You are on page 1of 11

Application of Low-Cost Accelerometers for Measuring the

Operational Efficiency of a Construction Equipment Fleet


Changbum R. Ahn, M.ASCE 1; SangHyun Lee, M.ASCE 2; and Feniosky Peña-Mora, M.ASCE 3

Abstract: Monitoring the operational efficiency of construction equipment offers great opportunities to enhance not only the productivity
but also the environmental performance of construction operations. However, existing enabling technologies still suffer from a lack of eco-
nomic feasibility, as well as technological compatibility with equipment fleets that are outdated or that consist of diverse manufacturers’
Downloaded from ascelibrary.org by Tufts University on 07/26/18. Copyright ASCE. For personal use only; all rights reserved.

models. In this context, this paper examines the feasibility of measuring the operational efficiency of equipment using low-cost acceler-
ometers. Acceleration data in three axes were collected from a real-world operation of excavators that performed various duty cycles. Multiple
features were calculated from acceleration data, and several classifiers using these features were tested to classify equipment operation into
engine-off, idling, and working modes. An accuracy of over 93% was obtained in the classification of excavators’ operation. This result has
demonstrated that the application of low-cost accelerometers has the potential to provide a robust system to automatically measure the
operational efficiency of construction equipment without any connection to its legacy engine system. DOI: 10.1061/(ASCE)CP.1943-
5487.0000337. © 2014 American Society of Civil Engineers.
Author keywords: Sustainable construction; Construction management; Resource tracking; Accelerometer; Operational efficiency;
Equipment idling.

Introduction (NOx), hydrocarbons (HC), particulate matter (PM), and black


carbon—cause immediate and adverse effects to public health
Construction projects have suffered from low productivity due to a and the ecosystem (EPA CAAAC 2006). Reducing exhaust emis-
variety of unexpected incidents, such as interruption and waiting, sions from construction equipment use can be achieved not only by
that are caused by dynamic outdoor working environments (Sakcs using cleaner equipment and fuel, but also by improving opera-
et al. 2010). As construction equipment is one of the main compo- tional efficiency, which means the reduction of engine idling (that
nents in a construction production system, the systematic measure- generates emissions without any production) (Ahn and Lee 2013).
ment and analysis of a construction equipment operation is The operational efficiency of construction equipment, which re-
essential for productivity improvement (Gong and Caldas 2011). fers to the ratio of the productive work time of equipment to its total
Accurate data of the process measurement are necessary not only operating time (Nichols and Day 2005), is a piece of key informa-
to control current projects, but also to update historical databases tion in the analysis of a construction equipment operation. It is
(Sardroud et al. 2010). Such updates allow for better planning of generally used in accounting for nonproductive equipment time,
future projects in terms of resource allocation, schedule, and cost. including the operator’s need for personal breaks and communica-
In addition, the analysis of a construction equipment operation tion with other personnel, as well as the side effect from interacting
is important in reducing the environmental impact generated from with equipment and crews (e.g., off-road trucks’ waiting in a queue
a construction project, since the use of construction equipment to be loaded by an excavator) (Lewis et al. 2011). In the practice
generates significant air pollutant emissions, including carbon of estimating construction equipment productivity and cost, 75%
emissions (CO2 ) and diesel exhaust emissions (EPA 2008). In par- (45 min=h) and 83% (50 min =h) of operational efficiencies are
ticular, diesel exhaust emissions generated from heavy-duty diesel commonly used to account mainly for the operators’ breaks, during
equipment—such as carbon monoxide (CO), nitrogen oxides which equipment is usually turned off (Peurifoy et al. 2011). On the
other hand, there exist little data on the average equipment idling
1 times during which equipment is running but not engaged in any
Assistant Professor, Construction Engineering and Management
Division, Charles Durham School of Architectural Engineering and
productive work (EPA 2009), although this is a crucial piece of
Construction, Univ. of Nebraska-Lincoln, 113 Nebraska Hall, Lincoln, information for productivity measurement and emission reduction.
NE 68588 (corresponding author). E-mail: cahn2@unl.edu The idling rate of construction equipment is highly dependent
2
Assistant Professor, Dept. of Civil and Environmental Engineering, on various characteristics (e.g., performed task type, equipment
Univ. of Michigan, 2350 Hayward St., Room 2340 G.G. Brown Building, fleet configuration, job site conditions) of an operation involving
Ann Arbor, MI 48109. E-mail: shdpm@umich.edu equipment. Generally, the idling rate of equipment in typical con-
3
Edwin Howard Armstrong Professor of Civil Engineering and struction operations is known to be quite higher than that of on-road
Engineering Mechanics and Professor of Earth and Environmental vehicles and off-road equipment use in other industries that are
Engineering and of Computer Science, Columbia Univ., City of New York, controlled under strict anti-idling laws (EPA 2006). For example,
510 Southwest Mudd Building, 500 West 120th St., New York, NY 10027.
Komatsu excavators in Colorado and Wyoming were reported to be
E-mail: feniosky@columbia.edu
Note. This manuscript was submitted on January 2, 2013; approved on idling approximately 35% of the time (Hagerty 2011). In a project
June 13, 2013; published online on June 15, 2013. Discussion period open located in an urban area, the idling rate of equipment is much
until September 30, 2014; separate discussions must be submitted for in- higher (up to 60 ∼ 70%) due to both the density of the job site
dividual papers. This paper is part of the Journal of Computing in Civil and traffic delays (Ahn and Lee 2013). In this context, monitoring
Engineering, © ASCE, ISSN 0887-3801/04014042(11)/$25.00. operational efficiency, which includes the measurement of the

© ASCE 04014042-1 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(2): 04014042


idling rate, offers great opportunities to enhance the productivity operating equipment efficiency (OEE) as the ratio of valuable
and environmental performance of construction operations. How- (nonidle) operating time to total operating (equipment engine is
ever, these monitoring efforts are still quite primitive due to the lack on) time of equipment, and provided the way to quantify the
of a practical monitoring method. Several emerging technologies amount of emissions from construction operations based on
exist, such as vehicle health monitoring systems (VHMS) of the measured OEEs:
original equipment manufacturers (OEM) or supplied by third par- X
ties, which allow for the accurate monitoring of equipment use. Emissions ¼ A × EFvaluable × ½OEE þ ð1 − OEEÞ × ρ
Still, they have compatibility issues across different manufacturers, equipment
and their implementation in the aftermarket is still costly (Monnot ð2Þ
and Williams 2011; Moore 2012). Furthermore, in the majority of
old equipment, it is hardly implemented without extensive modi- where A = activity (h); OEE = operating equipment efficiency;
fication, since the implementation of VHMS requires a connection EFvaluable = emission factor for valuable operating time (working
to legacy engine systems (Sharif and Lee 2010). In this context, the modes) (g=h); and ρ = the generalized ratio of idle to nonidle emis-
application of low-cost accelerometers has the potential to address sion rate of construction equipment (Lewis et al. 2012).
the challenges of existing technologies by providing a low-cost
Downloaded from ascelibrary.org by Tufts University on 07/26/18. Copyright ASCE. For personal use only; all rights reserved.

