Expert Systems With Applications 186 (2021) 115743

Contents lists available at ScienceDirect

Expert Systems With Applications

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

Building condition assessment using artificial neural network and

structural equations
Ahmed Gouda Mohamed a, *, Mohamed Marzouk b
a
Faculty of Engineering, Construction Engineering and Management, British University in Egypt, Egypt
b
Construction Engineering and Management, Structural Engineering Department, Faculty of Engineering, Cairo University, Egypt

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

Keywords: Building facilities condition assessment is considered a fundamental aspect of an effective decision-making
Building condition assessment maintenance management plan to fulfill service requirements. A noticeable dearth of studies is believed to
Artificial neural network have delivered condition assessment approaches for existing buildings; however, these approaches are still
Structural equation modeling
deemed premature, with some limitations in demand enhancement. This paper presents a novel physical con­
Maintenance management
Facilities management
dition assessment framework for existing educational buildings that contribute to the body of knowledge by
offering a state-of-the-art approach incorporating an Artificial Neural Network (ANN) predictive model and a
Structural Equation Model (SEM). The ANN predictive model aims to forecast the future condition-rating states
for each facility component in various building spaces. Simultaneously, the SEM determines the proportionate
weights of building facilities components. The primary objectives of this paper are to prioritize building com­
ponents for maintenance purposes and record the potential effects of several parameters influencing the con­
dition state of building components. These objectives can be achieved via four sequential modules: 1) scan to BIM
module; 2) condition assessment prediction module; 3) proportionate weight determination module; and 4)
entire space rating value module. Condition-monitoring data on six different buildings’ internal components are
analyzed to anticipate their future condition. The components carried out are: 1) wooden flooring tiles; 2)
gypsum board ceiling tiles; 3) wooden doors; 4) wooden windows, 5) split air conditioner units; and 6) desktop
computers. The overall coefficient of determination (R2) of the developed ANN models for the predicted six
components conditions are 0.99, 0.99, 0.927, 0.88, 0.97, and 0.972, respectively.

1. Introduction tasks. Building facilities components degrade over time due to external
and internal influences; most external influences are apparently due to
Buildings constitute a fundamental urban infrastructure physical climatic conditions, while user-driven consequences and building
asset, including numerous facilities. There are three main space features maintenance process causes the internal influences.
associated with building sheltering from climate and offering customers In fact, the research problem presented in this study is that,
security, safety, and privacy (Douglas, 1996). Despite the extreme eco­ currently, facility managers settled their decision-making on mainte­
nomic, cultural, and historical vitality of buildings, there is evidence nance and renovation approaches relying on sparse information
that they are degrading and are in a flaccid state due to insufficient or regarding the building facilities’ components’ current actual condition.
misleading information about their facilities conditions, required for These sparse data are insufficient to monitor and predict building fa­
building facilities maintenance management (Becerik-Gerber et al., cilities components’ signs of wear, tear, and Age, hindering bench­
2012). Hence, Condition Assessment (CA) of the building facilities marking enhancements outcomes and signifying any substantial
components is essential to assist the decision-making process regarding conditions alternations before influencing building facilities compo­
effective maintenance management plans and service requirements. nents efficiency. This issue highlights the demand for a condition
Unless there is an adequate CA. for building facilities components, they assessment model, which can portray a proactive alternative picture of
may deteriorate more rapidly and evolve into a less profitable asset, repair and maintenance requirements based on the future condition
coupled with a huge expenditures backlog in supporting their primary prediction of building facilities components (Grussing et al., 2016;

* Corresponding author.
E-mail addresses: [email protected] (A. Gouda Mohamed), [email protected] (M. Marzouk).

https://doi.org/10.1016/j.eswa.2021.115743
Received 8 May 2021; Received in revised form 6 August 2021; Accepted 7 August 2021
Available online 15 August 2021
0957-4174/© 2021 Elsevier Ltd. All rights reserved.
A. Gouda Mohamed and M. Marzouk Expert Systems With Applications 186 (2021) 115743

