Advanced Engineering Informatics 52 (2022) 101631

Contents lists available at ScienceDirect

Advanced Engineering Informatics

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

Full length article

A machine learning approach to predict production time using real-time

RFID data in industrialized building construction
Osama Mohsen *, Yasser Mohamed, Mohamed Al-Hussein
Department of Civil and Environmental Engineering, University of Alberta, Edmonton, Alberta T6G 1H9, Canada

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

Keywords: Industrialized building construction is an approach that integrates manufacturing techniques into construction
Industrialized building construction projects to achieve improved quality, shortened project duration, and enhanced schedule predictability. Time
Prefabricated construction savings result from concurrently carrying out factory operations and site preparation activities. In an industri­
Production time
alized building construction factory, the accurate prediction of production cycle time is crucial to reap the
Time prediction
advantage of improved schedule predictability leading to enhanced production planning and control. With the
RFID
Machine learning large amount of data being generated as part of the daily operations within such a factory, the present study
proposes a machine learning approach to accurately estimate production time using (1) the physical charac­
teristics of building components, (2) the real-time tracking data gathered using a radio frequency identification
system, and (3) a set of engineered features constructed to capture the real-time loading conditions of the job
shop. The results show a mean absolute percentage error and correlation coefficient of 11% and 0.80, respec­
tively, between the actual and predicted values when using random forest models. The results confirm the sig­
nificant effects of including shop utilization features in model training and suggest that predicting production
time can be reasonably achieved.

1. Introduction which encompasses automation, data exchange, and digitization in

Various terms, such as prefabricated construction, offsite construc­
tion, and modular construction, to name a few [1], all refer to the same
approach: a construction process that applies manufacturing principles
and techniques to the construction industry where products (in this case,
building components) go through a particular life-cycle from concept, to
design, to planning, to manufacturing, and then on-site assembly, as
shown in Fig. 1. The term “industrialized building construction” (IBC) is
used herein to denote any of the terms above.
Industrialized building construction (IBC) approaches, including
offsite, prefab, and modularization, are being used by more than 80% of
contractors in the U.S., underscoring the potential for future market
growth [2]. In fact, it is projected that modular construction will reach
$130 billion in market value in the U.S. and Europe by 2030 [3]. In
Canada, modular construction accounted for a market value of more
than $1 billion in 2019 [4]. Building construction enterprises have come
to recognize the value of information technologies and knowledge dis­
covery approaches [5]. The most recent evidence is the emerging
paradigm of Construction 4.0, a term coined in reference to Industry 4.0,
manufacturing [6–8]. Critical reviews of IBC were conducted that have
identified the major research topics, trends, and challenges in this field
[1,9,10]. Project planning and design, including workflow optimization
and project delivery process, have been among the most frequently
studied research themes in IBC over the past decade [11].
An IBC factory is one of the complex manufacturing environments
due to numerous factors, namely, (i) the diversity of building compo­
nents that leads, occasionally, to the complete customization of building
modules; (ii) the integration of many sub-systems (structural, electro­
mechanical, communications, finishing) that must be included in the
building; (iii) the large and heavy machinery and components involved
in the manufacturing process; and (iv) the numerous interrelated ac­
tivities including production, assembly, finishing, delivery and on-site
installation [12].
In any manufacturing process, as well as in the context of IBC, a job’s
cycle time (CT) is a key performance metric [13], defined as the time
taken to complete the production process of one product from start to
finish. A closely related term is lead time (LT), which is defined as the
time taken from order placement until the order is delivered to the

* Corresponding author.
E-mail address: omohsen@ualberta.ca (O. Mohsen).

