) was explored to rank the
0251 5 Medium 2,113,224 8 bridges by capturing the decision-maker’s preference uncertainty and
Maintenance
1202 6 Medium 1,277,602 8
risk attitude. While maintenance planning optimization was achieved by
Maintenance using Non-sorting Genetic Algorithms II (NSGA-II).
1001 3 Major 845,118 8 The framework was validated on an NBI data inventory of concrete
Maintenance and steel bridge networks in Washington DC, given performance and
1012 5 Medium 2,062,400 8
budget constraints. Based on the criteria in question, out of the 95
Maintenance
1109 7 Preventive 3,534,885 9 bridges, a group of 19 bridges was found to be optimal for maintenance
Maintenance treatment planning. Fig. 8 shows an example of before and after a bridge
1200 5 Medium 2,065,623 8 maintenance treatment is applied.
(RAMP Maintenance Fig. 8 depicts how a simple preventive treatment (painting) can
C)
1017(SB) 5 Medium 678,734 8
improve the condition of a bridge. The BMPF is generic and may include
Maintenance any other performance attributes or features, depending on the re­
1000 7 Preventive 365,389 9 searchers’, engineers’, or agency’s needs. Therefore, the proposed
Maintenance framework can enable transportation agencies to execute different
1200 5 Medium 14,747,810 8
maintenance planning scenarios by changing bridge network budget
(MAIN) Maintenance
544 7 Preventive 2,065,623 9 limits and performance targets such as bridge condition threshold.
Maintenance Additionally, by predicting bridge conditions, the framework can be
used for future budget planning. The most important function of a bridge
Number of Maintained Bridges: 19 Total Maintenance Cost: $49,995,042.
Remaining Budget: $4958 Average Condition: 8.5.
maintenance agency, engineer, or supervisor may be bridge mainte­
nance planning. The proposed Bridge Maintenance Planning Framework

Fig. 8. Before bridge maintenance painting (left) and after (right).