Grussing, 2018; Sadick & Issa, 2018). Grussing (2018) presented a framework for measuring the signifi­
The novel contribution of this research is to present a physical con­ cance of building component inspection data concerning the decision-
dition assessment model for existing educational building facilities making process for maintenance work planning. This framework uti­
components while supporting the mapping out of a long-term plan for lizes an information value decision tree and a Markov component
optimal maintenance and repair resolutions via an ANN predictive condition-forecasting model to estimate the likelihood of the building
model and SEM framework. The primary merit of this model is: 1) component failure. The proposed framework considers the last time
allowing facility managers to accurately predict the future condition inspection took place, the anticipated component condition state, the
data of building facilities components throughout the monitoring pro­ failure risk consequence, and the cost of risk mitigation associated with
cess, which facilitates determining the service level metrics for in­ the constructive repair or replacement practices.
terventions criteria and the foreseeable timing; 2) prioritizing building Furthermore, Sadick & Issa (2018) examined the development and
facilities components for maintenance purposes based on proportionate adoption of a condition evaluation instrument thoroughly concerning a
weighted method and rate of degradation; and 3) recording the potential building’s space level to evaluate the effect of building components’
effects of several parameters influencing the condition state of building physical conditions on indoor environmental quality, especially in
facilities components. A case study checks and confirms the proposed school buildings. This study examined 52 classrooms, either in new,
model implementation and its functional efficacy via selected space at rehabilitated, or non-rehabilitated schools. Findings of the condition
Cairo University in Egypt. evaluation information of the space level revealed that the highest
proportion of statistical significance discrepancies is amongst new and
2. Related works non-rehabilitated schools, while the lowest amongst rehabilitated and
non-rehabilitated. The condition evaluation scores of the space level and
2.1. Building condition assessment indoor environmental quality field analyses also affirmed that the
highest correlation exists amongst the main envelope of the condition
Assessment of condition is the foundation of ascertaining the rate of evaluation score of the space level and relative humidity in non-
regular maintenance appropriate for building facilities components. A rehabilitated schools.
condition assessment promotes a methodological approach needed for Atef and Bristow (2019) also proposed a tool to evaluate the prop­
assessing the capital assets of an entity for the facility repair purpose, agation of existing buildings and the failure of their components during
renovation, or substitution requirements, which will maintain their ca­ the operational phase. The tool consists of two major algorithms: spatial
pacity to handle the allocated tasks to serve. Abundant research initia­ and operational interdependency algorithms. The spatial algorithm
tives have been carried out for building condition assessment disintegrates a building into distinct venues and excerpts each venue’s
approaches in various contexts; instances involve distinct assessment components to be assembled into architectural, structural, mechanical,
types such as structural (Sandoval et al., 2017; Cerchiello et al., 2018; electrical, and plumbing disciplines. At the same time, the operational
Hait et al., 2020), environmental (Jung & Park, 2015; Lützkendorf, algorithm excerpts functional interdependency relationships among the
2018; Danivska et al., 2019), Building Information Modeling (BIM)- building’s distinctive systems to attach the building venues to their
based sustainable (Jalaei et al., 2019; Mellado et al., 2020), and physical functional dependencies.
(Elhakeem & Hegazy, 2005; Ahluwalia, 2008; Grussing, 2018; Atef & Nevertheless, Mohd Noor et al. (2019) portrayed a Building Condi­
Bristow, 2019). tion Assessment (BCA) implementation on a heritage building depend­
Several research attempts have been conducted to assess existing ing upon the maintainability, security, and functionality of structural
buildings’ physical conditions in the context of physical condition and architectural elements, in addition to original construction material
assessment. For example, Elhakeem & Hegazy (2005) created a visual sustainability. This research determines each defect ranking relative to
inspection scheme for buildings. This user-friendly approach can be its condition and maintenance requirement using a rating system via a
employed easily at sites to evaluate a condition to save time and costs comprehensive visual inspection and detailed notes for each current
concurrently. This visual guide scheme was designed for the building defect and drawbacks. Eventually, Piaia et al. (2020) endorsed the
component’s prompt evaluation by capturing images from previous and sustainable conservation and maintenance of cultural heritage buildings
different asset condition states. Furthermore, Langevine (2006) formu­ through a software platform for the on-site condition assessment and
lated a methodology to evaluate the building components status by asset management with an entrenched Building Information Model
investigating the building components weights through analyzing the (BIM) platform. This framework leverages the existing BIM information
proportionate significance of each component in the hierarchical to accelerate and improve the efficiency of building inspections so that
structure at each tier using thorough inspections developed at the bot­ conservation maintenance planning can be tailored and optimized
tom tier of the hierarchical building structure, and by employing a roll- concerning various requirements.
up process to evaluate the building condition.
On the other side, Grussing et al. (2006) adopted a probability dis­ 2.2. Artificial neural networks
tribution model to assess the building components’ lifecycle condition
and efficiency over time. This model adjusts itself utilizing attribute data Artificial Neural Network (ANN) is a refined information model of
gathered throughout the present and recorded inspections so that an biological neuron structures via a simple structure, which aims at
individual element in a building can precisely map the distinctive life­ emulating the thinking mechanism in the human brain. The human
cycle deterioration scheme. Ahluwalia (2008) established the embedded brain operates data via a sequence of interdependent elements or neu­
monitoring and condition evaluation framework to emphasize the most rons that transmit direct data through links. An ANN comprises
critical components required to be inspected. This framework presents a numerous linear or nonlinear information-processing paradigms linked
visual guidance system for building components, supported by a in a parallel scheme, influenced by the human brain processing infor­
graphical and descriptive repository, and provides a uniform building mation system. Concerning ANN modeling, each ANN processing unit
hierarchy for location-based inspections. Moreover, Hossny et al. (2013) collects massive inputs via weighted neuron connections in the other
introduced a framework model for educational buildings that in­ layer, executes adequate computations, and sends the output to other
corporates maintenance data to promote reliable inspection preparation processing layers. Optimization algorithms alter the weighted connec­
and expedite condition assessment for interconnected building compo­ tions among neurons throughout the learning phase to decrease the ANN
nents. The conducted model minimizes field monitoring by incorpo­ model errors that offer a more precise determinant or predictor. The
rating a degradation model that utilizes neural network techniques to Back Propagation Multi-Layer Perceptron (BP MLP) network is consid­
forecast various planning horizon repair approaches. ered one of the ANN model easy types, which decomposes an input

2
A. Gouda Mohamed and M. Marzouk Expert Systems With Applications 186 (2021) 115743

layer, hidden layers (one or more), and an output layer. External 3. Research gaps
reference supplies input neurons with input information, while hidden
neurons obtain inputs and outputs throughout the network, and they are To sum up, this section provided a comprehensive literature review
not apparent to the external structures. Output neurons give out signals in educational building condition assessment methodological ap­
from the network, measured by the system (Liu, 2012). proaches needed for assessing the capital assets of an entity for the fa­
Several research endeavors have been conducted in ANN imple­ cility repair purpose, renovation, or substitution requirements. The prior
mentations in ecological sciences to generate precise prediction models. literature concludes that previous approaches are still deemed prema­
For instance, Jahani and Saffariha (2020) developed a landscape model ture, with some limitations demanded to be enhanced. For instance: 1)
to anticipate the aesthetic prioritization and mental restoration values in the prior research approaches apropos of the physical condition
urban parks via employing artificial intelligence approaches like Multi- assessment of building are generic; indeed, it is very challenging to track
Layer Perceptron (MLP) neural network, Radial Basis Function Neural down the concerns within each building space; 2) the previous research
Network (RBFNN), and Support Vector Machine (SVM). The model studies do not assess a building relying on each facility component
recorded 11 landscape characteristics in 200 urban parks. The results condition state; 3) forecasting the building condition state concerning its
revealed that the SVM model depicts the most reliable assessment model internal spaces and their associated physical components using ANN
for anticipating the urban parks landscape score in aesthetic and mental model do not exists, and 4) Preceding models concerned with building
restoration potential prediction. Furthermore, Jahani et al. (2021) components physical condition did not contemplate the proportionate
proposed a decision support system to model and anticipate the land­ significance of each facility component to prioritize their importance
scape aesthetic value and quality in urban parks via employing ANN. level inside spaces.
The ANN model aims to anticipate the landscape aesthetical value and
prioritize the model’s crucial variables. In this research, the authors 4. Method
integrate the user’s perspective to quantify the landscape aesthetical
quality urban park and the ANN modeling approach to record the This research endeavor proposes a novel physical condition assess­
effective features and variables on the visual quality of the landscape. ment model for educational building facilities components that intends
Nevertheless, Jahani and Saffariha (2021) utilized machine learning to support clients’ and facility managers’ actions and decisions
approaches to discern trustworthy predictions for modeling the regarding the maintenance demands. The proposed model scheme relies
responsiveness of trees to wind pressure and developing novel tech­ on assessing a building facility’s components condition concerning in­
niques in forest management. Jahani and Saffariha (2021) proposed a ternal spaces and their associated physical components based on their
Tree Failure Model (TFM) via employing artificial intelligence ap­ proportionate weighted method. The developed model allows creating
proaches like Multi-Layer Perceptron (MLP) neural network, Radial an estimate of the recent maintenance services efficacy required for
Basis Function Neural Network (RBFNN), and Support Vector Machine determining the severity of the assigned and needed maintenance. Four
(SVM) models to identify impressionable trees in windstorm distur­ sequential modules achieve the developed model objectives to conduct
bances. Results revealed The MLP model depicts the highest precision of physical condition assessment per each educational building space, as
target trees classification in training and serves as a comparative impact shown in Fig. 1.
assessment model to lessen trees failure in wind circulations.
Numerous ANN modeling attempts have been endeavored to pre­ 4.1. Scan to BIM module
cisely predict energy consumption in buildings encompassing engi­
neering, statistical and artificial intelligence methods. For instance, Scan-to-BIM provides physical 3D laser scanning for indoor spaces or
Elbeltagi and Wefki (2021) proposed a methodological approach to environments to build a precise digital portrayal of these spaces. This
enhance anticipating the energy utilization inside residential buildings digital portrayal can then be utilized or implemented to distinct prac­
at the early design stages, relying on ANN modeling. The developed tices in which existing facilities demand to be captured, such as reno­
ANN model was operated by employing a dataset retrieved from energy vation, rehabilitation, or restoration activities. The scan-to-BIM process
consumption computations via simulating and integrating various can be sub-divided into three phases; these phases are: 1) automated
design scenarios with the input variables. A user-friendly interface was data acquisition and information capturing; 2) point clouds segmenta­
refined to anticipate energy consumption conveniently via an under­ tion process, and 3) modeling point clouds of building facilities. Auto­
standable and trustworthy decision support tool. mated data acquisition and information capturing are accountable for
Furthermore, Amasyali and El-Gohary (2021) developed a machine- data acquisition and scanning indoor building facilities using a terres­
learning approach based on an occupant-behavior-sensitive manner for trial laser scanner. The data acquisition process results in discrete re­
anticipating building energy consumption. An enormous set of energy- cords of point clouds, representing 3D spatiality information of indoor
use cases were modeled and simulated in EnergyPlus for the learning building facilities. The data acquisition process of building facilities
process. The model was trained via a huge dataset encompassing 3- results in a unified BIM model, incorporating data from multiple chan­
month hourly data for 5760 energy-use cases depicting distinct build­ nels that vary based on similar parameters. Thus, point clouds seg­
ing features, including outdoor weather conditions and occupant be­ mentation aims to segment and decompose point clouds into sets of
haviors. The developed approach employed four machine-learning workable clusters with similar properties via a region-growing algo­
algorithms, namely 1) Classification and Regression Trees (CART), 2) rithm. The purpose here is to break down the point clouds into practi­
Ensemble Bagging Trees (EBT), 3) Artificial Neural Networks (ANN), cable clusters and detect the scanned features acquiring the same
and 4) Deep Neural Networks (DNN) for testing and comparing with properties, which can be attained using a region-growing-based algo­
respect to the four algorithms anticipation precision and computational rithm. The region growing-based algorithm to identify building com­
efficacy. ponents from the point clouds is conducted using an object-oriented
Nevertheless, Chen et al. (2021) proposed two approaches to fit and programming language like C++ via the Microsoft visual studio
anticipate the electricity requirements concerning office buildings. The platform.
first approach stressed lessening occupants’ activities, including the Eventually, modeling point clouds of building facilities are respon­
working hours and non-working hours. The second approach employed sible for evolving segmented building facilities components into a
ANN and fuzzy logic techniques to fit the building baseload, peak load, semantically prosperous as-is 3D BIM model. Further details and ex­
and occupancy rate with multi-variables of weather variables. Actual planations regarding this module can be explicitly discussed in the
data retrieved from the University of Glasgow are utilized to verify the research endeavor developed by the authors (Gouda et al., 2020). The
proposed approach for building energy management purposes. output of the scan to BIM module is a 3D digital and physical