While different working modes (e.g., digging, moving) of


and nonintrusive monitoring system of the equipment operation. equipment yield different levels of emission rates, the variability
To this end, this paper evaluates the feasibility of using accelerom- of emission rates due to different working modes is very small com-
eter data to measure the operational efficiency of construction pared to the difference of emission rates between valuable and non-
equipment. The paper begins with the review of existing works valuable (idling) operation (Abolhasani et al. 2008). Therefore,
on the relationship between operational efficiency and environmen- such a variability of emission rates between different working
tal performance, enabling technologies for monitoring environmen- modes can be ignored in the development of the emission inventory
tal performance and equipment operational efficiency, and the at the project level (Ahn and Lee 2013). Then, Eq. (2) indicates that
applications of accelerometers in construction. The following sec- the environmental performance monitoring of construction opera-
tions of the paper describe the details of the experiments that were tions can be achieved with the measurement of two key pieces of
carried out and the processes to analyze the data from the experi- data: usage hours and the OEE of each piece of equipment. How-
ments. The paper ends with a discussion of the results and a con- ever, these pieces of data are not tracked in current practices of
clusion of the work. construction operations. While daily reports on the use of equip-
ment provide information on the equipment utilized (Ahn et al.
2011), these reports do not provide information on the usage hours
Background and OEE of each piece of equipment. This lack of disaggregated
data also hinders the accurate monitoring of environmental
Importance of Operational Efficiency in Reducing Air performance.
Pollutant Emissions
Ahn and Lee (2013) formulated the relationship between pollutant Enabling Technologies for Monitoring the
productivity and operational efficiency, as follows: Environmental Performance and Operational Efficiency
of Construction Vehicles
Production
Pollutant productivity ¼ Several emerging technologies exist that allow a more accurate
Pollutant monitoring of the environmental performance (Ahn et al. 2011).
Production Energy Fuel Portable emission measurement systems (PEMS) provide a means
¼ × × ð1Þ
Energy Fuel Pollutant of direct measurement of air pollutants emitted through the tailpipe
of construction equipment. PEMS are, however, too costly (in ex-
In Eq. (1), Pollutant productivity, which is production per cess of $100,000 per unit) to be employed for the simultaneous
unit of pollutant (i.e., cubic meters of material per Kg CO2 ), rep- monitoring of a number of energy/emission sources in a project,
resents the environmental performance of construction operations. so their use is limited to the measurement of engine emissions
Among the determinants of Pollutant productivity, Energy/Fuel for research purposes. In addition, vehicle health monitoring
indicates the fuel efficiency of a resource utilized in operations, systems (VHMS) of construction equipment are available in re-
which is the amount of energy generated by a resource per unit cently manufactured models of construction equipment, and allow
of fuel (i.e., kWh per liter of fuel). Fuel/Pollutant is highly related for the conversion of the electronic control unit (ECU) data from
to the environmental property of fuel consumed in resources, since on-board diagnostics (OBD) protocols (e.g., OBD-II, EOBD,
Fuel/Pollutant is an inverse form of the fuel’s intrinsic emission JOBD, and CAN bus) of construction equipment into useful infor-
intensity, which is the amount of pollutant emitted per unit of fuel mation (e.g., usage, OEE, and fuel consumption). However, old
consumed (i.e., Kg CO2 per liter of fuel). These two variables are equipment—which is the majority of equipment in use—does
hardly changeable once the construction phase starts since the not have OBD supports; therefore, VHMS implementation in such
change of equipment and fuel type for improving the environmental equipment is hardly possible (Sharif and Lee 2010). Even a piece of
performance rarely occurs in the middle of the operation. There- equipment with OBD supports requires extensive modification or
fore, Production/Energy—which represents the operational effi- the installation of additional devices. In the case of using aftermar-
ciency of a process or resource (the amount of production per ket devices, their price is typically in excess of $1,000 per machine
unit of energy usage, for example, cubic meters of material per (the price of OEM devices is usually much higher), and there are
kWh)—is the most important variable in determining the environ- also additional expenses for developing and operating integration
mental performance during construction (Ahn and Lee 2013). software (Moore 2012). In addition, there is a compatibility issue
To this end, monitoring the operational efficiency of resources em- between different manufacturers due to a lack of standardized
ployed in operations provides an efficient means of monitoring the protocol, and this creates another challenge in performing compre-
environmental performance. Ahn and Lee (2013) also defined the hensive monitoring for contractors who own, rent, or purchase

© ASCE 04014042-2 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(2): 04014042


machines from diverse manufacturers (Monnot and Williams 2011; Research Objective and Methodology
Moore 2012; Azar and McCabe 2012).
In recent years, several research studies have focused on an au- The objective of this research is to test the hypothesis that signals
tomated tracking of construction equipment operation. El-Omari captured by MEMS accelerometers that are installed on construc-
and Moselhi (2009) and Ergen et al. (2007) utilized radio- tion equipment can be used to analyze the OEE of that equipment,
frequency identification (RFID) combined with global positioning which indicates the ratio of the valuable operating time (nonidling)
system (GPS) technology for the automated location and tracking to the total operating time of the equipment. The measurement
of construction equipment. However, the GPS-based approach is of OEEs can be formulated as a problem that classifies the
still incapable of tracking the stationary operation of construction second-by-second operation of equipment into three activity modes
equipment, while the combination of GPS and RFID provides a (i.e., working, idling, and engine-off) based on acceleration data.
high accuracy in locating resources. On the other hand, Azar The underlying idea of the hypothesis is twofold: first, any nonsta-
and McCabe (2012), Heydarian et al. (2012), Gong and Caldas tionary operating of construction equipment (e.g., driving) will
(2009), and Zou and Kim (2007) introduced vision-based tech- create a notable level of acceleration compared to background
noise; second, any stationary operating of construction equipment
niques that track construction equipment operation by using video
(e.g., controlling excavators’ booms) will generate distinguishable
Downloaded from ascelibrary.org by Tufts University on 07/26/18. Copyright ASCE. For personal use only; all rights reserved.