https://doi.org/10.1016/j.aei.2022.101631
Received 4 August 2021; Received in revised form 15 April 2022; Accepted 9 May 2022
Available online 27 May 2022
1474-0346/© 2022 Elsevier Ltd. All rights reserved.
O. Mohsen et al.
customer. Accurate prediction of cycle time is crucial when determining
customer delivery dates, when scheduling resources and actions, and
when controlling daily operations. Scheduling in IBC depends on the
schedule of on-site activities, the factory production processes, and the
expected delivery dates of raw materials from suppliers. As such, pro­
duction CT constitutes a critical component in this supply chain, from
raw material to the end product. A more predictable production
schedule allows the production/project manager to mitigate delays and
implement alternative actions should there be any unexpected inter­
ruption to factory production. In addition, accurate prediction of CT
enables effective management of the fabrication shop’s capacity and
workload.
In the literature, several methods have been developed to predict
cycle time [14]. Simulation has been the most widely used approach in
that it can realistically capture the complexities of manufacturing;
however, practical disadvantages include the expense of model devel­
opment and maintenance and the emphasis on average performance
rather than on the prediction of CT for an individual product. On the
other hand, the most fundamental analytical approach to predicting the
average CT is Little’s law [15], where cycle time is deterministically
calculated. Such a static model does not provide good predictions
because it does not capture the stochastic and make-to-order nature of
IBC manufacturing; nevertheless, it can be used as a benchmark. Despite
a large amount of data being collected in IBC factories and the advances
over the last decade to collect and analyze data automatically, produc­
tion data are not being fully utilized to improve decision-making in IBC.
The main goal of the present study is to devise a framework for ac­
curate prediction of CT that exploits the knowledge embedded in the
raw data from production tracking systems. The present study represents
a middle approach between simple statistical methods (e.g., Little’s law)
and more sophisticated simulation modeling techniques. More specif­
ically, the present study aims at extracting knowledge about the status of
the workload of the production environment that is not readily attain­
able using existing production tracking data. Such extracted knowledge
is then used to augment the inputs to machine learning (ML) models to
improve the prediction of CT.
The remainder of this paper is organized as follows. Section 2 pro­
vides a brief review of the relevant literature concerning IBC and the
application of ML in both construction and production cycle time pre­
diction. The study objectives and research methods are presented in
Sections 3 and 4. Section 5 describes the data preparation and feature
engineering experiments, while Section 6 describes our experimentation
with ML models, including a summary and discussion of the results.
Lastly, the conclusions, including limitations and future work, are pre­
sented in Section 7.
2. Related literature
2.1. Knowledge discovery in construction
The term “Knowledge Discovery in Databases” (KDD) had emerged
by the late-1980s to collectively refer to the various methods for iden­
tifying valid, practical, and nontrivial patterns and similarities in raw
Fig. 1. Simplified high-level workflow in industrialized building construction.
2
Advanced Engineering Informatics 52 (2022) 101631
data, such as knowledge extraction, information discovery, information
harvesting, and data pattern processing [16,17]. Several studies have
investigated the application of KDD in the construction industry. For
example, KDD processes are used to identify the causes of delays in
drainage pipeline installation, focusing mainly on data preparation [18];
to identify the key factors that contribute to delays in construction
projects [19]; to estimate future resource requirements in multi-project
industrial construction [20], and to forecast construction project time
and cost using support vector regression methods [21]. A detailed
literature review is presented of the current applications and future
potential of data analytics in construction, affirming that adopting these
technologies in construction lags its broader application in other fields
[22]. More recently, various ML algorithms are being used in con­
struction to detect non-certified work on sites [23], classify workers
encountering high fatality risk accidents [24], verify the personal pro­
tective equipment compliance of workers [25,26], and forecast profit
margins for new projects using an applied machine learning process
[27]. It is evident from the literature that ML application in many areas
of the construction industry is gaining increasing acceptance and
adoption; however, these techniques have seen limited applications in
project planning and scheduling in IBC.
To enable the application of KDD and ML, the availability of large
datasets is crucial. One technology for automatic real-time data acqui­
sition typical to IBC is the radio frequency identification (RFID) tracking
system. RFID is used as a sensing mechanism to locate product tags using
the electromagnetic field of antennas connected to RFID readers, which
populate a central database with tracking information. Within any
specific setup, each tag contains unique identification information. The
RFID system setup used in the present study is shown in Fig. 3; it consists
of three main components: (a) scanning antennas, (b) a transceiver that
reads and interprets product timestamps, and (c) transponders/tags that
have been programmed with information and are attached to the objects
being tracked (i.e., wall panels in the present study). Valero [28] and
Valero and Adán [29] provided a comprehensive review of recent RFID
applications in construction.
2.2. Machine learning for cycle time predictions
An overview of machine learning applications in manufacturing was
conducted and indicated that future research would be focused on
design, shop floor control, supply chain management, and developing
solutions that can be easily integrated with existing systems [30].
Similarly, Choudhary et al. [31] reviewed the literature about data
mining in manufacturing, focusing on categorizing the articles into the
five primary functions of data mining, namely, description, association,
classification, prediction, and clustering. In recent years, a few studies
have focused on predicting LT and CT using ML approaches in
manufacturing environments other than IBC. For example, a regression
tree is used to estimate LT in a hypothetical make-to-order
manufacturing factory [32], a hybrid method of clustering and regres­
sion trees is used to predict CT in a semiconductor factory [33], and in
the context of wafer fabrication, a CT range is estimated using fuzzy
backpropagation network, principal component analysis, and fuzzy c-
O. Mohsen et al.
means [34]. Another study used genetic programming to predict process
CT based on system status information [35]. More recently, Gyulai et al.
[36] compared analytical and ML techniques for LT prediction in the
optics industry, and Lingitz et al. [37] compared the accuracy of several
regression algorithms when predicting LT in semiconductor
manufacturing; both suggested that random forests are the highest ac­
curacy models in terms of lower error measures.
Although IBC is an interdisciplinary field at the interface of
manufacturing and construction, relatively few ML studies within this
domain have been conducted. IBC typically employs a make-to-order
production approach where products (i.e., building components or
modules) are only built after a confirmed customer order. Products are
also highly customized, and their production must be coordinated with
site delivery; all the above factors make the prediction of production
cycle time a complex task. Previous studies applying ML to predict CT
have mainly used simulated data. In contrast, the real-time RFID data
utilized in this study is more prone to noise than simulated data and,
hence, is more challenging to manipulate and prepare for modeling.
Actual RFID production data also provides a more accurate and realistic
knowledge representation of the production system.
3. Study objectives
The present study builds upon the previous research to achieve an
accurate prediction of cycle time in manufacturing environments using
ML in the context of IBC, which differs from other manufacturing en­
vironments as described above. The goal is to develop data-driven pre­
dictive models to estimate the cycle time required to accomplish a highly
customized production process in a wall panel manufacturing facility
and to explore the effect of different product and production process
attributes on the predictability of CT. The primary input data to the
target models include 1) physical characteristics of wall panels obtained
from building information models (BIM), 2) RFID timestamps used to
track panel production, and 3) a set of engineered features that depicts
the real-time loading conditions of the job shop.
More specifically, the present study seeks to achieve three objectives:
Fig. 2. Framework of research approach.
3
Advanced Engineering Informatics 52 (2022) 101631
(1) to accurately predict the CT of a production process using ML models
and real-time data; (2) to examine the effect of augmenting the dataset
with domain-specific engineered features that capture the real-time
loading conditions of the fabrication shop; and (3) to investigate the
effect of different grouping strategies, used to combine several features,
on the prediction accuracy.
4. Research approach and methods
The research approach is based on an integrated KDD framework
consisting of several interconnected phases, as shown in Fig. 2. This
framework is developed based upon commonly used frameworks for
data mining, such as the cross-industry standard process model for data
mining (CRISP-DM) [38]; Oracle’s architecture development process
[39]; the work published by Fayyad et al. [16], Davenport et al. [40],
Dietrich et al. [41]; and the IBM foundational methodology for data
science [42].
Phase (1) involves identifying and collecting data by determining
some criteria for data inclusion suitable for the present analysis. This
phase comprises two main steps: identifying data and acquisition
mechanisms and identifying appropriate data storage media. In this
research, data is acquired from a central database containing (a) the
physical properties of wall panels as stored in BIM and (b) the RFID
timestamps captured by the panel tracking system.
Phase (2) involves the iterative tasks of cleaning and preparing the
data once it has been captured and stored. Essentially, this phase en­
deavors to answer the following two questions: Is the data collected
representative of the problem to be solved? What should efficient
cleaning steps be followed to represent the data in a format suitable for
prediction models?
Phase (3) is subdivided into two tasks. The first is the construction of
engineered features to reflect the real-time loading conditions of the job
shop. In the second task, the newly generated features are combined
with the cleaned dataset from phase (2), and the resulting integrated
data is then directed back through phase (2) for further preparation. The
output of this iterative process is a “tidy dataset.” Tidy data, it should be
O. Mohsen et al.
noted, is a common term used in data science and has been formally
defined by Wickham [43] as “a standard way of mapping the meaning of
a dataset to its structure” in which each variable/feature forms a col­
umn, each observation/record forms a row, and each type of observa­
tional unit forms a table. To realize the CT, two factors should be
quantified: the quantified product composition (PC), which represents
physical properties, and the shop loading conditions (SL). Accordingly,
based on PC and SL, the predicted value of CT, CTpred, can be formulated
as shown in Equation (1):
CT pred = f (PC, SL) (1)
where f stands for the mapping function from PC and SL to CTpred.
Phase (4) involves mining the dataset by applying various ML
regression techniques to the different versions of the tidy data. These
versions are developed for the purpose of experimenting with different
aspects of data aggregation and integration. Phase (5) involves evalu­
ating the developed models using five performance metrics. Model
evaluation typically consists of two main phases: the diagnostic phase,
which ensures the model is working as intended, and the statistical
testing phase, which is applied to ensure that the data is being properly
handled and interpreted within the model.
The following sections describe each of these phases in greater detail,
including the results.
5. Identifying and preparing the data
5.1. Process description
The production of prefabricated wall panels at an IBC factory located
in Edmonton, Canada, is analyzed. The production floor layout is shown
in Fig. 3, where the scope of the present study is restricted to the
multiwall panel (MWP) production route, which consists of four
consecutive workstations: framing station (FS), buffer table (BT),
sheathing station (SS), and multifunction bridge (MFB). A wall panel
goes through these stations in sequence, and no two panels are processed
at one station simultaneously. RFID antennas are located between
consecutive workstations where timestamps are recorded by reading
RFID tags attached to wall panels as they pass by.
Fig. 3. Production floor layout and location of RFID antennas between workstations.
4
Advanced Engineering Informatics 52 (2022) 101631
The floor layout is configured such that each MWP is framed, prep­
ped, sheathed, and nailed between antenna locations A1 to A5. Then,
each MWP is cut into several single panels and transferred to a butterfly
table, where antenna A6 records when the single wall panel starts the
following production phase. This study focuses on the first phase of
production: the MWP production starting at antenna A1 (located at the
front of FS) and ending at antenna A5 (located at the end of MFB).
5.2. Data acquisition and description
The raw data analyzed in this study covers a period of 42 months
between February 2015 and August 2018. The raw data is composed
primarily of two datasets. The first dataset (DF1) is the “RFID Readings”
dataset that contains timestamps for each MWP along the production
route. It consists of 416,948 records and ten attributes. The second
dataset (DF2) is the “Multipanel” dataset that contains descriptive at­
tributes of each MWP, including the physical properties of the panels. It
consists of 39,703 records and 37 attributes. Fig. 4 illustrates the flow of
the data identification and data cleaning steps. Initial data identification
is performed to set the study’s limits and to include only the information
relevant to the current analysis. These steps are expert rules depending
mainly on the specific application. For example, in DF1, steps include (a)
removing records before September 11, 2015, since the factory layout
and RFID system setup was changed, and (b) only keeping RFID records
of antennas A1 through A5 as these correspond to the scope of the study.
As for DF2, the steps include (a) removing panels that are manually
manufactured and (b) excluding any attribute where all data are
missing, or there is only one dominant value.
In addition, as part of this phase of the methodology, descriptive
statistics and data visualization are applied to assess the content and
quality and to gain initial insights into the data. Fig. 5 shows the total
number of RFID readings per month, per week, and using a 14-day
rolling window. Periods during which there is a low number of RFID
readings, such as from July 2016 to September 2016, might indicate low
productivity periods due to market recessions. Such periods represent
abnormal operating conditions and should be excluded from the analysis
as they may skew the results.
However, a more consistent indicator of daily productivity is to
count the number of panels produced each day over the study period.
O. Mohsen et al. Advanced Engineering Informatics 52 (2022) 101631