10
H. Jaafaru and B. Agbelie Automation in Construction 141 (2022) 104460

may pay off in a variety of ways, including increased production rates on [15] Pablo L. Durango, Samer M. Madanat, Optimization of inspection and maintenance
decision for infrastructure facilities under performance model uncertainty: a quasi-
individual bridge projects, reduced “down time” between projects, more
Bayes approach, Transp. Res. A Policy Pract. 42 (8) (2008) 1074–1085, https://doi.
efficient use of “down time” for training workers, and improved bridge org/10.1016/j.tra.2008.03.004.
inventory condition. [16] A. Ferreira, A. Antunes, L. Picado-Santos, Probabilistic segment-linked pavement
Future studies should explore detailed representations of construc­ management optimization model, J. Transp. Eng. 128 (6) (2002) 568–577, https://
doi.org/10.1061/(asce)0733-947x(2002)128:6(568).
tion and material costs in their models. Further improvement to the [19] FHWA report, Bridge Preservation Guide – Maintaining a Resilient Infrastructure to
BMPF may consider other structural elements of bridges while defining Preserve Mobility, Trid.Trb.Org, 2018. https://trid.trb.org/view/1640085
specific maintenance treatment. A traffic flow study can also be imple­ [Accessed 6 June 2022].
[20] Dan M. Frangopol, Min Liu, Maintenance and management of civil infrastructure
mented to determine accurate user delay costs in work zones. Other based on condition, safety, optimization, and life-cycle cost, Struct. Infrastruct.
optimization methods such as particle swarm and hybrid meta-heuristics Eng. 3 (1) (2007) 29–41, https://doi.org/10.1080/15732470500253164.
optimization should be explored as they may perform better than ge­ [21] D. Frangopol, Y. Dong, S. Sabatino, Bridge life-cycle performance and cost:
analysis, prediction, optimization, and decision-making, Struct. Infrastruct. Eng. 13
netic algorithms. The machine learning model for performance predic­ (10) (2017) 1239–1257. Informa UK Limited, https://doi.org/10.1080/15
tion can also be improved by using different techniques to increase 732479.2016.1267772.
accuracy. Other forms of MCDA ranking methods such as AHP and [22] T. Fwa, W. Chan, K. Hoque, Multi-objective optimization for pavement
maintenance programming, J. Transp. Eng. 126 (5) (2000) 367–374. American
TOPSIS can also be explored as well. These improvements may present Society of Civil Engineers (ASCE), https://doi.org/10.1061/(asce)0733-947x
more accurate results and estimation of the required budget and per­ (2000)126:5(367).
formance objectives for the stakeholders. [23] D. Goldberg, J. Holland, Genetic algorithms in search optimization and machine
learning, Machine 3 (1998) 95–99. American Library Association, https://doi.
org/10.5860/choice.27-0936.
Declaration of Competing Interest [24] L. Guelman, Gradient boosting trees for auto insurance loss cost modeling and
prediction, Expert Syst. Appl. 39 (2012) 3659–3667, https://doi.org/10.1016/j.
eswa.2011.09.058.
None. [25] S. Haykin, Neural networks, a comprehensive foundation, Int. J. Neural Syst. 05
(04) (1994) 363–364, https://doi.org/10.1142/s0129065794000372.
[26] G.E. Hinton, R. Salakhutdinov, Reducing the dimensionality of data with neural
Acknowledgement networks, Science 313 (2006) 504–507. American Association for The
Advancement Of Science (AAAS), https://doi.org/10.1126/science.1127647.
Some or all data, models, or code that support the findings of this [27] Y. Huang, H. Huang, A model for concurrent maintenance of bridge elements,
Autom. Constr. 21 (2012) 74–80. Elsevier BV, http://10/1016/j.autcon.2011.05.0
study are available from the corresponding author upon reasonable 14.
request. [28] M.A. Hurt, S.D. Schrock, Highway Bridge Maintenance Planning and Scheduling,
1st edition, Butterworth-Heinemann, Cambridge, MA, 2016, pp. 289–310. ISBN-
10: 0128020695.
Appendix A. Supplementary data [29] R.L. Keeney, H. Raiffa, Decisions with Multiple Objectives Preferences and Value
Tradeoffs, Cambridge University press, New York, 1993. ISBN 0-521-44185-4.
Supplementary data to this article can be found online at https://doi. [30] R.L. Keeney, D. von Winterfeldt, M13 practical value models, in: W. Edwards,
R. Miles, D. von Winterfeldt (Eds.), Advances in Decision Analysis: From
org/10.1016/j.autcon.2022.104460. Foundations to Applications, Cambridge University Press, New York, 2007,
pp. 232–252, https://doi.org/10.1017/CBO9780511611308.
References [31] S. Krishnamurty, Normative decision analysis in engineering design, Decis. Making
Eng. Design 4 (4) (2006) 21–33.
[32] Y. Lecun, Y. Bengio, G.E. Hinton, Deep learning, Nature 521 (2015) 436–444.
[1] B. Agbelie, Economic Impacts of Transportation Infrastructure Investment across
Springer Science and Business Media LLC, https://doi.org/10.1038/nature14539.
Countries: An Empirical Analysis, Doctoral Dissertation, Purdue University, West
[33] C. Lee, S. Kim, GA-based algorithm for selecting optimal repair and rehabilitation