3
A. Gouda Mohamed and M. Marzouk Expert Systems With Applications 186 (2021) 115743

Fig. 1. Proposed physical condition assessment model.

representation of the indoor building components to be extensively other than other methods concerning addressing complicated nonlinear
employed for various approaches. The condition-monitoring data of problems and enduring faults and noise.
these identified building components and their inspection records are The data employed in this research are collected from the Faculty of
considered the significant input variables of the ANN network to run the Engineering, Cairo University, Egypt, to operate the ANN model. The
model and prioritize building facilities components for maintenance database concerning condition monitoring has amassed data over the
purposes. last obtainable five years. The condition monitoring data related to
building various components were recorded from inspectors’ visual
observation. In this research, condition-monitoring data on six different
4.2. Condition assessment prediction module buildings’ internal components are analyzed to anticipate their future
condition state. The components carried out are: 1) wooden flooring
This section introduces the construction of a Back Propagation Multi- tiles; 2) gypsum board ceiling tiles; 3) wooden doors; 4) wooden win­
Layer Perceptron (BP-MLP) ANN model to forecast future condition dows, 5) split air conditioner units; and 6) desktop computers. A set of
ratings for facilities components in educational building spaces, which is 35 distinct spaces from 4 different buildings at the Faculty of Engi­
a satisfactory sign for their deterioration paradigm. MLP is considered neering, Cairo University, are utilized to compute the future condition
one of the ANN model easy types, which decomposes an input layer, ratings from their input data. These 35 spaces involve 35 wooden
hidden layers (one or more), and an output layer. Input neurons are flooring tiles, 35 gypsum board ceiling tiles, 35 wooden doors, 50
supplied with input information from an external reference. Hidden wooden windows, 40 split air conditioner units, and 50 desktop
neurons obtain inputs and outputs throughout the network, and they are computers.
not apparent to the external structures. Output neurons give out signals The amassed dataset encompasses condition information for these
from the network, measured by the system. In MLP, every neuron in the components via a seven-point index scale as a monitoring tool, as clearly
layer is attached to the entire neurons in the succeeding layer via the depicted and discussed in Fig. 2. The collected dataset comprises further
weighted connections. The arrows’ direction illustrates that the data is physical parameters such as components length and width, last inspec­
forwarded via the neurons network from the input layer to the output tion date, expected service life, installation date, annual average daily
layer in a forward path; the forward interconnection of this data is usage, Age, and the number of observation intervals per component.
named feed-forward framework. Since ANN can acquire information from the previous observations and
ANN model learns from the input and output parameters mapping then implement the knowledge to successfully handle calculations to
via a learning mechanism called error backpropagation, which includes figure out the numerical values for unforeseen data. The knowledge is
a forward and backward pass through the network. With a defined target attained via the learning process, entitled the training phase. At this
output, the input vector is implemented to the network neurons in the phase, the network is refined by utilizing a training dataset, portrayed as
forward pass, and its impact extends via the network, layer by layer. The the inspection records for the components mentioned above. The ANN
network connection weights are adjusted during the forward pass. The network learns and defines the relation between input and output pa­
network feedback is the array of inputs generated at the end of the rameters during the learning phase. The relation is construed via uti­
forward pass. On the other hand, the connection weights are adapted lizing the interconnection strengths amidst nodes through weights.
using the error backpropagation algorithm during the backward pass. These weights are stored to be modeled to predict the future condition
the backpropagation algorithm is a probabilistic gradient regression state of the pre-defined components. This research utilizes MATLAB
approach employed to mitigate square error. The algorithm delivers a Neural Network Toolbox (MATLAB R2018a) for developing the ANN
manner to compute the error slope function utilizing the chain differ­ model.
entiation rule (Jahani & Saffariha, 2021).
The ANN is selected to carry out this research because it can 4.2.1. Collected data processing
accommodate several parameters that substantially impact the output This model uses the Facility Component Average Rating (FCAR), as
results. ANN can acquire information from accessible data and establish shown in Eq.1, to construe the condition of the components from the
the relationship between the output and exogenous input variables. data assembled. By employing the computed FCAR per component, the
Besides, an analysis of the weights determined by the ANN model may data are screened to be refined from any enhancement in components
be beneficial in determining the most significant parameters influencing condition during the inspection. Any enhancement in the FCAR value
the model output. Moreover, ANNs have numerous remarkable features