Fig. 4. Flowchart of initial data identification and data preprocessing steps.

Fig. 5. Number of RFID readings per month, per week, and 14-days rolling window.

For example, there were approximately 1,400 panels manufactured
when the daily production rate was between 35 and 40 panels/day,
Fig. 6. Working days are excluded if the daily production is less than five
or more than 75 panels/day as production beyond these limits repre­
sents abnormal operating conditions. Approximately 97% of the total
study period is covered by excluding these data. This exercise aims to
determine which production days represent anomalous operations and,
therefore, exclude the records during these days from further analysis.
Furthermore, Table 1 summarizes the statistics for the attributes in
DF2. Some attributes have extreme values compared to the mean and
median. For example, ‘NailCount’ has an extreme maximum value of
2,466, which is considered an outlier. It can also be noted that the
attribute ‘LargeDoors’ is expected to have minimal influence on the
predictability of CT as most of the panels have no large doors installed.
This can be inferred from the value of zero for the mean, standard de­
viation, and interquartile range.
5.3. Data cleaning and preprocessing
Data preprocessing is an essential step in improving the quality of the
data itself and, consequently, the quality of the data mining results. Data
preprocessing techniques include cleaning, integration, transformation,
and reduction [44]. The main preprocessing steps applied to each
dataset in the present study (i.e., DF1 and DF2) are summarized in
Table 2.
As an MWP passes over the RFID antenna, a timestamp is recorded in

5
O. Mohsen et al. Advanced Engineering Informatics 52 (2022) 101631

Fig. 6. Total panel counts for each daily production throughput.

Table 1
Statistics summary of attributes in DF2 dataset.
mean std min 25% 50% 75% max

Length 8999 3315 1203 6087 10,404 11,878 12,338

Width 2525 158 1607 2467 2467 2505 3235
Height 117 29 89 89 140 140 314
Window 0 1 0 0 0 0 9
LargeWindow 0 1 0 0 0 0 5
Door 1 1 0 0 0 1 10
LargeDoor 0 0 0 0 0 0 3
Sheetfull 1 2 0 0 0 0 17
SheetPartial 3 4 0 0 0 4 31
Cutzone 2 2 0 0 1 3 17
Drillhole 4 3 0 2 4 6 26
Stud 15 7 0 9 15 21 215
DStud 0 0 0 0 0 0 6
LStud 2 2 0 0 1 3 18
MStud 0 1 0 0 0 0 14
Block 1 3 0 0 0 0 42
Backing 2 4 0 0 0 3 50
NailCount 327 401 0 0 190 566 2466
Nailline 33 37 0 0 20 63 447

the attribute “InitialReadDateTime.” When the panel leaves the RFID

antenna range, the attribute “LastReadDateTime” records another
timestamp. Ideally, when there is no interruption or waiting in the
production line, the two timestamps should be similar, indicating that
the panel has not been sitting idle over the antenna. The idle time (IT)
for panel i can be defined by Equation (2):
ITpanel = tlastRead, i − tinitialRead, i
The intermediate CT can be described as the processing time plus the
idle time. Time differences between consecutive antenna locations
indicate the time during which the panel is processed at each of the four
intermediate stations. Total CT is calculated, according to Equation (3),
as the sum of these individual station times:
∑
4
ICT = tinitialRead, i+1 − tinitialRead, i (3)
i=1
where tinitialRead,i+1 , and tinitialRead,i are the timestamps at the end and start
antenna locations, respectively. It should be noted that the cycle time as
measured in the present study encompasses processing time, idle time,
(2) and the time spent on regular breaks during a typical working shift.
Fig. 7 shows the histogram and the kernel density smoothing curve of
both the measured CT (ActualCT) and the CT without the break times
(ActualCT-BT) for all MWPs. There are three break times during a typical
working shift in the case production factory: two 15-minute coffee
breaks and one 30-minute lunch break. If the production of the panel
falls within these break times, the measured CT is reduced accordingly.
In this manner, actual production times are calculated, and any adverse