Lafayette, IN, 2013, https://docs.lib.purdue.edu/dissertations/AAI3604694/
methods for reinforced concrete (RC) bridge decks, Autom. Constr. 16 (2) (2007)
(Accessed 22 March 2022).
153–164, https://doi.org/10.1016/j.autcon.2006.03.001.
[2] B. Agbelie, Impact of highway nonuser revenue on equity ratio: exploratory
[34] H. Liu, Y. Zhang, Bridge condition rating data modeling using deep learning
analysis, J. Infrastruct. Syst. 27 (3) (2021) 04021017, https://doi.org/10.1061/
algorithm, Struct. Infrastruct. Eng. (2020) 1–14, https://doi.org/10.1080/
(ASCE)IS.1943-555X.0000631.
15732479.2020.1712610.
[3] B. Agbelie, S. Labi, J. Fricker, Y. Qiao, Z. Zhang, K.C. Sinha, Lifecycle decision
[35] C. Liu, A. Hammad, Y. Itoh, Maintenance strategy optimization of bridge decks
framework for steel bridge painting, J. Bridg. Eng. 22 (11) (2017) 1–6, https://doi.
using genetic algorithm, J. Transp. Eng. 123 (2) (1997) 91–100, https://doi.org/
org/10.1061/(ASCE)BE.1943-5592.0001140.
10.1061/(asce)0733-947x(1997)123:2(91).
[4] B.R. Agbelie, S. Labi, K.C. Sinha, Estimating the marginal costs of bridge damage
[36] A.G. Matin, R.V. Nezaf, A. Golroo, A comparative study on using meta-heuristic
due to overweight vehicles using a modified equivalent-vehicle methodology and
algorithms for road maintenance planning: insights from field study in a
in-service data on life-cycle costs and usage, Transp. Res. A Policy Pract. 95 (2017)
developing country, J. Traffic Transp. 4 (5) (2017) 477–486, https://doi.org/
275–288, https://doi.org/10.1016/j.tra.2016.11.008.
10.1016/j.jtte.2017.06.004.
[5] A. Agrawal, A. Kawaguchi, Z. Chen, Deterioration rates of typical bridge elements
[37] M. Middleton, Data Analysis Using Microsoft Excel: Updated for Windows 95,
in New York, J. Bridg. Eng. 15 (4) (2010) 419–429, https://doi.org/10.1061/(asce)
Wadsworth Publishing, Belmont, CA, USA, 1996. ISBN-10: 0534244262.
be.1943-5592.0000123.
[38] G. Morcous, Z. Lounis, Maintenance optimization of infrastructure networks using
[7] L. Brieman, Random forests, Mach. Learn. 45 (2001) 5–32, https://doi.org/
genetic algorithms, Autom. Constr. 14 (1) (2005) 129–142, https://doi.org/
10.1023/a:1010933404324.
10.1016/j.autcon.2004.08.014.
[8] Z.A. Bukhsh, I. Stipanovich, A.G. Doree, Network level bridge maintenance
[39] V. Patidar, S. Labi, K.C. Sinha, National Academies of Sciences, Engineering, and
planning using multi-attribute utility theory, Struct. Infrastruct. Eng. 15 (7) (2019)
Medicine 2007. Multi-Objective Optimization for Bridge Management Systems,
872–885, https://doi.org/10.1080/15732479.2017.1414858.
The National Academies Press, Washington, DC, 2007, https://doi.org/10.17226/
[9] Z.A. Bukhsh, I. Stipanovich, A.G. Doree, Multi-year maintenance planning
23147.
framework using multi-attribute utility theory and genetic algorithms, Eur. Transp.
[40] J. Rezaei, M. Shahbakhti, B. Bahri, A. Aziz, Performance prediction of HCCI
Res. Rev. 12 (2020) 3, https://doi.org/10.1186/s12544-019-0388-y.
engines with oxygenated fuels using artificial neural networks, Appl. Energy 138
[10] T. Chen, C. Guestrin, XGBoost, in: Proceedings of the 22nd ACM SIGKDD
(2015) 460–473, https://doi.org/10.1016/j.apenergy.2014.10.088.
International Conference on Knowledge Discovery and Data Mining, 2016.
[41] S. Rose, B. Mathew, K. Isaac, A. Abhaya, Risk based probabilistic pavement
[11] Saleh Abu Dabous, Sabah Alkass, Decision support method for multi-criteria
deterioration prediction models for low volume roads, Int. J. Pavement Eng. 19 (1)
selection of bridge rehabilitation strategy, Constr. Manag. Econ. 26 (8) (2008)
(2016) 88–97, https://doi.org/10.1080/10298436.2016.1162308.
883–893, https://doi.org/10.1080/01446190802071190.
[43] K.C. Sinha, S. Labi, B.R. Agbelie, Transportation infrastructure asset management
[12] A. de Almeida, R. Ferreira, C. Cavalcante, A review of the use of multicriteria and
in the new millennium: continuing issues, and emerging challenges and
multi-objective models in maintenance and reliability, IMA J. Manag. Math. 26 (3)
opportunities, Transportmetrica A: Transp. Sci. 13 (7) (2017) 591–606, https://
(2015) 249–271, https://doi.org/10.1093/imaman/dpv010.
doi.org/10.1080/23249935.2017.1308977.
[13] K. Deb, A. Pratap, S. Agarwal, T. Meyarivan, A fast and elitist multi-objective
[44] M. Volovski, E. Bardaka, Z. Zhang, B. Agbelie, S. Labi, K.C. Sinha, Indiana State
genetic algorithm: NSGA-II, IEEE Trans. Evol. Comput. 6 (2) (2002) 182–197,
Highway Cost Allocation and Revenue Attribution Study and Estimation of Travel
https://doi.org/10.1109/4235.996017.
by Out-of-State Vehicles on Indiana Highways, Joint Transportation Research
[14] Pablo L. Durango, Samer M. Madanat, Optimal maintenance and repair policies in
Program Publication No. FHWA/IN/JTRP-2015/12). West Lafayette, IN, 2015,
infrastructure management under uncertain facility deterioration rates: an
https://doi.org/10.5703/1288284315709.
adaptive control approach, Transp. Res. A Policy Pract. 36 (9) (2002) 763–778,
https://doi.org/10.1016/S0965-8564(01)00038-6.