4
A. Gouda Mohamed and M. Marzouk Expert Systems With Applications 186 (2021) 115743

Fig. 2. The condition index scale for the facility components.

per each component signifies that maintenance intervention has been standardized using Z-score (Eq. (2)) to meet the transfer function criteria
carried out. These recorders will be exempted from the analysis process utilized in creating the ANN model, Sigmoid Feature. In Eq. (2), z depicts
to reliably experience the future deterioration trend in components the standardized or scaled values, μ is the mean or average value of each
condition. Table 1 depicts an original sample from the dataset amassed input variable, X is the non-standardized or raw data, and σ depicts the
for training and testing the selected components. As previously standard deviation of each input variable.
mentioned, the adopted condition state scale is a seven-point index, in
X− μ
which condition state 1 implies that the component condition state z= (2)
σ
ranges from 99% to 100%. This index means that the component seems
to be either new or has been retained lately and does not reveal any
4.2.3. Input and output layer
indications of deterioration. In this case, the action required is preven­
The input variables matrix for each facility component is designed
tion Maintainance, as seen in Fig. 2.
and represented as n1 × n3, in which n1 is the input parameter number to
Moreover, condition state 2 implies that the component condition
the network (n = 5), and n3 is the number of components. The input
state ranges from 92% to 99%. This index implies that trivial sub-
number prescribes the size and node number on the ANN model’s input
components may endure slight deterioration, and the action required
layer. The output dimensions use the same configuration as the input
is very slight serviceability. The other condition state from 3 to 7 are
dimensions. The network output layer comprises seven nodes, initiating
explicitly depicted in Fig. 2, including their condition index range, lin­
from condition state 1 to condition state 7, as represented in Fig. 2. The
guistic depiction, and action required.
ANN model output size is n2 × n3, in which n2 is the number of output
parameters (n = 7), and n3 is the number of components, as shown in
( )
7P1 + 6P2 + 5P3 + 4P4 + 3P5 + 2P6 + 1P7
FCAR = x100 (1) Fig. 3.
7
FCAR is the facility component average rating value, and Pi is the 4.2.4. Hidden layer and activation function
facility component percentage with a condition rating of “i divide by During this stage, the neuron number in the hidden layer is deter­
100”, as “I” ranges from 1 to 7. mined to achieve optimum outcomes. Eq. (3) computes the number of
hidden neurons in a single layer commonly adopted in several research
4.2.2. Input parameters identification efforts (Xu & Chen, 2008). Where n depicts the optimum neuron number
The ANN model output precision relies heavily upon the accurate in a single hidden layer, Ip is the input layer size, and Op is the output
identification of the input variables. To identify the significant input layer size.
variables of the network, factors listed in prior research endeavors (ISO,
2011; ISO, 2012; Grussing et al., 2016); besides, around 25 facility n=
2 × (Ip + Op)
(3)
managers’ feedback to identify the substantial variables for educational 3
building facilitates components, which act as the network input pa­ In the network properties, the transfer function used for the hidden
rameters. The final selected input parameters are Age, Observation In­ layer is “TANSIG,” which is accountable as a transition function from the
terval (OI), Last Inspection Date (LID), Expected Service Life (ESL), all input layer to the hidden one and transitioning from the hidden layer to
are in years, and Annual Average Daily Usage (AADU) as a percentage the output one. The hyperbolic tangent sigmoid function is utilized for
value. The inspection records of these input variables concerning the activation, which has a broad set of output values that range from − 1 to
trained components are collected as numerical values. These numerical 1.
values are scattered. Their deviation is enormous; thus, they need to be

Table 1
Sample of the data collected for wooden door components.
Component I.D Condition State FCAR Age OI LID ESL AADU

1 2 3 4 5 6 7

1 WD_4 0.56 0 0.03 0.18 0.05 0.15 0.03 75.8 14 6 4 26 73

2 WD_9 0 0 0 0.78 0.22 0 0 53.8 14 8 4 26 62

3 WD_13 0.76 0.20 0.01 0.01 0.02 0 0 95.5 3 12 1 37 28
4 WD_19 0.03 0.72 0.24 0 0.01 0 0 82.2 9 10 2 31 43
5 WD_24 0.10 0 0 0 0.75 0.12 0.03 46.3 17 3 5.5 23 88
6 WD_29 0.21 0 0.02 0.01 0.60 0.14 0.02 53.1 14 4 5 26 85
7 WD_32 0.03 0 0 0 0.82 0.10 0.05 41.6 17 2 6 23 93

5
A. Gouda Mohamed and M. Marzouk Expert Systems With Applications 186 (2021) 115743

√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√∑
√n
√ (yi − Å⋅i)2
√
RMSE = i=1 (6)
n
∑n
i=1 |yi − Å⋅i|
MAE = (7)
n