6
O. Mohsen et al. Advanced Engineering Informatics 52 (2022) 101631

Table 2
Main preprocessing steps applied to DF1 and DF2.
Preprocessing Step
“RFID Readings” dataset (DF1)
1 Keeping “InitialReadDateTime” and
“LastReadDateTime”.
2 Discarding “LocationSourceAntenna”,
“LocationTagID”, “TagID”, “WallNumber”,
and “LastReadDate”.
3 Transforming the dataset: each row stores
timestamps for each MWP at each station;
the original dataset stores each timestamp
for each panel at each station in a separate
row.
4 Replacing a missing value of timestamps at
individual stations with the mean value
provided that the timestamps at all other
stations are not missing.
5 Records, where the production of a panel
started on one day and finished on a
subsequent day, are discarded.
“Multipanel” dataset (DF2)
1 Keeping the dimension attributes
(“Length”, “Width”, and “Height”) within a
specific range. Panel lengths range
between 1,203 and 12,338 mm; widths
range between 1,607 and 3,235 mm; and
heights range between 89 and 314 mm.
2 Nominal values of the attribute ‘Type’ are
converted to numbers using the “dummy
coding” technique, commonly known as
binarization. There are five unique values:
‘EXT’, ‘INT’, ‘STR’, ‘GAR’, and ‘MEC’.
3 Discarding the attributes “Job”,
“Component”, “wall”, “Siding”,
“SidingLine”, “Model”, “Floor”, “Unit”,
“GarageDoor”, “Sequence”,
“Basementwall”, “position”, and
“ProductionJob”.
4 Missing values are either replaced using
functions defined by fitting curves to the
known values or discarded if they account
for more than 90% of the total count of an
attribute.
Reasoning
Initial and last timestamps used
for time difference calculations.
Contain either redundant or
irrelevant information.
To prepare the data in a tidy
format, and to facilitate CT
calculations.
Replacing missing values so that
the dataset is consistent and
complete.
Discarding outliers for consistent
cycle time calculations.
Only panel manufactured at the
automatic production line are
kept. Smaller panels, that are
manually produced, are
discarded.
For all unique values of the
nominal attribute, a new
attribute is created. The attribute
corresponding to the value of the
record gets a value of 1; other
attributes are set to 0.
These attributes are irrelevant as
they contain redundant
information, have one unique
value, have few prevalent values,
or have many missing values.
Replacing missing values so that
the dataset is consistent and
complete.
attributes of panels that preceded the panel, i, in the production line but
whose production is still underway, i.e., work in progress (WIP). These
WIP panels represent the current state of system utilization at the
moment production of the current panel, i, begins. If we let Xn×m be a
matrix in which each row, n, represents one WIP panel preceding the
current panel, and each column, m, represents an existing feature, x. If
we consider θ to be a binary vector of size m, where each element cor­
responds to one of the existing features, then a value of 1 corresponds to
that feature being included when calculating the new engineered feature
of the current panel; otherwise, the feature is ignored. This can be
expressed by Equation (4):
⎡ ⎤
x11 &x12 ⋯&x1n
Si = θ⋅X = [θ1 θ2 ⋯ θm ]⎣
T
⋮&⋱&⋮ ⎦ (4)
xm1 &xm2 ⋯&xnm
SL features are developed for the present study to capture the system
status based on the total production during one work shift. The con­
structed features are ˝WPi ˝, ˝WPLengthi ˝ and ˝WPTimei ˝ and they store
the number, length, and processing time, respectively, of the WIP
panels. These features represent the set, Si, in Equation (4), with each
being generated by setting the appropriate values of the binary vector, θ.
Given i = {1, 2, …, n}, where n is the total number of WIP associated
with the current panel, the attributes above are calculated based on
Equations (5) through (8):
WPi = |A| = cardinality of set A (5)
{ }
A = (x, y) : ∃ x, y ∈ DF 2 | x.T1 > y.T1 and x.T1 < max(y.T2 , y.T3 , y.T4 )
(6)
∑
m
{ }
WPLengthi = x.Lj : ∃ x ∈ A where m = |A| (7)
j=1
∑
m
{ }
WPTimei = x.PT j : ∃ x ∈ A where m = |A| (8)
j=1
where: (x, y) the pair of timestamps; x is the timestamp at antenna A1 of
the current panel i, and y is the timestamp of any other panel concur­
rently being processed;

effect of break times on the accuracy of the CT predictions is mitigated.
Hence, ‘ActualCT-BT’ rather than ‘ActualCT’ is used to develop the
prediction models.
5.4. Feature engineering and selection
Feature engineering (FE) has been widely used in data-driven
modeling to improve the quality of feature sets. FE mainly encom­
passes feature selection, feature extraction, and feature construction
[45]. In the present research, each record in the tidy dataset stores the
physical properties of a single multiwall panel and the respective actual
cycle time, which is the target value to be predicted. One of the main
contributions of the present study is to capture the system status by
constructing new features that objectively measure the loading condi­
tions of the job shop (i.e., SL features). In essence, these new features are
developed to augment the existing features of a product unit based on
the units that immediately precede it along the production line. The CT
in the IBC is largely governed by the SL features (i.e., a measure of how
busy the fabrication floor is when the production of a new wall panel
begins). The inclusion of SL features in model development significantly
enhances the model accuracy and, hence, CT predictability. This is
mainly due to the fact that these features capture shop utilization, which
otherwise (i.e., if only the physical features such as length, width, etc.,
are used) would be unaccounted for.
Let Si represents the set of newly constructed features of a product, i.
Each feature ski ∈ Si is generated in such a manner as to capture
DF denotes the tidy dataset;
Tk is the timestamp at antenna location Ak, k = {1, 2, 3, 4};
x.Lj is the length of panels that satisfy the condition; and
x.PTj is the processing time of the panels that satisfy the condition.
A Python script is developed to iterate records in DF and create the
engineered features. The schematic in Fig. 8 represents the process of
calculating the length of WIP panels, ˝WPLengthi ˝, that precede panel i
in a typical work shift. In this example, tp,fs is the timestamp associated
with the panel when its production begins at the framing station, while
tp− 1,ss and tp− 2,mfb are the timestamps for two panels that are ahead of the
current panel but are still in production.
Moreover, based on the newly constructed features, different ver­
sions of the dataset are generated, each containing a different set of
attributes to examine the effect of varying attribute subsets on the pre­
diction accuracy. These new attributes are created by combining exist­
ing physical characteristics to reduce both the dataset dimensions and to
examine the effect of features granularity on the accuracy of ML models.
Table 3 summarizes the experiments conducted with different attribute
subsets. Results are summarized in Table 4 and discussed in Section
6.3.3.

7
O. Mohsen et al. Advanced Engineering Informatics 52 (2022) 101631

Fig. 7. Total cycle time (between A1 and A5 in minutes).

Fig. 8. Engineered feature calculations for Panel i at the time its production begins.

6. Experiments, results, and discussion

Table 3
Subsets created by combining attributes of the tidy dataset.
6.1. Machine learning regression models
Name Size (records × Description
attributes)
The three main categories of ML are supervised learning, unsuper­
Raw data 1,094 × 23 Contains only physical properties of panels vised learning, and reinforcement learning. The CT prediction under
Engineered 1,094 × 28 In addition to the physical properties of
investigation in the present study is a regression supervised learning
panels, it contains engineered features
capturing the system status (WP, WPLength,
task, and there exist numerous ML models to predict a continuous var­
WPTime, Panels/Day). iable of this nature. Among these, four are selected for our study: linear
Engineered 1,094 × 20 Combining similar properties of a panel into regression (LR), k-nearest neighbor (KNN), random forest (RF), and
combined1 single attribute: neural networks (NN). A successful model should provide reasonably
SheetFP = Sheetfull + SheetPartial
accurate estimates of the CT to be used in factory planning and should be
TotalStuds = Stud + 2*DStud + 2*LStud +
3*MStud easy to interpret. Representative references describing these methods
BlockBack = Block + Backing include Hastie et al. [46], Witten et al. [47], and, in the context of Py­
totalWD = Window + LargeWindow + Door thon language, Raschka and Mirjalili [48]. For the treatment of regres­
+ LargeDoor
sion trees and random forests, the interested reader may refer to
Engineered 1,094 × 17 Combining all types of components
combined2 (openings, stud, sheets, supportive elements)
Breiman et al. [49] and Breiman [50]. The four models mentioned above
into one attribute: are selected from among the many available ML models available for
Components = totalWD + TotalStuds + regression based on the following considerations: (1) they range from
SheetFP + BlockBack simple to implement and easily interpretable (i.e., LR and KNN) to more
sophisticated black-box models (i.e., RF and NN); (2) they have been