11
H. Jaafaru and B. Agbelie Automation in Construction 141 (2022) 104460

[45] Y. Wang, P.M. Goodrum, Use of conceptual road user costs for a rapid roadway [48] F. Wang, Z. Zhang, R. Machemehl, Decision-making problem for managing
construction decision making system, in: Construction Research Congress 2005: pavement maintenance and Rehabilitation Projects, J. Transp. Res. Board 1853 (1)
Broadening Perspectives, 2005, pp. 1–11, https://doi.org/10.1108/ (2003) 21–28, https://doi.org/10.3141/1853-03.
09699980910938000. [49] S. Amin, U. Tamima, L.E. Amador-Jimenez, Optimal pavement management:
[46] Y. Zhang, A. Haghani, A gradient boosting method to improve travel time resilient roads in support of emergency response of cyclone affected coastal areas,
prediction, Transp. Res. Part C: Emerg. Technol. (2015) 308–324, https://doi.org/ Transportation Research Part A: Policy and Practice 119 (2019) 45–61, https://doi.
10.1016/j.trc.2015.02.019. org/10.1016/j.tra.2018.11.001.
[47] R. Wang, L. Ma, C. Yan and J. Mathew, Condition deterioration prediction of [50] V. Svetnik, A. Liaw, C. Tong, J. Culberson, R. Sheridan, B. Feuston, B. Random
bridge elements using Dynamic Bayesian Networks (DBNs), 2012 International Forest: A Classification and Regression Tool for Compound Classification and QSAR
Conference on Quality, Reliability, Risk, Maintenance, and Safety Engineering, Modeling. J. Chem. Inform. Comput. Sci. 43(6), (2003) 1947-1958,doi: 10.1021/
(2012) 566-571,doi: 10.1109/ICQR2MSE.2012.6246298. ci034160g.
[51] ASCE’s 2021 Infrastructure Report Card, Bridges (2022). https://infrastructurerep
ortcard.org/cat-item/bridges/.

12

You might also like