4.2.7. Testing the ANN models

The creation of an ANN model engages partitioning the data gath­
ered into three clusters. The first cluster involves the training set, which
is responsible for measuring and updating the weights and biases of the
network. The second cluster is the validation set, while the third cluster
commits to testing refined models with actual context data. This
research divides collected data into two clusters, in which most of the
data are employed for model training to compute the weights and biases.
In contrast, the remaining data are adopted to test the trained network.
The network training process exploits around 70% of the input data
Fig. 3. The architecture of the ANN predictive model. using the Matlab random function, while the network testing process
adopts the remaining data. Finally, the model precision on the test data
creates a reasonable estimate of the model efficiency on entirely invis­
4.2.5. Learning and training function
ible records and reveals the network’s real predictive potentials. After­
The ANN employs the Bayesian Regularization function to prevent
ward, the “TRAIN” function reverts the trained neural network
the developed model from being prone to overfitting behavior during
(optimized weights and bias of the connector) as well as the network
training the network. Bayesian regularization is a technique to lessen the
error. Besides, the ANN configures a 0.01-minute goal, 1000 times
complication of the model and is appropriate for noisy and extensive
epoch, 0.6 learning rate, and 0.4 moments for the model training.
information. This function is responsible for updating weight and bias
values relevant to the Levenberg Marquardt optimization approach
(Roweis, 1996). The Bayesian regularization method reduces the inte­ 4.3. Proportionate weight determination module
gration of squared errors and weights and dictates the right integration
for creating a distinct, optimized, and predictive network. Moreover, The proportionate weights determination of building facilities com­
another approach to ensuring the developed ANN does not overfitting ponents is still recognized as a complicated mission across entire
behavior is to stop the training process early at the optimum perfor­ building facilities categories and prioritizing the building facilities
mance epoch (Amasyali & El-Gohary, 2021). components’ importance levels, which differs from one space category
The “TRAINBR” training function is selected in the network prop­ to another. This research consequently determines each facility com­
erties, which is accountable for the neural network training process, ponent’s proportionate weight computing process within each specific
depending on the parameters dataset, along with momentum and space category. A systematic questionnaire survey is constructed for
learning frequency. Moreover, the “LEARNGDM” learning function is experts to determine the proportionate weights of building facilities’
opted responsible for weight and bias calculations and fluctuations for components.
the backpropagation.
For training purposes, the network adopts the “TRAIN” function for 4.3.1. Questionnaire survey design
training the neural network, which uses the training input and output For data acquisition and knowledge attainment, facilities managers,
information as arguments. The training process amends the neural asset management specialists, and academic staff, from Egypt are
network weights and ensures that the network and fine-tune models are considered for responding to the questionnaire survey (years of experi­
not over-fitted. A validation set requires to be added to the network to ence ranges from 10 to 30). The sample size population requested the
verify if the error is within the range (this set is not utilized immediately questionnaire survey response calculated using an infinite and finite
to change the weights but is employed to provide an optimal number of population sample size, developed by Godden (2004). For descriptive
hidden units or the setting of the backpropagation algorithm stop point). sampling statistics dealing with uncertainty, Eq. (8) computes the
sample size for an infinite population.
4.2.6. ANN model validation Z2xPx(1 − P)
The developed ANN models performance is appraised by simulating SS = (8)
C2
a set of test data not employed in the training process. The Coefficient of
Determination (R2-squared, Eq. (4)), Mean Squared Error (MSE, Eq.5), SS depicts the sample size for an infinite population; Z represents the
Root Mean Squared Error (RMSE, Eq.6), and Mean Absolute Error (MAE, probability that a sample will fall within a certain distribution; P depicts
Eq.7) are used as statistical indicators to anticipate future performance the population percentage opting a choice, shown as a decimal value,
or test assumptions relying upon other pertinent data and to attain and C depicts the confidence interval, expressed as a decimal value. The
optimal optimized performance of the ANN. In these equations, yi de­ infinite population formula computes the sample size. The sample size
picts the targets, ŷi represents the network outputs, ỹi depicts the mean derived from that calculation is adopted to calculate a new sample size
of target values, and n is the samples number (Saffariha et al., 2021). for a finite population using Eq. (9). New SS depicts the sample size for a
finite population, and Pop represents the population multiplied by
∑n
yi)2
yi − ̃
(̂ personnel working.
R - Squared = ∑i=1 n
(4)
yi)2
i=1 (yi − ̃ SS
NewSS = ( ) (9)
∑n 2 1 + SS− 1
i=1 (yi− Å⋅i)
(5)
Pop
MSE =
n
The questionnaire’s general question is as follows: “What is the sig­
nificance of each building facilities component in the analysis of the

6
A. Gouda Mohamed and M. Marzouk Expert Systems With Applications 186 (2021) 115743

space condition?”. The questionnaire survey requests the respondents to

submit their responses with a 7-point scale, as follows: 1) very weak 5. Model implementation
significance; 2) weak significance; 3) slight significance; 4) neutral sig­
nificance; 5) strong significance; 6) very strong significance; and 7) The validity approach affirms the proposed model for condition
absolute significance, respectively. Four main spaces category in the assessment by integrating it via an existing real context. Condition-
educational buildings (classrooms, computer labs, staff offices, and monitoring data on six different buildings’ internal components are
lecture halls) are designated for the questionnaire to prioritize and rank analyzed to anticipate their future condition. The components carried
their components proportionate importance, which alters based on the out are: 1) wooden flooring tiles; 2) gypsum board ceiling tiles; 3)
typology of buildings. wooden doors; 4) wooden windows, 5) split air conditioner units; and 6)
desktop computers. A set of 35 distinct spaces from 4 different buildings
4.3.2. Proportionate weights calculations at the Faculty of Engineering, Cairo University, are utilized to compute
Data derived from questionnaire responses are processed through the future condition ratings from their input data. These 35 spaces
employing Confirmatory Factor Analysis (CFA), which tackles the involve 35 wooden flooring tiles, 35 gypsum board ceiling tiles, 35
interdependence among different building facilities components. CFA is wooden doors, 50 wooden windows, 40 split air conditioner units, and
an empirical approach used in determining the linkage between the 50 desktop computers.
variable and its indicators by analyzing the correlation. CFA is one of the
phases in Structural Equation Modeling (SEM) that computes the in­ 5.1. Condition assessment prediction
dicators coefficients, equivalent to the regression coefficient between
indicators Y with variable X. The coefficient is called the “loading fac­ Throughout the ANN model, the input layer and output layer are
tor,” which is a value between “0” (defined as weak) and “1” (defined as defined to forecast the future condition ratings for each facility
perfect fit). When the value approaches point 1, this implies that in­ component. The input variables matrix size is [5 × 35] for the wooden
dicators are significant. flooring tiles, [5 × 35] for the gypsum board ceiling tiles, [5 × 35] for
Moreover, CFA is adopted to identify the component indicator sig­ the wooden doors, [5 × 50] the wooden windows, [5 × 40] for the split
nificance of a variable. Building spaces, for instance, are defined as air conditioner units, and [5 × 50] for the desktop computers. The input
variables in the CFA problem, and space components, including doors, variables matrix size is variant attributable to the acquirable amassed
windows, air conditioning, floors, etc., are represented as indicators. A data over the last obtainable five years, concerning the condition
loading factor depicts the results of CFA on each indicator’s significance. monitoring data recorded from inspectors’ visual observation. The
By employing loading factors, the ranking process for each facility output dimensions use the same configuration as the input dimensions,
component per each space category is identified, relying on elicited in which the network output layer comprises seven nodes, initiating
questionnaire survey responses. Fig. 4 shows modeling the indicators from condition state 1 to condition state 7. Table 2 shows a sample from
and variables relation in CFA format; besides, Eq. (10) computes the the input parameter values for the picked components generated from
loading factor of each space component. Y represents each variable’s inspection collected records.
indicators, F is the loading factor for each indicator, X is the variable,
and Z is the error term. 5.1.1. ANN model results and discussions
The input variables are standardized and scaled using the stan­
Y = F.X + Z (10)
dardization Z-score Equation. Table 3 outlines the collected and antic­
ipated condition state per each component as output values. The
4.4. Entire space rating value module network output layer comprises seven nodes that depicted the seven
condition states presented earlier. The developed ANN models are
This module exploits the anticipated condition-rating states of validated using the Coefficient of Determination (R2), Mean Squared
building facility components, conducted using the ANN predictive Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute
model and prior defined the proportionate significance of each facility Error (MAE), in which the optimal outcomes of the ANN models and the
component to calculate the entire rating of each building space using Eq. best structure in training and test datasets are depicted in Table 4. The
(11). Where ER is the entire rating value per each building space, SPiis coefficient of determination (R2), displayed in Table 4, represents the
the name of each building space, FCAR is the facility component average ANN networks’ precision in predicting each component’s condition
rating value within the space, as “j” represents each space component, state, relying on the input variables. Fig. 5 displays a sample of the
and W.F.Comp is the proportionate weight of each facility component regression plots for models refined using ANN for six distinct building
category within the space. components: wooden flooring tiles, gypsum board ceiling tiles; wooden
∑ doors; wooden windows, and split air conditioner units.
ER(SPi) = FCARj × W.F.Comp. (11) After exploiting the set of collected data for training the ANN model,
j
as stated earlier, the saved network architecture is employed to antici­
pate the condition rating for the selected components. These compo­
nents took advantage of the saved connection weights and biases to
compute the network output. The ages of these components are fore­
casted for the next five years (assuming no maintenance intervention);
the observation interval and AADU are retained the same, as their future
data are not feasible. Eventually, the network generated the condition
rating of the elected components for the prescribed dataset, as displayed
in Table 3. The FCAR is computed, using Eq. (1), from the ANN condition
rating outcome per each of the components, and the findings are dis­
played vividly in Fig. 6. From Fig. 6, it can be seen that the FCAR of the
selected components after five years frequently remains lower than their
initial collected values, which implies the suitability of the proposed
ANN to foresee and anticipate the components condition rating from the
defined input variables that mimic the components degradation
Fig. 4. Confirmatory factor analysis model. behavior. Fig. 7 depicts the average numerical value of collected and