8
O. Mohsen et al. Advanced Engineering Informatics 52 (2022) 101631

Table 4 models include support vector machines and deep learning networks,
Performance measures of the models for the datasets described in Table 3. which were not thoroughly examined due to an initial screening that
ML model LR KNN RF NN discarded the models that had a long run time given the large size of the
Hyperparameters – k = 11, p 400 trees, (52, 52) dataset used in the present study. It should also be noted that the main
=2 MAE 2,000 itr. focus of the present study is to evaluate the impact of engineered fea­
Raw data R2 on 0.27 0.91 0.80 0.24 tures on ML model performance in general.
training With the tidy dataset being developed, the selected models are
MAE 9.2 6.8 5.8 7.7 trained and tested. Per Witten et al. [47], the dataset is divided into
MedAE 6.7 3.6 3.1 4.8
MAPE 23.6% 17.3% 17.0% 19.7%
three subsets: a training set, a validation set, and a test set. The test set
CC − 0.03 0.13 0.47 0.19 contains the records from a given day for which the CT is to be predicted.
Engineered R2 on 0.46 0.99 0.90 0.42 The training and validation sets contain the production data from pre­
training vious days, and this data is used for training and validation using a 5-fold
MAE 7.1 6.1 5.3 6.7
cross-validation technique, as shown in Fig. 9. The models are trained
MedAE 5.6 3.4 3.5 4.4
MAPE 17.8% 14.9% 13.7% 16.4% and tested using Scikit-Learn, an open-source Python library for ML
CC 0.34 0.47 0.71 0.44 [60].
Engineered R2 on 0.43 0.99 0.90 0.40
combined 1 training
MAE 6.6 6.1 4.9 6.9
6.2. Comparison of model performances
MedAE 4.2 3.4 2.8 5.0
MAPE 16.4% 14.9% 12.3% 16.7%
CC 0.40 0.47 0.73 0.39 The performance of the models is evaluated based on the following
Engineered R2 on 0.41 0.99 0.90 0.39 measures: mean absolute error (MAE), median absolute error (MedAE),
combined 2 training mean absolute percentage error (MAPE), and correlation coefficient
MAE 6.1 6.0 4.4 6.5
(CC), which are calculated as shown in Equations (9) through (12):
MedAE 4.3 3.3 3.2 4.3
MAPE 15.0% 14.9% 11.0% 15.6%
1∑ n
CC 0.53 0.47 0.80 0.44 Mean Absolute Error (MAE) = |ai − pi | (9)
n i=1

implemented in previous studies to predict production cycle time as Median Absolute Error (MedAE) = mediani=1, n (|ai − pi |) (10)
discussed earlier in Section 2.2; and (3) they have been widely used in
n ⃒ ⃒
the literature to investigate construction problems. For example, LR was 100 ∑ ⃒a i − pi ⃒
Mean Absolute Percentage Error (MAPE) = ⃒ ⃒ (11)
employed to quantify construction delays [51], and to predict con­ n i=1 ⃒ ai ⃒
struction costs [52,53]; KNN methods have been applied in the predic­
∑ ∑ ∑
tion of construction cost index [54], and planning of construction n ai pi − ai pi
technical specifications for deep foundations [55]; RF models were used Correlation Coefficient (CC) = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑ 2 ∑ 2 ̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
̅ √ ∑ ∑ ̅
n ai − ( ai ) n p2i − ( pi )2
to predict ground surface settlement due to tunnel construction [56],
and to improve the accuracy of BIM-enabled clash detection [57]; NN (12)
was used to predict labor productivity [58], and estimate schedule to
completion of construction projects [59]. Other potentially powerful where ai is the actual CT of panel i, pi is the predicted CT of panel i, and n
is the total number of panels in the data sample. These measures are

Fig. 9. Data splitting and model evaluation.