cameras. However, vision-based techniques provide only limited


patterns of acceleration signals compared to the idling and engine-
information based on the visual scope of cameras, and still have
off modes. The former idea has already been demonstrated by the
difficulties in providing high accuracies when there are high levels
application of accelerometers to detect passenger vehicle motion
of noise (e.g., entity overlaps, moving backgrounds, and varying
(MacDonald 1990; Baldwin et al. 2004), but the latter idea needs
light conditions), which are common in dynamic construction
to be tested due to limited previous studies on the vibration of con-
surroundings.
struction vehicles.
In this context, the initial experiment is conducted under an in-
Applications of Accelerometers in Construction structed environment in order to analyze patterns of accelerometer
data from the stationary operating of construction equipment. An
An accelerometer is an electromechanical device that measures exploratory data analysis of the experiment result is conducted to
proper acceleration. The recent advent of small-sized, low-cost mi- discover distinguishable patterns of accelerometer data captured
croelectromechanical (MEMS) accelerometers has resulted in the from the stationary operating. Multiple features are extracted from
utilization of accelerometers in various applications, including raw accelerometer data, and the statistical significance is investi-
the sensing of a smartphone’s movement and gesture changes. gated in differences between distributions of features extracted
Within the construction industry, MEMS accelerometers have been during three activity modes. Next, experiments to collect acceler-
used for the health monitoring of structures by sensing the struc- ometer data in real-world applications are conducted in order to
tures’ vibrations (Spencer et al. 2004). In addition, several studies evaluate the feasibility of the proposed approach in various real-
(Joshua and Varghese 2011; Cheng et al. 2013a, b) have used ac- world duty cycles. Data are collected from different excavators that
celerometer data for the activity analysis of construction workers. perform diverse stationary and nonstationary operations. The col-
Akhavian and Behzadan (2012) utilized a set of inertial measure- lected data are segmented into a fixed window size, and features are
ment sensors (magnetometer, gyroscope, and accelerometer) to de- extracted from segmented data based on the analysis of the experi-
tect the three-dimensional orientation of construction equipment. ment result. The extracted features are trained and tested with
Accelerometers have also been used for the control and condi- different classification algorithms that are found in the Waikato
tion monitoring (e.g., engine fault detection) of internal combustion Environment for Knowledge Analysis (WEKA) Machine Learning
(IC) engines in vehicles, including construction machines (Antoni Algorithms Toolkit (Hall et al. 2009). The performances of differ-
et al. 2002). The basic underlying idea of such applications is that ent classifiers are evaluated.
every moving component or physical process involved in the op- Acceleration signals from construction vehicles under the study
eration of an engine produces its own unique vibration signal, were measured using the accelerometer mounted inside the cabin
which is referred to as the vibration signature. It is assumed that of vehicles, and the operation of construction vehicles during
vibration signatures created by the same operating engine have the study was videotaped. The accelerometer used (LIS331DL,
the same features under the same conditions. For the conditioning STMicroelectronics 2013) was the one embedded in a smartphone,
monitoring of IC engines, expensive conventional accelerometers which can sense acceleration in the x-, y-, and z-directions. The
are generally used, since condition monitoring requires a high level sensitivity of the accelerometer was 16.2 mg ðmilli-gÞ=digit, and
of precision in sensing vibrations (Albartar et al. 2008). Compared its measurement range was 2 g. The raw data from the acceler-
to conditioning monitoring, measuring equipment operational effi- ometer is collected without any data processing (e.g., the removal
ciency by detecting the activity of equipment (e.g., engine-off, idle, of outliner) through the smartphone application developed for the
working) is assumed to be a simpler application that can be achiev- experiments. The signals acquired by the accelerometer are
able with MEMS accelerometers. This is because different activ- sampled at a rate of 100 Hz. The mounting location of the accel-
ities exhibit a notable difference in vibration signatures, while erometer varies by vehicle, but is generally mounted on a rigid
the condition monitoring of IC engines (e.g., engine fault) relies block around the control system within the cabin, as illustrated
on a more subtle difference. In this context, the paper evaluates in Fig. 1(a). Video recordings of vehicle operations are used to label
the feasibility of using MEMS accelerometers as a monitor of op- actual operational modes of second-by-second vehicle operation; a
eration efficiency. However, a challenge in using MEMS acceler- vehicle is determined to be idling if it does not show any physical
movement for more than 3 s while its engine is on.
ometers as a monitor of operational efficiency in this research is
that we aim to achieve the nonintrusive measurement of different
duty cycles (e.g., attaching an accelerometer on equipment exterior Experimental Analysis
or placing it inside the cabin), while the condition monitoring of IC
engines requires the installation of the accelerometer directly to the An experiment is designed and carried out to analyze signals from
engine block. an excavator’s stationary operating. The operation of the excavator

© ASCE 04014042-3 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(2): 04014042


Fig. 1. (a and b) Devices mounted inside an excavator cabin and axes of acceleration relative to smartphone; (c) excavator in experimental analysis
and axes of acceleration relative to equipment
Downloaded from ascelibrary.org by Tufts University on 07/26/18. Copyright ASCE. For personal use only; all rights reserved.

during the experiment is strictly instructed, such that the excavator different activity modes, since newer vehicle models adopt better
does not have any movement except the up-and-down movement of vibration control technologies for the operator comfort.
its boom. It is also instructed to be idle for several seconds before Fig. 2 shows the x-axis readings during the entire period of the
and after operating its boom, in order to capture the signals from experiment, and Fig. 3 illustrates the examples of three axes’ read-
idling. In addition, a recently manufactured excavator, as illustrated ings from 1-s time segments that represent engine-off, idling, and
in Fig. 1(c), is chosen to provide a more restricted experiment envi- stationary operating modes. It is observed that different activity
ronment that creates the least difference of vibration levels between modes create different levels of acceleration in three axes, and that