7
A. Gouda Mohamed and M. Marzouk Expert Systems With Applications 186 (2021) 115743

Table 2
Input parameter sample for the selected components.
Component Category Component I.D Age Observation Interval Last Inspection Date Expected Service Life AADU

Wooden Flooring Tiles WFT_11 12 9 2.5 26 50

WFT_17 3 12 1 35 28
Gypsum Ceiling Tiles GCT_7 17 3 5.5 23 87
GCT_13 9 10 2 31 43
Wooden Doors WD_13 9 10 2 31 43
WD_24 14 7 3.5 26 62
Wooden Windows WW_2 14 7 3.5 21 62
WW_21 6 11 1.5 29 37
Split Air Conditioner SAC_14 17 2 6 13 93
SAC_19 7 11 1.5 23 38
Desktop Computers DC_5 2 9 2 13 28
DC_20 7 12 3 8 34

Table 3
Output sample for the selected components.
ID C1 C2 C3 C4 C5 C6 C7 C1 C2 C3 C4 C5 C6 C7

Collected Condition State Predicted Condition State

(Input) (Output)

WFT_11 0 0 0.85 0.10 0.01 0.03 0.01 0.03 0.02 0.52 0.32 0.02 0.01 0.05
WFT_17 0.75 0.22 0.02 0.01 0 0 0 0.4 0.43 0.15 0.01 0.01 0 0
GCT_7 0.1 0.08 0.12 0.23 0.01 0.07 0.4 0.02 0.04 0.06 0.1 0.08 0.10 0.68
GCT_13 0.26 0.74 0 0 0 0 0 0.21 0.48 0.26 0.01 0.03 0.01 0
WD_13 0.02 0.71 0.24 0.01 0.01 0.01 0 0 0.67 0.22 0.05 0.04 0 0.02
WD_24 0.01 0 0.11 0.84 0 0 0.04 0 0 0 0.85 0.08 0.03 0.04
WW_2 0.01 0.01 0.16 0.67 0.14 0.01 0 0 0 0.05 0.36 0.59 0 0
WW_21 0.64 0.24 0.08 0.01 0.01 0.02 0 0.01 0.61 0.31 0 0 0.01 0.06
SAC_14 0 0.17 0.01 0.15 0.67 0 0.01 0 0 0 0 0 0.9 0.1
SAC_19 0.67 0.16 0.08 0.05 0.02 0.02 0 0.18 0.70 0.04 0.04 0.02 0.01 0.02
DC_5 0.52 0.48 0 0 0 0 0 0.07 0.45 0.08 0.1 0.1 0.1 0.1
DC_20 0.69 0.05 0.07 0.00 0.09 0.07 0.04 0.01 0.38 0.56 0.05 0 0 0

Table 4
The results of ANN models validations.
Component Category Test Set Training Data

R2 MSE RMSE MAE R2 MSE RMSE MAE

Wooden Flooring Tiles 0.948 0.036 0.190 0.167 0.997 0.014 0.118 0.084
Gypsum Ceiling Tiles 0.959 0.158 0.397 0.762 0.944 0.033 0.182 0.153
Wooden Doors 0.959 0.084 0.290 0.533 0.923 0.023 0.152 0.111
Wooden Windows 0.945 0.074 0.272 0.357 0.862 0.017 0.130 0.108
Split Air Conditioner 0.986 0.124 0.352 0.574 0.97 0.019 0.138 0.106
Desktop Computers 0.97 0.177 0.421 0.855 0.965 0.027 0.164 0.175

predicted FCAR; for instance, the average FCAR of the data amassed introduced by Olden et al. (2004) to rank the significance of the input
regarding the 50 wooden windows, computed using Eq. (1), is 72.22; on parameters in descending order by categorizing the entire significance
the contrary, the predicted FCAR after five years time horizon (without measure of each parameter. The number of hidden neurons, weights
any maintenance intervention) is 61.42. From Fig. 7, it can be inferred between the input parameters and hidden neurons, and weights between
that the deterioration rate of wooden flooring tiles, gypsum board hidden neurons and the output neurons are essential factors to compute
ceiling tiles, wooden doors, wooden windows, split air conditioner units, each input parameter’s entire significance measure. The results
and desktop computers are 12.56%, 22.69%, 16.85%, 14.95%, 16.38%, retrieved from connection weight analysis reveal that the Annual
and 19.78 respectively. Average Daily Usage (AADU) acquire the highest significance in iden­
The proposed approach is the first research endeavor to predict the tifying the component’s condition state, which Age follows, Observation
building’s components condition state via an ANN predictive and SEM Interval(OI), Expected Service Life (ESL), and Last Inspection Date (LID).
framework, which is dissimilar to the previous research works that aim Table 5 depicts the input parameters’ entire significance as the ANN
attention at ANN implementations in ecological sciences to generate model derivative of the wooden flooring tiles component.
precise prediction models (Jahani & Saffariha, 2020, Jahani et al., 2021,
and Jahani & Saffariha, 2021), and ANN interventions in predicting 5.2. Proportionate weight determination
energy consumption in buildings (Elbeltagi & Wefki, 2021; Amasyali &
El-Gohary, 2021; Chen et al., 2021). The developed ANN models are The questionnaire survey comprises two parts: part one concerns the
validated using the R2, MSE, RMSE, and MAE to monitor their accuracy. questionnaire results acquisition; part two deals with analyzing the
Compared with previous research endeavors validation values, their gathered results to identify the proportionate weight per each facility
values are satisfactory and compelling to predict each component’s component in each space category. Eq. (8) and (9) calculate the infinite
condition state, relying on the input variables. and finite population sample size, as follows:
This research adopts the connection weight analysis (CWA)

8
A. Gouda Mohamed and M. Marzouk Expert Systems With Applications 186 (2021) 115743

Fig. 5. A sample of the coefficient of determination plots for the ANN models. (a) R-squared value for wooden flooring tiles; (b) R-squared value for gypsum board
ceiling tiles; and (c) R-squared value for wooden doors.