9
O. Mohsen et al.
selected as per the literature analysis on performance metrics in ML
regression algorithms [61].
The resulting performance measures for the four datasets described
in Table 3 are given in Table 4. The four models’ hyperparameters have
been tuned using the training and validation datasets to attain better
performing and more robust models. The performance of ML algorithms
depends highly on finding an optimal set of hyperparameters. Various
strategies can be used to search for the best combination of hyper­
parameters; these include grid search and random search [62]. In the
present study, a combination of manual and brute force search using the
GridSearchCV method available from the Scikit-Learn library is used
to search for the optimal combination of a predefined set of values for
each model’s hyperparameters. The most important values are listed in
Table 4 under the corresponding model’s name.
6.3. Observations on model performances
6.3.1. General trends
Referring to Table 4, a distinct difference can be observed between
the MAE and the MedAE values. The MedAE is smaller than the MAE for
all models, a trend aligned with the distribution pattern of actual CT,
which is positively skewed (moderately skewed right), as shown in
Fig. 7. This behavior is attributable to the presence of a few outlier re­
cords that are significantly higher than the average values. These
extreme values cause the distribution of CT to be widespread. Another
observation is that achieving a higher value of R2 on the training set does
not necessarily translate into better performance on the testing set, as
can be seen for the performance of KNN, where the model is observed to
have overfitted the training data. As this example underscores, the
model performance should not be evaluated based on the metrics of the
training set.
Across all models and datasets, the MAPE range between 24%
(worst) and 11% (best), and the CC range between 0.0 (no correlation)
and 0.8 (strong positive correlation). Overall, the RF models have better
performance measures in terms of lower MAE, MedAE, MAPE, and
higher correlation coefficients. However, it should be noted that RF
requires slightly more computational time. An advantage of RF over
other methods is the ability to handle both categorical and continuous
attributes, and attributes need not be scaled to common units. When
testing several configurations of RF models, using approximately 400
trees and MAE as the evaluation criterion when splitting the nodes
yielded the best performance.
6.3.2. Engineered features
The production system utilization status is valuable information to
add to the ML models to improve the predictability of CT. The raw
dataset, ‘Raw data’ in Table 3, includes only the physical properties of
the panels; the engineered features that represent the system utilization
status are only included in the other three datasets. Using the raw data
without the SL features leads to inferior model performance. The MAPE
is found to be as high as 24%, while the CC is low for all models.
Meanwhile, an improvement in MAPE of as much as 9% is achieved
when augmenting the datasets with the proposed SL features. Moreover,
the correlation coefficients have increased for all models to as high as
0.80 in the case of the RF model.
Intuitively, augmenting the feature set with engineered features
should result in better predictability of the developed ML models
because these features provide additional information about how the
production floor is being utilized when the production of a new product
begins. Based on the presented results, this assumption is proven to be
valid, with performance improvements having been consistently ach­
ieved for all models.
6.3.3. Combining features
The performance of all models does slightly improve by reducing the
granularity of the features. For example, the mean absolute percentage
10
Advanced Engineering Informatics 52 (2022) 101631
error of the LR and RF models dropped from 17.8% and 13.7% in the
case of the first version of the engineered dataset, ‘Engineered’, to 15.0%
and 11.0% in the case of the reduced dataset, ‘Engineered combined 2’,
respectively. In the aggregated dataset ‘Engineered combined 2’, all
components of a wall panel (studs, windows, doors, backing, and
sheathing sheets) are combined and represented by one feature, namely,
“Components.”
In addition to requiring less computational time, it can be concluded
that reducing the dimensionality of the data by combining similar at­
tributes does improve the performance in terms of lower prediction er­
rors and higher correlation coefficients. This might be because the
combined features have a weak correlation with the target value when
computed individually. In contrast, when these features are aggregated,
the correlation with the target value increases. For example, when
comparing the correlation coefficients in Table 4, we observe an in­
crease of 19% for LR models and 9% for RF models between the second
dataset ‘Engineered’ and the fourth dataset ‘Engineered combined 2’.
However, in the case of KNN and NN models, we observe no improve­
ment. It should be noted that these observations are specific to the
current case study and are not necessarily true for other cases. Further
investigation is required to explore the effect of combining features on
model performance in the context of different cases.
6.3.4. Lookback timeframe for model training
The training/validation set is examined in reference to various
lookback timeframes as part of the ML model development process. The
goal here is to investigate the effect of the number of days used to train
the model on the model performance and the predictability of cycle
time. On average, using a lookback timeframe of 10 to 20 days for
training and validation is found to result in better performance measures
compared to shorter or longer lookback timeframes. Fig. 9 illustrates the
manner in which the splitting of the data is carried out. It is worth noting
that further investigation is required to examine the full effect of varying
lookback timeframes on different days of the week. This task will be
undertaken as part of future work.
6.4. Comparison with Little’s law
The simplest and most widely used analytical method to predict cycle
time is Little’s law, which can be applied to dynamic systems by dividing
the time into finite intervals. Equation (13) presents Little’s law, where
any of the average CT, system throughput (TH), and average units in the
system (WIP or work-in-progress) can be calculated by making obser­
vations about the other two quantities over the given time interval [15]:
WIP = CT × TH (13)
To compare these quantities, the time interval in the present study is
set to one work shift. The average daily production, shown in Fig. 6, is
found to be 39.3 panels/day. Given a 7.5-hour work shift, this translates
into a system throughput of about 5 panels/hour. When the observed
panel enters the production route, the average number of WIP panels is
3.5 panels. Using Little’s law, the cycle time is then calculated as 42 min,
close to the calculated average of 40.9 min, as shown in Fig. 7.
Assuming that this average number is used as the cycle time for all
panels, the performance measures used to compare ML models are
calculated for the given dataset: MAE = 11 min, MedAE = 8 min, and
MAPE = 30%. These values are inferior to those obtained from any data-
driven ML models. Although this analytical method can be used to
capture the average performance of the system by estimating CT based
on the average daily production, error rates are high, and the variability
in the production can not be efficiently captured as in the case of
applying ML models to predict the CT.
7. Conclusions
This study focuses on the accurate prediction of cycle time in a
O. Mohsen et al.
panelized wall manufacturing shop using the physical properties of the
panels and real-time loading conditions of the fabrication shop con­
structed from raw RFID timestamps. Real-time actual data, although it
poses several challenges compared to the simulated data used in most
previous studies in this domain, as it has a considerable amount of noise,
offers the advantage of depicting the actual production system more
realistically than does simulated data. Therefore, extensive data iden­
tification, preprocessing, and cleaning are carried out to obtain the tidy
dataset used for model training, validation, and testing. A methodology