0.14
0.12

0.1

0.08
Acceleration (g)

0.06

0.04

0.02

-0.02

-0.04

-0.06
10 20 30 40 50 60 70 80
Time (s)

Fig. 2. x-axis acceleration signals in experimental analysis

0.2 0.2 0.2

X-axis
0 0 0
Y-axis

-0.2 -0.2 -0.2


Acceleration (mg)

-0.4 -0.4 -0.4

-0.6 -0.6 -0.6

-0.8 -0.8 -0.8

Z-axis
-1 -1 -1

20 40 60 80 100 20 40 60 80 100 20 40 60 80 100


Engine-Off Idling Working

Fig. 3. Acceleration signals from three axes for engine-off, idling, and stationary operating modes

© ASCE 04014042-4 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(2): 04014042


the signal patterns in the three axes are quite similar in terms of the background noise (unwanted signal). The result indicates
of increasing and decreasing trends. The raw data are divided into that the distribution of average resultant acceleration, standard
1-s segments, and then features are generated based on the 100 deviation, and peak values are highly correlated with activity
readings contained within each 1-s segment. Any overlapping modes of time frames, while mean and correlation features do
between data frames is not considered at this point. not provide a strong discrimination between different activity
A total of 15 time-domain features are generated from each data modes. In the distribution of average resultant acceleration, stan-
frame. All of the features except the mean features are computed dard deviation, and peak values, the engine-off time frames have
from the detrended data (removing the mean of the signal). The lower values than other activity mode time frames, while the work-
features extracted are the following: ing time frames tend to have higher values compared to others.
1. Average resultant acceleration: This feature is evaluated by In particular, the engine-off mode time frames are clearly distin-
averaging the resultant acceleration values that are determined guishable in the distribution of those features, since any overlap
by the sum of the root mean square acceleration values of the with other groups is not found except one data point, which is as-
three axes. This feature also represents a short-term energy of sumed to be a result of external noise (e.g., unintended knock on the
signals. sensor). These features also provide a good basis for classifying
idling and working data frames, although some overlaps between
Downloaded from ascelibrary.org by Tufts University on 07/26/18. Copyright ASCE. For personal use only; all rights reserved.

2. Mean: This feature is evaluated by averaging the raw samples


over the window for each axis, in order to represent the DC the idling group and the work group are found. The features that are
component (the mean value of the waveform) of the signal. A extracted from y- and z-axis data are found to have similar patterns
total of three mean features are generated from three axes. as the feature extracted from x-axis data. For example, Fig. 5 illus-
3. Standard deviation: This feature represents the standard devia- trates the values of standard deviation and peak features in the x-,
tion value of accelerations. A total of three standard deviation y-, and z-axes. While standard deviation and peak values in the
features are generated from three axes. y- and z-axes have different scales than x-axis features, they still
4. Peak: This feature represents the maximum absolute value of provide a distinguishable power between the groups of data frames.
accelerations in each data frame. A total of three peak features The statistical significance of difference in distributions of average
are generated from three axes. resultant acceleration, standard deviation (three axes), and peak
5. Correlation: This feature measures the correlation of accelera- (three axes) features between three groups is tested with the analy-
tion between each pair of axes. A total of three correlation fea- sis of variance (ANOVA), and is confirmed.
tures are generated. The result of this experiment indicates two things. First, any
Fig. 4 illustrates a scatterplot that depicts the values of several engine operation of construction equipment, including idling,
features: average resultant acceleration, x-axis mean (AvgX), x-axis creates distinguishable patterns of acceleration signals compared
standard deviation (StdX), x-axis peak (PeakX), and correlation to the background noise that is generated during the engine-off
between x- and y-axes (CorrXY). The data points that indicate mode. Second, the stationary operating of an excavator—even
the time frames during the engine-off mode represent the features under a highly restricted environment—creates distinguishable

Fig. 4. Scatterplot of average resultant acceleration, x-axis mean (AvgX), x-axis standard deviation (StdX), x-axis peak (PeakX), and correlation
between x- and y-axes (CorrXY)

© ASCE 04014042-5 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(2): 04014042


Downloaded from ascelibrary.org by Tufts University on 07/26/18. Copyright ASCE. For personal use only; all rights reserved.

Fig. 5. Scatterplot of standard deviation and peak values in x-, y-, and z-axes

patterns of acceleration signals, as compared to the idling of an required very tumultuous actions of the excavator, placing wooden
excavator. trench boxes involved relatively modest actions. Experiment 2 col-
lected data from a large-sized excavator that bored holes for dew-
atering using a vibratory pile driver. The performed work repeated a
Real-World Operation Experiments working cycle that consisted of moving, locating the pile driver,
driving the pile, and pulling out the pile. While driving the pile
This section describes the experiments conducted to classify a real- generated an excessive level of vibration, moving and locating
world operation of excavators using accelerometer data, in order to the pile driver involved very modest levels of vibration. Due to
measure their operational efficiency. Excavators’ operations ana- the excessive vibration during driving the pile, the additional
lyzed in this section include various types of stationary and non- mounting device [Fig. 1(b)] was used to install the smartphone
stationary operation of excavators. to the side window inside the cabin. A wheeled excavator was
chosen for Experiments 3 and 4, in order to identify the effect
of different undercarriage types of construction vehicles. Two in-
Data Collection dependent data collections are made for this excavator to analyze
The acceleration data is collected from four individual experiments. the effect of different duty cycles that the excavator performs. In
The details of excavators and their duty cycle performed during the Experiment 3, the excavator mainly cleared and moved waste,
data collection are presented in Table 1. Experiment 1 collected and processed debris with another skid-steer loader. In Experiment
data from a medium-sized excavator that performed utility work 4, the excavator mainly demolished existing ground pavement, with
(placing wooden trench boxes and backfilling). While backfilling its stabilizers down. In each experiment, only idling and working

Table 1. Equipment Specifications and Their Performed Work in the Experiments


Experiment Equipment specifications Duration Observed OEE
number Model (HP, model year) Performed work (min) (%)
1 CAT 321C LCR Crawler type (148 hp, 2010) Install utility and backfill 29 80.2
2 CAT 330D Crawler type (270 hp, 2004) Drill dewatering holes 40 81.3
3 Volvo EW180C Wheeled type (180 hp, 2008) Clear debris 33 69.9
4 Volvo EW180C Wheeled type (180 hp, 2008) Demolish existing surface pavement 50 74.6