• An infinite population sample size = (1.962 × 0.3 × 0.3)/0.082 = 54 ascertain the four spaces components framework. After creating the SEM
• A finite population sample size = (54)/(1 + (54 – 1)/2,500) = 53 framework, depending on recapped questionnaire survey findings,
components inside the four designated spaces are assessed using CFA to
The questionnaires are delivered to approximately 90 field experts, determine the components loading factor inside each space category.
whether facility managers or academic staff, aged around 10 and 30 CFA outcomes portrayed the ranking and preference of components
years of professional experience. These experts are elected to cover all inside each space category, as shown in Fig. 9, via loading factors. Fig. 9
aspects and dimensions of specified building spaces. According to pop­ summarizes the questionnaire survey findings that inferred the sub­
ulation sample size equations, the questionnaire answers percentage is stantial proportionate weight per each of the selected components inside
approximately 68.88%, which is convenient and regarded as compelling the four different space categories. Results revealed that the “desktop
support for this research study. computers” and “split air conditioner units” components acquire the
highest proportionate weight among other components predominantly.
5.2.1. The computational process of components proportionate weights This outcome attributes the fact that the educational building spaces
The questionnaire survey findings are imported into Amos software, category relies heavily, in its functionality, on desktop computers and
developed by IBM. Firstly, four Structural Equation Modeling (SEM) split air conditioner units.
frameworks for classrooms, computer labs, staff offices, and lecture halls
components are designed (see Fig. 8) to attain and reach the “goodness-
of-fit” that is demanded by the SEM implementation to validate and

9
A. Gouda Mohamed and M. Marzouk Expert Systems With Applications 186 (2021) 115743

Fig. 6. A sample of the collected and predicted FCAR.

Fig. 8. SEM framework for typical computer lab components.

Fig. 7. Average FCAR of the selected components.

Table 5
Input parameters’ entire significance of the wooden flooring tiles component
ANN model.
No. Input Parameter Entire Significance

1 Annual Average Daily Usage 1.12

2 Age 0.98
3 Observation Interval 0.87
4 Expected Service Life 0.57 Fig. 9. Proportionate weights of the selected components inside the four-
5 Last Inspection Date 0.49 space category.

6. Conclusion weighted method, and record the portable influences of multiple pa­
rameters impacting the building facilities’ condition state. At the same
This paper portrays a novel physical condition assessment model for time and different from prior research endeavors, this model supports
existing educational building facilities components via an ANN predic­ generating an assessment of the latest maintenance services effective­
tive and SEM framework. This model aims to anticipate the condition ness designed for identifying the severity of the designated and
state of the building facilities components, prioritize building facilities demanded maintenance. The model objectives could be attained via four
components for service-level interventions based on proportionate consecutive modules: 1) scan to BIM module; 2) condition assessment