that draws upon the data mining frameworks commonly used in the
literature is proposed and systematically applied to develop several ML
models for CT prediction. The approach uses real-time RFID data to
construct new features depicting the real-time loading conditions of the
job shop. It seeks to map the relationship between features representing
the current state of shop utilization, the product properties, and the CT
based on historical data. The results show that all the developed ML
models (LR, KNN, RF, and NN) yield predictions of production CT that
are more accurate than the predictions of average CT as calculated using
Little’s law. In addition, all the developed models showed improved
prediction accuracy when engineering features that represent shop uti­
lization were added to the training dataset. The random forests model,
which uses an ensemble of regression decision trees, provides a flexible
method and is found to achieve the best performance among the various
techniques considered, with a MAPE of 11% and a CC of 0.80.
The present study demonstrates the applicability and accuracy of ML
for predicting production CT in IBC using real-time shop utilization data.
This approach can help production managers make sound decisions
regarding dynamic scheduling of orders and shop capacity planning. We
show how this improved accuracy of CT prediction is largely attribut­
able to the augmentation of the dataset with engineered features that
capture shop utilization. This is evident in the improved performance (i.
e., lower error measures and higher CC) of all ML models tested in this
study compared to using only the physical properties of products.
Moreover, the results show that all models perform slightly better when
the granularity of the dataset is reduced by combining multiple physical
attributes into unified features (e.g., panel components = studs + win­
dows + doors + sheathing + backing), which is found to have a higher
correlation with the cycle time. We also show that a lookback timeframe
of 10–20 days results in more accurate predictions compared to using all
available historical data. However, further study is required to examine
the effect of varying lookback timeframes on the robustness and accu­
racy of the ML models.
The present research study uses historical data obtained from an
RFID data acquisition system. This data contains considerable noise and
requires extensive data cleaning and preparation. In addition, due to the
existing RFID system setup, some data is not collected that might contain
valuable information, especially the features associated with waiting
times for each panel throughout the production process. Therefore,
future research directions include developing an integrated data mining
and simulation modeling that can generate the missing data regarding
waiting and idle times of panels while in production.
Moreover, future research is to be directed towards investigating the
application of ML and data mining techniques in other similar indus­
trialized building construction settings and how accurate prediction of
CT impacts the scheduling of on-site activities. The goal is to examine
the applicability and consistency of using ML models to predict CT in
mass customization production as in the IBC industry. It is worth noting
that the dataset used in the present study represents one fabrication
facility, and future work should also be directed towards confirming the
results obtained through this study by repeating the work for other
cases. The details of the engineered feature development reflecting the
loading conditions of the job shop will depend on the specific case study,
and some custom development may be required. However, due to the
similarities among IBC manufacturing environments, with some minor
customization, the presented features can give a representative picture
of the current state of shop utilization in a given IBC factory and thereby
11
Advanced Engineering Informatics 52 (2022) 101631
improve the predictability of CT.
Declaration of Competing Interest
The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influence
the work reported in this paper.
Acknowledgments
This research was funded through a Collaborative Research Devel­
opment Grant by the Natural Science and Engineering Research Council
of Canada (Grant File No. CRDPJ 536466-18).
References
[1] Z. Li, G.Q. Shen, X. Xue, Critical review of the research on the management of
prefabricated construction, Habitat Int. 43 (2014) 240–249, https://doi.org/
10.1016/j.habitatint.2014.04.001.
[2] USG Corporation and U.S. Chamber of Commerce, Commercial Construction Index
Q1 2018, 2018. https://www.uschamber.com/report/usg-us-chamber-commerce
-commercial-construction-index-2018-q1 (accessed April 2, 2021).
[3] N. Bertram, S. Fuchs, J. Mischke, R. Palter, G. Strube, J. Woetzel, Modular
construction: From projects to products, 2019. https://www.mckinsey.com/bus
iness-functions/operations/our-insights/modular-construction-from-projects-to
-products (accessed April 2, 2021).
[4] Modular Building Institute, 2020 Canadian Commercial Modular Construction
Annual Report, 2020. https://www.modular.org/HtmlPage.aspx?name=2020
-MBI-annual-reports (accessed April 2, 2021).
[5] Z. You, C. Wu, A framework for data-driven informatization of the construction
company, Adv. Eng. Inform. 39 (2019) 269–277, https://doi.org/10.1016/j.
aei.2019.02.002.
[6] H. Lasi, P. Fettke, H.-G. Kemper, T. Feld, M. Hoffmann, Industry 4.0, Bus. Inf. Syst.
Eng. 6 (2014) 239–242, https://doi.org/10.1007/s12599-014-0334-4.
[7] M. Hermann, T. Pentek, B. Otto, Design Principles for Industrie 4.0 Scenarios, in:
49th Hawaii Int. Conf. Syst. Sci., Koloa, HI, 2016, pp. 3928–3937. https://doi.org/
10.1109/HICSS.2016.488.
[8] K. Schwab, The Fourth Industrial Revolution, Crown Business, New York, 2017.
[9] A. Aapaoja, H. Haapasalo, The Challenges of Standardization of Products and
Processes in Construction, in: 22nd Annu. Conf. Int. Gr. Lean Constr., Oslo,
Norway, 2014, pp. 983–993. https://doi.org/10.13140/2.1.3993.7600.
[10] M.R. Hosseini, I. Martek, E.K. Zavadskas, A.A. Aibinu, M. Arashpour, N. Chileshe,
Critical evaluation of offsite construction research: A Scientometric analysis,
Autom. Constr. 87 (2018) 235–247, https://doi.org/10.1016/j.
autcon.2017.12.002.
[11] R. Jin, S. Gao, A. Cheshmehzangi, E. Aboagye-Nimo, A holistic review of offsite
construction literature published between 2008 and 2018, J. Clean. Prod. 202
(2018) 1202–1219, https://doi.org/10.1016/j.jclepro.2018.08.195.
[12] M.A. Mullens, Factory Design for Modular Homebuilding: Equipping the Modular
Factory for Success, Constructability Press, Winter Park, FL, 2011.
[13] M.E. Pfund, S.J. Mason, J.W. Fowler, Semiconductor Manufacturing Scheduling
and Dispatching, in: Handb. Prod. Sched., Kluwer Academic Publishers, Boston,
2006, pp. 213–241. https://doi.org/10.1007/0-387-33117-4_9.
[14] S.-H. Chung, H.-W. Huang, Cycle time estimation for wafer fab with engineering
lots, IIE Trans. 34 (2002) 105–118, https://doi.org/10.1080/
07408170208928854.
[15] W.J. Hopp, M.L. Spearman, Basic Factory Dynamics, in: Fact. Phys. Found. Manuf.
Manag., 2nd ed., McGraw-Hill/Irwin, Boston, MA, 2000, pp. 213–247.
[16] U. Fayyad, G. Piatetsky-Shapiro, P. Smyth, From Data Mining to Knowledge
Discovery in Databases, AI Mag. 17 (1996) 37–54, https://doi.org/10.1609/aimag.
v17i3.1230.
[17] P. Cabena, R. Stadler, J. Verhees, A. Zanasi, P. Hadjnian, Discovering data mining
from concept to implementation, Prentice Hall, New Jersey, 1998.
[18] L. Soibelman, H. Kim, Data Preparation Process for Construction Knowledge
Generation through Knowledge Discovery in Databases, J. Comput. Civ. Eng. 16
(2002) 39–48, https://doi.org/10.1061/(ASCE)0887-3801(2002)16:1(39).
[19] H. Kim, L. Soibelman, F. Grobler, Factor selection for delay analysis using
Knowledge Discovery in Databases, Autom. Constr. 17 (2008) 550–560, https://
doi.org/10.1016/j.autcon.2007.10.001.
[20] A.M. Hammad, An Integrated Framework for Managing Labour Resources Data in
Industrial Construction Projects: A Knowledge Discovery in Data (KDD) Approach,
PhD, University of Alberta, 2009 https://www.collectionscanada.gc.ca/obj/
thesescanada/vol2/002/NR55814.PDF (accessed April 2, 2021).
[21] M. Wauters, M. Vanhoucke, Support Vector Machine Regression for project control
forecasting, Autom. Constr. 47 (2014) 92–106, https://doi.org/10.1016/j.
autcon.2014.07.014.
[22] M. Bilal, L.O. Oyedele, J. Qadir, K. Munir, S.O. Ajayi, O.O. Akinade, H.A. Owolabi,