© ASCE 04014042-6 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(2): 04014042


modes were observed since the excavators were kept running and from acceleration signals in three axes. Spectral entropy is calcu-
were not turned off. Therefore, the key focus of the experiments lated as the normalized information entropy of the discrete
was to detect the idling operation from various duty cycles of ex- FFT component magnitudes of the signal. Spectral centroid is cal-
cavators that create different patterns of vibration. culated as the weighted mean of the frequencies present in the sig-
nal with their magnitudes as the weights. These frequency-domain
features may support the discrimination of time frames that have
Data Processing and Classification similar values of time domain features. In fact, the spectral
The accelerometer data are segmented into 128-sample windows centroid in the z-axis is found to provide a strong discriminative
with 64 samples overlapping between consecutive windows power. Fig. 6 illustrates a scatterplot of the normalized spectral
(50% overlapping). Feature extraction on sliding windows with centroid values in Experiment 1. While the z-axis spectral centroid
50% overlap demonstrated success in past studies (Bao and Intille of working time frames has values varying from 0.3 to 0.6, most
2004). The window size of 128 samples is chosen to enable the idling time frames have a value less than 0.01. This tendency is
speedy computation of the fast Fourier transform (FFT) used for also found in spectral centroid features extracted from other
some of the features. The change of the window size from 100 sam- experiments.
Downloaded from ascelibrary.org by Tufts University on 07/26/18. Copyright ASCE. For personal use only; all rights reserved.

ples (1 s), which was used in labeling the second-by-second oper- Classification of the extracted features is conducted using sev-
ation of excavators from video recordings, to 128 samples, requires eral supervised classifiers that are found in the WEKA Machine
merging and redefining the labels of segments. When merging and Learning Algorithms Toolkit (Hall et al. 2009). The following clas-
relabeling segments, a merged segment is labeled as idling when sifiers are used to classify activity modes:
more than 80% of samples in the segment were originally annotated 1. Naïve Bayes: Simple Bayesian networks that are composed of
as idling via a video recording. The merge process with the 80% directed acyclic graphs with only one parent (representing the
threshold rule for determining an idling label is found to create less unobserved node) and several children (corresponding to ob-
than a 1% difference in the overall idling duration compared to the served nodes) with a strong assumption of independence
original idling duration determined by video recording in each among child nodes.
experiment. 2. Instance-based learning (IBL): k-nearest neighbor (KNN)
In addition to average resultant acceleration, mean, standard classifier that is based on the principle that the instances within
deviation, peak, and correlation, two additional frequency-domain a data set will generally exist in close proximity to other in-
features—spectral entropy and spectral centroid—are extracted stances that have similar properties.

Fig. 6. Scatterplot of spectral centroid values in x-, y-, and z-axes

© ASCE 04014042-7 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(2): 04014042


3. Decision tree (J48): The implementation of the C4.5 decision
tree building algorithm that chooses one attribute of the data
that most effectively splits its set of samples into subsets en-
riched in one class or the other at each node of the tree, based
on information entropy.
4. Multilayer perceptron: A feedforward artificial neural network
model that maps sets of input data onto a set of appropriate
output with one or more layers between the input and output
layers.

Results and Analyses


The percentage accuracy of classifiers is evaluated through the
10-fold cross-validation. The statistical results of classifier perfor- Fig. 7. Classification accuracies with different window sizes
Downloaded from ascelibrary.org by Tufts University on 07/26/18. Copyright ASCE. For personal use only; all rights reserved.

mances are obtained from 10 runs of the 10-fold cross-validation.


Mean and standard deviation for classification accuracy in each
experiment is summarized in Table 2. performance of the Naïve Bayes classifier has greatly improved
Overall, the classification accuracies of IBL, J48, and multilayer with a bigger window size. With a window size of 1,024 samples,
perceptron are over 93% in any data set, while the accuracy of
the classification accuracy is found to be over 95% with any clas-
Naïve Bayes is quite lower than other classifiers. A paired t-test
sifier. In the results from the data sets of other experiments, the
result indicates that the classification performances of IBL, J48,
greatest accuracy is also found in the classification using a window
and Multilayer perceptron are not significantly different in most
size of 1,024 samples. The highest classification accuracies in
cases. In addition, the classification accuracies of IBL, J48, and
Experiments 1, 3, and 4 with a window size of 1,024 samples
Multilayer perceptron do not fluctuate greatly over different experi-
are 95.34% (with J48), 96.13% (with multilayer perceptron),
ments. This indicates that the classification accuracies based on ac-
and 98.67% (with Naïve Bayes), respectively.
celerometer signals are not greatly affected by an excavator’s
Another point of interest is the need for the training process
undercarriage type (track-type versus wheel-type), size, age, or
as an excavator performs different duty cycles. For example, in
duty cycle. This result also implies that the effect of the mounting
Experiments 3 and 4, data were collected from the same excavator,
location, orientation, and method of the accelerometer is not
but the excavator performed different duty cycles in the two
notable; even in Experiment 2 where the additional mounting de-
experiments. In order to analyze the effect of a change in duty
vice is used, any significant difference in classification accuracy
cycle, the classifiers are trained using the data set of Experiment
compared to other experiments is not found.
3, and their performances in Experiment 4 are evaluated. The clas-
In addition, the effect of the window size in the data segmenta-
sification performances in Experiment 3 are also evaluated in a
tion is analyzed. The data are segmented with different window
way that uses the data set of Experiment 4 as training data. The
sizes: 256, 512, and 1,024 samples. Different window sizes in
result is presented in Table 3. The overall accuracies in both cases
the data segmentation certainly affect the minimum idling duration
that can be recognized, since the data segmentation with different are lower than the original result in Table 2, but the greatest
window sizes involves merging and relabeling data segments, as accuracies of 95.1 and 86.6% obtained in Experiments 3 and 4
explained in the previous section. For example, with a window size are still significant. In particular, the performance of the Naïve
of 128 samples (1.2 s) and the 80% threshold rule for determining Bayes classifier remains steady in spite of using the training and
an idling label, the segment relabeled as idling in the merging pro- test data sets extracted from different experiments. It is noteworthy
cess has at least 102 samples (1 s) that were originally annotated as that the accelerometer was newly installed for each experiment, and
idling. Then, with a window size of 1,024 samples (10.2 s), the that the mounting location and orientation of the accelerometer
segment should have at least 819 samples (8 s) that were originally changed slightly over the experiments. Thus, a change of the accel-
annotated as idling in order to be relabeled as idling. This means erometer location and orientation may greatly contribute to the
that shorter than 8 s of a continuous pause of equipment motion in classification errors found in Table 2. In turn, this result indicates
this window size is highly likely to be labeled as working. The per- that a satisfactory accuracy can be obtained using pretrained clas-
formances of classifiers in different window sizes are evaluated sifiers in spite of a slight change in the accelerometer location and
using 10 times 10-fold cross-validation. Fig. 7 illustrates its result orientation.
in Experiment 2. All of the classifiers’ accuracies have increased In addition, the discriminating powers of features in each experi-
in a bigger window size, with one exception (IBL classifier from ment are evaluated using the correlation-based feature selection
512 sample window to 1,024 sample window). In particular, the