10
A. Gouda Mohamed and M. Marzouk Expert Systems With Applications 186 (2021) 115743

prediction module; 3) proportionate weight determination module; and Cerchiello, V., Ceresa, P., Monteiro, R., & Komendantova, N. (2018). Assessment of social
vulnerability to seismic hazard in Nablus, Palestine. International Journal of Disaster
4) entire space rating value module. The second module is liable for
Risk Reduction, 28, 491–506. https://doi.org/10.1016/j.ijdrr.2017.12.012
forecasting future condition ratings for each facility component in Chen, S., Ren, Y., Friedrich, D., Yu, Z., & Yu, J. (2021). Prediction of Office Building
various building spaces using ANN predive model. The third module is Electricity Demand using Artificial Neural Network by Splitting the Time Horizon for
culpable for determining the proportionate weights of building facilities Different Occupancy Rates. Energy and AI, 100093-1 - 100093-12. doi.org/10.1016/
j.egyai.2021.100093.
components and prioritizing their importance levels that vary from one Danivska, V., Heywood, C., Christersson, M., Zhang, E., & Nenonen, S. (2019).
space category to another using the CFA model. Eventually, the fourth Environmental and social sustainability–emergence of well-being in the built
module is chargeable for computing the entire rating value of each environment, assessment tools and real estate market implications. Intelligent
Buildings International, 11(3–4), 1–15. https://doi.org/10.1080/
building space relying on facility components’ condition-rating states, 17508975.2019.1678005
conducted using the ANN predictive model, and defined the propor­ Douglas, J. (1996). Building performance and its relevance to facilities management.
tionate significance of each facility component type inside each space. Facilities, 14, 23–32. https://doi: 10.1108/02632779610112508.
Elbeltagi, E., & Wefki, H. (2021). Predicting energy consumption for residential buildings
The developed condition assessment model was incorporated using ANN through parametric modeling. Energy Reports, 7, 2534-2545. doi.org/
throughout an actual application context to test and validate the do- 10.1016/j.egyr.2021.04.053.
ability of the proposed model. The case study findings and discussions Elhakeem, A., & Hegazy, T. (2005). Towards a visual guidance system for condition
assessment of the building infrastructure. In Proceedings of 1st CSCE Specialty
reveal the proposed model applicability in supporting clients’ and fa­ Conference on Infrastructure Technologies, Management, and Policies, CSCE,
cility managers’ actions and decisions regarding the maintenance de­ Toronto, Ontario, Canada, June (pp. 2-4).
mands via assessing a building facility’s components condition Godden, B. (2004). Sample size formulas. Journal of Statitics, 3 (66). http://williamgodd
en.com/samplesizeformula.pdf.
concerning internal spaces and their associated physical components
Gouda Mohamed, A., Abdallah, M.R., & Marzouk, M. (2020). BIM and semantic web-
based on the proportionate weighted method. based maintenance information for existing buildings. Automation in Construction,
Even though the current research endeavors have accomplished their 116, 103209-1 – 103209-20. doi:10.1016/j.autcon.2020.103209.
primary objective, there are some limitations for further research at­ Grussing, M. N. (2018). Optimized building component assessment planning using a
value of information model. Journal of Performance of Constructed Facilities, 32(4),
tempts to improve and broaden this research field. The proposed model 04018054–1–04018054–10. https://doi: 10.1061/(ASCE)CF.1943-5509.0001198.
can be extended by considering distinct building types, concede several Grussing, M. N., Liu, L. Y., Uzarski, D. R., El-Rayes, K., & El-Gohary, N. (2016). Discrete
challenges, considerations, and issues. Moreover, forecasting the con­ markov approach for building component condition, reliability, and service-life
prediction modeling. Journal of Performance of Constructed Facilities, 30(5),
dition state of the building facilities components utilizing the ANN 04016015–1–04016015–9. https://doi:10.1061/(ASCE)CF.1943-5509.0000865.
model can be further improved by considering the uncertainties and Grussing, M. N., Uzarski, D. R., & Marrano, L. R. (2006). Condition and reliability
challenges correlated with the building facilities’ components degra­ prediction models using the weibull probability distribution. In Applications of
advanced technology in transportation (pp. 19-24). https://doi: 10.1061/40799
dation during their lifecycle. (213)4.
Hait, P., Sil, A., & Choudhury, S. (2020). Seismic damage assessment and prediction
Funding using artificial neural network of RC building considering irregularities. Journal of
Structural Integrity and Maintenance, 5(1), 51–69. https://doi.org/10.1080/
24705314.2019.1692167
The research did not obtain any particular grant from governmental, Hossny, O., Elhakeem, A., & Hussein, D. (2013). An asset management system for
private, or non-profit funding institutions. maintenance and repair of educational buildings. International Journal of Scientific
& Engineering Research, 4(6), 2053–2064. https://www.ijser.org/researchpaper/
An-Asset-ManagementSystem-for-Maintenance-and-Repair-of-Educational-Buildin
CRediT authorship contribution statement gs.pdf.
ISO. (2011). Buildings and constructed assets — Service life planning Part 1: General
principles and framework. International Standards. https://www.iso.org/standard/4
Ahmed Gouda Mohamed: Conceptualization, Methodology, Soft­ 5798.html.
ware, Validation, Formal analysis, Investigation, Resources, Data cura­ Jalaei, F., Zoghi, M., & Khoshand, A. (2019). Life cycle environmental impact assessment
tion, Writing - original draft, Writing - review & editing, Visualization, to manage and optimize construction waste using Building Information Modeling
(BIM). International Journal of Construction Management, 1–18. https://doi.org/
Supervision, Project administration. Mohamed Marzouk: Conceptual­
10.1080/15623599.2019.1583850
ization, Methodology, Software, Validation, Formal analysis, Investi­ Jahani, A., & Saffariha, M. (2020). Aesthetic preference and mental restoration
gation, Resources, Data curation, Writing - original draft, Writing - prediction in urban parks: An application of environmental modeling approach.
review & editing, Visualization, Supervision, Project administration. Urban Forestry & Urban Greening, 54, 126775-1 - 126775-14 . doi.org/10.1016/j.
ufug.2020.126775.
Jahani, A., Allahverdi, S., Saffariha, M., Alitavoli, A., & Ghiyasi, S. (2021).
Declaration of Competing Interest Environmental modeling of landscape aesthetic value in natural urban parks using
artificial neural network technique. Modeling Earth Systems and Environment, 1-10.
doi.org/10.1007/s40808-020-01068-2.
The authors declare that they have no known competing financial Jahani, A., & Saffariha, M. (2021). Modeling of trees failure under windstorm in
interests or personal relationships that could have appeared to influence harvested Hyrcanian forests using machine learning techniques. Scientific Reports,
11(1), 1-13. doi.org/10.1038/s41598-020-80426-7.
the work reported in this paper. Jung, H., & Park, M. (2015). Sustainable building assessment through carbon efficiency
based on energy rating in apartments in Korea. International Journal of Sustainable
References Building Technology and Urban Development, 6(2), 114–124. https://doi.org/10.1080/
2093761X.2015.1036953
Langevine, R. (2006). A decision support system for the maintenance management of
Ahluwalia, S. (2008). A framework for efficient condition assessment of the building
buildings. Edmonton, Alberta, Canada: University of Alberta.
infrastructure. School of Civil and Environmental Engineering. Ph.D. Thesis. University
Liu, G. P. (2012). Nonlinear identification and control: A neural network approach.
of Waterloo.
Springer Science & Business Media. https://www.springer.com/gp/book/9781
Atef, A., & Bristow, D. (2019). Capturing the spatial and operational interdependencies
852333423.
among building systems using building information modelling. Structure and
Lützkendorf, T. (2018). Assessing the environmental performance of buildings: Trends,
Infrastructure Engineering, 15(12), 1613–1629. https://doi.org/10.1080/
lessons and tensions. Building Research & Information, 46(5), 594–614. https://doi.
15732479.2019.1636285
org/10.1080/09613218.2017.1356126
Amasyali, K., & El-Gohary, N. (2021). Machine learning for occupant-behavior-sensitive
Mellado, F., Wong, P. F., Amano, K., Johnson, C., & Lou, E. C. W. (2020). Digitisation of
cooling energy consumption prediction in office buildings. Renewable and Sustainable
existing buildings to support building assessment schemes: Viability of automated
Energy Reviews, 142, 110714-1–110714-10. https://doi.org/10.1016/j.
sustainability-led design scan-to-BIM process. Architectural Engineering and Design
rser.2021.110714
Management, 16(2), 84–99. https://doi.org/10.1080/17452007.2019.1674126
Becerik-Gerber, B., Jazizadeh, F., Li, N., & Calis, G. (2012). Application areas and data
Mohd Noor, S., Shing Mei, C., Syahrizal Ibrahim, I., Nabilah Sarbini, N., Hanim Osman,
requirements for BIM-enabled facilities management. Journal of Construction
M., & Azeyah Khiyon, N. (2019). Heritage building condition assessment: A case
Engineering and Management, 138(8), 431–442. https://doi: 10.1061/(ASCE)
study from Johor Bahru, Malaysia. In IOP Conference Series: Earth and
CO.1943-7862.0000433.
Environmental Science (Vol. 220, No. 1, p. 012024). IOP Publishing. https://doi: 10
BSI ISO. (2012). Buildings and constructed assets - Service life planning Part 2 : Service
.1088/1755-1315/220/1/012024.
life prediction procedures. International Standards. https://www.iso.org/standard/5
1826.html.

11
A. Gouda Mohamed and M. Marzouk Expert Systems With Applications 186 (2021) 115743

Olden, J. D., Joy, M. K., & Death, R. G. (2004). An accurate comparison of methods for Journal of Building Engineering, 20, 520–530. https://doi.org/10.1016/j.
quantifying variable importance in artificial neural networks using simulated data. jobe.2018.08.018
Ecological modelling, 178(3-4), 389-397. doi.org/10.1016/j. Saffariha, M., Jahani, A., Jahani, R., & Latif, S. (2021). Prediction of hypericin content in
ecolmodel.2004.03.013. Hypericum perforatum L. in different ecological habitat using artificial neural
Piaia, E., Maietti, F., Di Giulio, R., Schippers-Trifan, O., Van Delft, A., Bruinenberg, S., & networks. Plant Methods, 17(1), 1-17. doi.org/10.1186/s13007-021-00710-z.
Olivadese, R. (2020). BIM-based cultural heritage asset management tool. innovative Sandoval, C., Valledor, R., & Lopez-Garcia, D. (2017). Numerical assessment of
solution to orient the preservation and valorization of historic buildings. accumulated seismic damage in a historic masonry building. A case study.
International Journal of Architectural Heritage, 1–24. https://doi.org/10.1080/ International Journal of. Architectural Heritage, 11(8), 1177–1194. https://doi.org/
15583058.2020.1734686 10.1080/15583058.2017.1356945
Roweis, S. (1996). Levenberg-marquardt optimization. Lecture Notes, University of Xu, S., & Chen, L. (2008). A novel approach for determining the optimal number of
Toronto, Toronto, Canada. https://cs.nyu.edu/~roweis/notes/lm.pdf. hidden layer neurons for FNN’s and its application in data mining. In 5th
Sadick, A. M., & Issa, M. H. (2018). Assessing physical conditions of indoor space International Conference on Information Technology and Applications. ICITA2008
enclosing elements in schools in relation to their indoor environmental quality. ISBN: 978-0-9803267-2-7 (pp. 683-686). https://eprints.utas.edu.au/6995/1/02-au-
xu.pdf.

12