H.A. Alaka, M. Pasha, Big Data in the Construction Industry: A review of Present
Status, Opportunities, and Future Trends, Adv. Eng. Informatics. 30 (2016)
500–521, https://doi.org/10.1016/j.aei.2016.07.001.
O. Mohsen et al.
[23] Q. Fang, H. Li, X. Luo, L. Ding, T.M. Rose, W. An, Y. Yu, A deep learning-based
method for detecting non-certified work on construction sites, Adv. Eng.
Informatics. 35 (2018) 56–68, https://doi.org/10.1016/j.aei.2018.01.001.
[24] J. Choi, B. Gu, S. Chin, J.-S. Lee, Machine learning predictive model based on
national data for fatal accidents of construction workers, Autom. Constr. 110
(2020) 1–14, https://doi.org/10.1016/j.autcon.2019.102974.
[25] N.D. Nath, A.H. Behzadan, S.G. Paal, Deep learning for site safety: Real-time
detection of personal protective equipment, Autom. Constr. 112 (2020) 1–20,
https://doi.org/10.1016/j.autcon.2020.103085.
[26] J. Wu, N. Cai, W. Chen, H. Wang, G. Wang, Automatic detection of hardhats worn
by construction personnel: A deep learning approach and benchmark dataset,
Autom. Constr. 106 (2019) 1–7, https://doi.org/10.1016/j.autcon.2019.102894.
[27] M. Bilal, L.O. Oyedele, Guidelines for applied machine learning in construction
industry—A case of profit margins estimation, Adv. Eng. Informatics. 43 (2020),
101013, https://doi.org/10.1016/j.aei.2019.101013.
[28] E. Valero, A. Adán, C. Cerrada, Evolution of RFID Applications in Construction: A
Literature Review, Sensors. 15 (2015) 15988–16008, https://doi.org/10.3390/
s150715988.
[29] E. Valero, A. Adán, Integration of RFID with other technologies in construction,
Measurement 94 (2016) 614–620, https://doi.org/10.1016/j.
measurement.2016.08.037.
[30] J.A. Harding, M. Shahbaz, A.K. Srinivas, Data Mining in Manufacturing: A Review,
J. Manuf. Sci. Eng. 128 (2006) 969–976, https://doi.org/10.1115/1.2194554.
[31] A.K. Choudhary, J.A. Harding, M.K. Tiwari, Data mining in manufacturing: a
review based on the kind of knowledge, J. Intell. Manuf. 20 (2009) 501–521,
https://doi.org/10.1007/s10845-008-0145-x.
[32] A. Öztürk, S. Kayalıgil, N.E. Özdemirel, Manufacturing lead time estimation using
data mining, Eur. J. Oper. Res. 173 (2006) 683–700, https://doi.org/10.1016/j.
ejor.2005.03.015.
[33] P. Backus, M. Janakiram, S. Mowzoon, G.C. Runger, A. Bhargava, Factory Cycle-
Time Prediction With a Data-Mining Approach, IEEE Trans. Semicond. Manuf. 19
(2006) 252–258, https://doi.org/10.1109/TSM.2006.873400.
[34] T. Chen, R. Romanowski, Precise and Accurate Job Cycle Time Forecasting in a
Wafer Fabrication Factory with a Fuzzy Data Mining Approach, Math. Probl. Eng.
2013 (2013) 1–14, https://doi.org/10.1155/2013/496826.
[35] B. Can, C. Heavey, A Demonstration of Machine Learning for Explicit Functions for
Cycle Time Prediction Using MES Data, in: 2016 Winter Simul. Conf., Washington,
DC, 2016, pp. 2500–2511. https://doi.org/10.1109/WSC.2016.7822289.
[36] D. Gyulai, A. Pfeiffer, G. Nick, V. Gallina, W. Sihn, L. Monostori, Lead time
prediction in a flow-shop environment with analytical and machine learning
approaches, IFAC-PapersOnLine. 51 (2018) 1029–1034, https://doi.org/10.1016/
j.ifacol.2018.08.472.
[37] L. Lingitz, V. Gallina, F. Ansari, D. Gyulai, A. Pfeiffer, W. Sihn, L. Monostori, Lead
time prediction using machine learning algorithms: A case study by a
semiconductor manufacturer, Proc. CIRP 72 (2018) 1051–1056, https://doi.org/
10.1016/j.procir.2018.03.148.
[38] P. Chapman, J. Clinton, R. Kerber, T. Khabaza, T. Reinartz, C. Shearer, R. Wirth,
CRISP-DM 1.0 Step-by-Step Data Mining Guide, 2000. https://the-modeling-agen
cy.com/crisp-dm.pdf (accessed April 2, 2021).
[39] P. Heller, D. Piziak, R. Stackowiak, A. Licht, T. Luckenbach, B. Cauthen, A. Misra,
J. Wyant, J. Knudsen, An Enterprise Architect’s Guide to Big Data — Reference
Architecture Overview, 2016. http://www.oracle.com/technetwork/topics/ent
arch/articles/oea-big-data-guide-1522052.pdf (accessed April 2, 2021).
[40] T.H. Davenport, J.G. Harris, R. Morison, Analytics at Work: Smarter Decisions,
Better Results, Harvard Business Press, Boston, Massachusetts, 2010.
[41] D. Dietrich, B. Heller, Y. Beibei, Data Science & Big Data Analytics: Discovering,
Analyzing, Visualizing and Presenting Data, John Wiley & Sons, Indianapolis, IN,
2015.
[42] J.B. Rollins, Foundational Methodology for Data Science, IBM Anal. (2015).
https://tdwi.org/~/media/64511A895D86457E964174EDC5C4C7B1.PDF
(accessed April 2, 2021).
12
Advanced Engineering Informatics 52 (2022) 101631
[43] H. Wickham, Tidy Data, J. Stat. Softw. 59 (2014) 1–23. https://doi.org/10.18637/
jss.v059.i10.
[44] S. Chakrabarti, E. Cox, E. Frank, R.H. Güting, J. Han, X. Jiang, M. Kamber, S.
S. Lightstone, T.P. Nadeau, R.E. Neapolitan, D. Pyle, M. Refaat, M. Schneider, T.
J. Teorey, I.H. Witten, Data Mining Know It All, Morgan Kaufmann, Burlington,
MA, 2009.
[45] Y. Li, C. Yang, Domain knowledge based explainable feature construction method
and its application in ironmaking process, Eng. Appl. Artif. Intell. 100 (2021),
104197, https://doi.org/10.1016/j.engappai.2021.104197.
[46] T. Hastie, R. Tibshirani, J. Friedman, The Elements of Statistical Learning, 2nd ed.,
Springer New York, New York, NY, 2009. https://doi.org/10.1007/978-0-387-
84858-7.
[47] I.H. Witten, E. Frank, M.A. Hall, Data Mining: Practical Machine Learning Tools
and Techniques, 3rd ed., Morgan Kaufmann, Burlington, MA, 2011. https://doi.
org/10.1007/s00170-004-2497-5.
[48] S. Raschka, V. Mirjalili, Python Machine Learning - Second Edition, 2nd ed., Packt
Publishing, Birmingham, 2017.
[49] L. Breiman, J.H. Friedman, R.A. Olshen, C.J. Stone, Classification And Regression
Trees, 1st ed., Routledge, 1984. https://doi.org/10.1201/9781315139470.
[50] L. Breiman, Random forests, Mach. Learn. 45 (2001) 5–32, https://doi.org/
10.1023/A:1010933404324.
[51] A.H. Al-Momani, Construction delay: a quantitative analysis, Int. J. Proj. Manag.
18 (2000) 51–59, https://doi.org/10.1016/S0263-7863(98)00060-X.
[52] D.J. Lowe, M.W. Emsley, A. Harding, Predicting Construction Cost Using Multiple
Regression Techniques, J. Constr. Eng. Manag. 132 (2006) 750–758, https://doi.
org/10.1061/(ASCE)0733-9364(2006)132:7(750).
[53] T.M. Zayed, D.W. Halpin, Productivity and Cost Regression Models for Pile
Construction, J. Constr. Eng. Manag. 131 (2005) 779–789, https://doi.org/
10.1061/(ASCE)0733-9364(2005)131:7(779).
[54] J. Wang, B. Ashuri, Predicting ENR’S Construction Cost Index Using the Modified K
Nearest Neighbors (KNN) Algorithm, in: Constr. Res. Congr. 2016, American
Society of Civil Engineers, Reston, VA, 2016, pp. 2502–2509. https://doi.org/
10.1061/9780784479827.249.
[55] Y. Zhang, L. Ding, P.E.D. Love, Planning of Deep Foundation Construction
Technical Specifications Using Improved Case-Based Reasoning with Weighted k-
Nearest Neighbors, J. Comput. Civ. Eng. 31 (2017) 04017029, https://doi.org/
10.1061/(ASCE)CP.1943-5487.0000682.
[56] J. Zhou, X. Shi, K. Du, X. Qiu, X. Li, H.S. Mitri, Development of Ground Movements
Due to a Shield Tunnelling Prediction Model Using Random Forests, in: Geo-China
2016, American Society of Civil Engineers, Reston, VA, 2016, pp. 108–115.
https://doi.org/10.1061/9780784480106.014.
[57] Y. Hu, D. Castro-Lacouture, Clash Relevance Prediction Based on Machine
Learning, J. Comput. Civ. Eng. 33 (2019) 1–15, https://doi.org/10.1061/(ASCE)
CP.1943-5487.0000810.
[58] G. Heravi, E. Eslamdoost, Applying Artificial Neural Networks for Measuring and
Predicting Construction-Labor Productivity, J. Constr. Eng. Manag. 141 (2015)
04015032, https://doi.org/10.1061/(asce)co.1943-7862.0001006.
[59] M.-Y. Cheng, Y.-H. Chang, D. Korir, Novel Approach to Estimating Schedule to
Completion in Construction Projects Using Sequence and Nonsequence Learning,
J. Constr. Eng. Manag. 145 (2019) 04019072, https://doi.org/10.1061/(asce)
co.1943-7862.0001697.
[60] G. Varoquaux, L. Buitinck, G. Louppe, O. Grisel, F. Pedregosa, A. Mueller, Scikit-
learn: Machine Learning in Python, J. Mach. Learn. Res. 12 (2011) 2825–2830. htt
p://jmlr.csail.mit.edu/papers/volume12/pedregosa11a/pedregosa11a.pdf.
accessed April 2, 2021.
[61] A. Botchkarev, A New Typology Design of Performance Metrics to Measure Errors
in Machine Learning Regression Algorithms, Interdiscip. J. Information,
Knowledge, Manag. 14 (2019) 045–076. https://doi.org/10.28945/4184.
[62] J. Bergstra, Y. Bengio, Random Search for Hyper-Parameter Optimization, J. Mach.
Learn. Res. 13 (2012) 281–305. https://jmlr.csail.mit.edu/papers/volume13/be
rgstra12a/bergstra12a.pdf. accessed April 2, 2021.