Table 3. Classification Accuracies Using Training and Test Data Sets


Extracted from Different Experiments
Table 2. Classification Accuracies Using Different Classifiers
% of instances correctly predicted
% of instances correctly predicted
(average  standard deviation) Training set: experiment 3 Training set: experiment 4
Experiment Multilayer Classifiers Test set: experiment 4 Test set: experiment 3
number Naïve Bayes IBL J48 perceptron
Naïve Bayes 95.1 86.6
1 71.71  2.89 94.60  1.47 94.78  1.29 94.14  1.75 IBL 94.3 81.2
2 89.92  1.44 93.37  1.37 94.80  1.08 93.56  1.01 J48 76.4 75.0
3 86.50  1.90 94.18  1.18 95.01  1.17 94.93  1.13 Multilayer 94.5 84.9
4 96.03  0.98 96.75  0.84 97.48  0.70 97.62  0.84 perceptron

© ASCE 04014042-8 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(2): 04014042


Table 4. Selected Features with High Discriminative Powers from idling to working modes or from working to idling modes.
Experiment This can be explained by the possible variation of engine submodes
number Selected features during excavators’ idling. During idling, excavators’ engines typ-
ically go through four submodes: low idle, transient between low
1 Peak (x- and z-axes), correlation (between x- and y-axes),
spectral centroid (x-axis) idle and high idle, high idle, and transient between high idle and
2 ARA, standard deviation (z-axis), peak (y-axis), spectral nonidle (Abolhasani et al. 2008). When the operator is ready to start
centroid (x- and y-axes) using the bucket, he/she increases the engine idle speed to a high-
3 Mean (x axis), standard deviation (x-, y-, and z-axes), idle mode, which is run at a higher engine revolutions per minute
peak (y-axis) (RPM) than a low-idle mode. Therefore, during short idling periods
4 Standard deviation (z-axis), peak (x-, y-, and z-axes), or transient periods, excavators may have run at high-idle or tran-
correlation (between x- and y-axes), spectral centroid sient mode between high and nonidle. On the other hand, during
(x- and z-axes) long idling periods, excavators may have run mostly at a low-idle
mode. From the practical point of view, a short pause (3 to 5 s) of
the equipment motion can barely be considered nonvaluable oper-
ation, since it often occurs between the change of motion (body
Downloaded from ascelibrary.org by Tufts University on 07/26/18. Copyright ASCE. For personal use only; all rights reserved.

(CFS) method (Hall et al. 2009). The CFS method finds a subset of
features that are highly correlated with the classification and with direction or initiating boom use) of equipment. When we define
other features, yet uncorrelated with one another. The Best First idling with a longer period of continuous motion pause, the clas-
sification accuracy greatly improves, by allowing the discard of er-
search approach (Hall et al. 2009) is used as the search approach
rors in short idling periods. The classification accuracy with the
of a subset of features. The features selected through CFS in each
practical definition of idling (longer period of continuous motion
experiment are presented in Table 4. Overall, Standard deviation,
pause) can be anticipated from the results of using a bigger window
Peak, and Spectral centroid are found to have strong discriminating
size in the data segmentation, since the window size is highly rel-
powers among features. This result agrees with the finding in the
evant to the minimum idling duration to be recognized. As shown
exploratory data analysis of the previous section.
in Fig. 7, when the minimum idling duration to be recognized in-
creased to around 8 s (1,024-sample window size), the classifica-
tion accuracies of all the classifiers improved to over 95%. In
Discussion addition, a bigger window size in the data segmentation helps alle-
viate the effect of external noises (e.g., unintended knock on the
The primary purpose of classifying accelerometer signals into dif-
sensor), since averaging feature values over the window smoothes
ferent activity modes of equipment is to automatically measure the
out a spike in the signals due to external noises.
OEEs of equipment. The acceptable range of the classification ac-
Monitoring the operational efficiency of equipment has drawn
curacy thus needs to be investigated in terms of the effects of the
the attention of contractors, as a means of locating ways to improve
classification accuracy on the accuracy in the OEE measurement.
their productivity. Recently, the reduction of equipment idling
Table 5 summarizes the possible deviations in OEE measurement
through monitoring OEE is more emphasized due to growing in-
when using the results of the high-performance classifiers (IBL,
terests in the environmental sustainability of construction projects
J48, and Multilayer perceptron) in each experiment. Since Type I
(EPA 2006). While a vehicle health monitoring system that allows
errors (false idling) and Type II errors (false working) in the clas-
for the accurate tracking of equipment operation spreads out along
sification offset each other in the OEE measurement, the accuracy
the distribution of recent models of construction equipment,
in the OEE measurement is usually higher than the classification
contractors still have a clear need to appropriate technology to
accuracy. Over 93% of the classification accuracy is found to create comprehensively track the OEE of their equipment fleet that is
less than a 2% deviation between the observed and measured OEEs outdated or that consists of various manufacturers’ models
in most cases, except for the Multilayer perceptron in Experiment (Azar and McCabe 2012). The presented approach provides a
2. This level of error rate in the OEE measurement is considered to measure to monitor the OEE of equipment with any age or manu-
be acceptable, taking into account that even the data segmentation facturer. The biggest advantage of the presented approach over
process (e.g., label merge due to a bigger window size) may cause other existing enablers is its relative low cost and effort to construct
up to a 1% difference in the observed OEE values. In particular, the a monitoring system for many pieces of equipment employed
IBL classifier is found to provide more accurate results (less than a in a large-scale project. The cost of a MEMS accelerometer
1% deviation) in the OEE measurement, while its classification ac- including data processing units is very low (around $50 per unit,
curacy is similar to that of J43 and Multilayer perceptron in STMicroelectronics 2013) compared to other existing enablers.
most cases. Also, the presented approach does not require any connection with
The classification errors are mostly found in the short idling an embedded system that controls equipment (i.e., it is nonintru-
periods that continue from 3 to 5 s, and in the transient periods sive) and its accuracy is quite robust in spite of the possible varia-
tion of the mounting location and method of the accelerometer.

Table 5. Measured OEEs Using Low-Performance (Naïve Bayes) and


High-Performance (IBL) Classifiers Conclusions
Measured OEE (deviation)
Automated measurement of the operational efficiency of con-
Experiment Observed Multilayer struction equipment provides key information in identifying
number OEE (%) IBL J48 perceptron
opportunities to improve the productivity and environmental
1 80.2 80.1% (−0.1) 81.5% (þ1.3) 80.4% (þ0.2) performance of construction operations. This paper presented a
2 81.3 81.7% (þ0.4) 82.9% (þ1.6) 83.5% (þ2.2) promising approach to measuring the operational efficiency of con-
3 69.9 69.4% (−0.5) 71.3% (þ1.4) 70.7% (þ0.8) struction equipment using low-cost accelerometers. The results in a
4 74.6 74.2% (−0.4) 75.4% (þ0.8) 75.4% (þ0.8)
real-world operation of excavators indicated that the presented

© ASCE 04014042-9 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(2): 04014042


approach can achieve an accuracy of over 93%; this level of accu- Multiclass Classification of Equipment Operation
racy is acceptable for practical applications since it created less than
While the operation of equipment has been classified mainly into
a 2% deviation in OEE measurements. It was also reported that the
the working and idling modes in this paper, the further analysis of
classification accuracy could be greatly enhanced with the practical
experiment data indicates a promising opportunity for the multi-
definition of equipment idling.
class classification of equipment operations. It has been found that
The approach to measuring the operational efficiency of con-
the different actions (e.g., digging, swinging loaded, dumping,
struction equipment using an accelerometer has clear advantages
moving forward/backward, rotating cabin) of an excavator create
over current enabling technologies in terms of technological com-
quite different patterns of acceleration signals. In particular, it is
patibility with equipment fleets that are outdated or that consist of
found that the classification of different actions during working
diverse manufacturers’ models, as well as the economic feasibility
mode can be greatly facilitated through the transformation of the
of implementation. The application of the proposed approach will
coordinates of the acceleration measurements from the device to
contribute not only to improving the environmental performance,
the desired coordinates that reflect the relevant axes of vehicle
but also to helping monitor and enhance the productivity of overall
manipulation (e.g., forward and sideward). Automated recognition
operations by detecting excessive nonvaluable operation.
of such equipment actions will provide an effective way to measure
Downloaded from ascelibrary.org by Tufts University on 07/26/18. Copyright ASCE. For personal use only; all rights reserved.

the cycle time of each piece of equipment for a detailed productiv-


ity analysis.
Potential Applications and Future Directions

The main focus of the presented approach is to automatically clas-


Acknowledgments
sify equipment operation into working, idling, and engine-off
modes using MEMS accelerometer data, in order to measure the The authors would like to acknowledge the Turner Construction
operational efficiency of equipment. While the feasibility of the Company, in particular James Barret (National Director, Integrated
presented approach has been tested mainly for excavators in this Building Solution), for their considerable help in collecting data.
paper, its feasibility for other types of equipment needs to be further
tested. Since the nonstationary operation of any type of construc-
tion equipment creates clearly distinguishable accelerometer data, References
its feasibility in construction equipment that mostly performs non-
stationary operation (e.g., off-road trucks, scrapers, dozers) is quite Abolhasani, S., Frey, H. C., Kim, K., Rasdorf, W., Lewis, P., and Pang,
promising. However, its feasibility for equipment that involves a S. H. (2008). “Real-world in-use activity, fuel use, and emissions for
great deal of stationary operation (e.g., cranes, loaders) needs to nonroad construction vehicles: A case study for excavators.” J. Air
be further investigated by future research. In addition, the preva- Waste Manage. Assoc., 58(8), 1033–1046.
lence of MEMS accelerometers in existing smartphones and the Ahn, C., Lee, S. H., and Peña-Mora, F. (2011). “Carbon emissions
quantification and verification strategies for large-scale construction
promising capability of the presented methodology in classifying
projects.” Proc., Int. Conf. on Sustainable Design and Construction
equipment operation imply a set of applications for the system. (ICSDC), ASCE, Reston, VA.
A list of the potential applications of the presented approach will Ahn, C. R., and Lee, S. (2013). “Importance of operational efficiency
be explained subsequently. to improve environmental performance of construction operations.”
J. Constr. Eng. Manage., 10.1061/(ASCE)CO.1943-7862.0000609,
404–413.
Real-Time Monitoring Using a Smartphone Akhavian, R., and Behzadan, A. H. (2012). “An integrated data collection
Smartphones are now prevalent in construction companies and analysis framework for remote monitoring and planning of con-
struction operations.” Adv. Eng. Inform., 26(4), 749–761.
(Constructech 2011). Since most current smartphones have MEMS
Albartar, A., Mekid, S., Starr, A., and Pertuszkiewicz, R. (2008). “Suitabil-
accelerometers, utilizing a smartphone as a sensor unit will provide ity of MEMS accelerometers for condition monitoring: An experimental
an efficient way to construct a comprehensive real-time monitoring study.” Sensor, MDPI, 8(2), 784–799.
system of an equipment fleet. The information on the measured Antoni, J., Daniere, J., and Guillet, F. (2002). “Effective vibration analysis
operational efficiency of each piece of equipment can be transferred of IC engines using cyclostationarity. Part I—A methodology for
wirelessly using the smartphone’s cellular network in real time, condition monitoring.” J. Sound Vib., 257(5), 815–837.
with the location information of equipment that is detected through Azar, E. R., and McCabe, B. M. (2012). “Automated visual recognition
a GPS unit embedded in the smartphone. This will provide of dump trucks in construction videos.” J. Comput. Civ. Eng., 26(6),
construction managers with more vivid and detailed information 769–781.
on on-going operations. Baldwin, K. C., Duncan, D. D., and West, S. K. (2004). “The driver monitor
system: A means of assessing driver performance.” Johns Hopkins APL
Tech. Digest, 25(3), 269–277.
Defining a Baseline Operational Efficiency for Typical Bao, L., and Intille, S. S. (2004). “Activity recognition from user-annotated
acceleration data.” Proc., Pervasive Computing 2004, Springer, Berlin,
Construction Operations
1–17.
Although operational efficiency is a key piece of information in Cheng, T., Migliaccio, G., Teizer, J., and Gatti, U. (2013a). “Data fusion
reliably estimating productivity and costs related to construction of real-time location sensing and physiological status monitoring for
equipment, there are only limited data available for estimators. Data ergonomics analysis of construction workers.” J. Comput. Civ. Eng.,
collection through the applications of the proposed approach will 10.1061/(ASCE)CP.1943-5487.0000222, 320–335.
Cheng, T., Teizer, J., Migliaccio, G., and Gatti, U. (2013b). “Automated
facilitate the definition of acceptable operational efficiencies for
task-level activity analysis through fusion of real time location sensors
each type of equipment in various types of construction operations. and worker’s thoracic posture data.” Autom. Constr., 29, 24–39.
This will not only help estimators to develop more reliable esti- Constructech. (2011). Constructech IT Playbook 2011, Specialty
mates that may result in a successful bid, but also allow field man- Publishing, Carol Stream, IL.
agers to have a solid basis on which to evaluate the performance of El-Omari, S., and Moselhi, O. (2009). “Integrating automated data acquis-
their on-site equipment operation. ition technologies for progress reporting of construction projects.”

© ASCE 04014042-10 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(2): 04014042


Proc., 26th ISARC, International Association for Automation and Lewis, P., Leming, M., Frey, H. C., and Rasdorf, W. (2011). “Assessing
Robotics in Construction. effects of operational efficiency on pollutant emissions of nonroad
Environmental Protection Agnecy (EPA). (2006). Compilation of state, diesel construction equipment.” J. Transp. Res. Board, 2233(2),
county, and local anti-idling regulations, EPA420-B-06-004, U.S. 11–18.
Environmental Protection Agency, Office of Transportation and Air Lewis, P., Leming, M., and Rasdorf, W. (2012). “Impact of engine
Quality, Ann Arbor, MI. idling on fuel use and CO2 emissions of nonroad diesel construction
EPA. (2008). Quantifying greenhouse gas emissions in key industrial equipment.” J. Manage. Eng., 10.1061/(ASCE)ME.1943-5479
sectors, EPA 100-R-08-002, Sector Strategies Division, US EPA, .0000068, 31–38.
Washington, DC. MacDonald, G. A. (1990). “A review of low cost accelerometers for vehicle
EPA. (2009). Potential for reducing greenhouse gas emissions in the dynamics.” Sensor Actuator Phys., 21(1–3), 303–307.
construction sector, US EPA, Washington, DC 〈http://www.epa.gov/ Monnot, J. M., and Williams, R. C. (2011). “Construction equipment
sectors/pdf/construction-sector-report.pdf〉 (Apr. 10, 2013). telematics.” J. Constr. Eng. Manage., 10.1061/(ASCE)CO.1943-7862
EPA Clean Air Act Advisory Committee (CAAAC). (2006). Recommen- .0000281, 793–796.
dations for reducing emissions from the legacy diesel fleet, US EPA, Moore, W. (2012). “Practical telematics.” Construction equipment, 〈http://
Washington, DC 〈http://www.epa.gov/diesel/documents/caaac-apr06 www.constructionequipment.com/practical-telematics〉 (May 5, 2013).
.pdf〉 (Jul. 24, 2010). Nichols, H., and Day, D. (2005). Moving the Earth: The workbook of
Downloaded from ascelibrary.org by Tufts University on 07/26/18. Copyright ASCE. For personal use only; all rights reserved.

Ergen, E., Akinci, B., East, B., and Kirby, J. (2007). “Tracking components excavation, McGraw-Hill, New York.
and maintenance history within a facility utilizing radio frequency
Peurifoy, R. L., Schexnayder, C. J., Shapira, A., and Schmitt, R. (2011).
identification technology.” J. Comput. Civ. Eng., 10.1061/(ASCE)
Construction planning, equipment, and methods, 8th Ed., McGraw-
0887-3801(2007)21:1(11), 11–20.
Hill, New York.
Gong, J., and Caldas, C. H. (2009). “Construction site vision workbench: A
Sacks, R., Radosavljevic, M., and Barak, R. (2010). “Requirements for
software framework for real-time process analysis of cyclic construction
building information modeling based lean production management
operations.” Proc., ASCE Int. Workshop on Computing in Civil Engi-
systems for construction.” Autom. Constr., 19(5), 641–655.
neering, ASCE, Reston, VA, 64–73.
Gong, J., and Caldas, C. H. (2011). “An object recognition, tracking, and Sardroud, J. M., Limbachiya, M., and Saremi, A. (2010). “Ubiquitous
contextual reasoning-based video interpretation method for rapid pro- tracking and locating of construction resource using GIS and RFID.”
ductivity analysis of construction operations.” Autom. Constr., 20(8), Proc., 6th GIS Conf. Exhibition, (GIS 88), National Cartographic
1211–1226. Center of Iran, Tehran, Iran.
Hagerty, J. R. (2011). “Big brother keeps an eye on heavy-equipment Sharif, M., and Lee, S. (2010). “Measurement of CO2 emissions from con-
fleet.” Wall Street J., 〈http://online.wsj.com/article/SB1000142405 struction vehicles.” Proc., CSCE 2010 Annual Conf./11th Int. Environ-
2748703509104576329881589249572.html〉 (Jan. 3, 2012). mental Structures Specialty Conf., Canadian Society for Civil
Hall, M., Frank, E., Holmes, G., Pfahringer, G., Reutemann, P., and Witten, Engineering (CSCE), Montreal.
I. A. (2009). “The WEKA data mining software: An update.” SIGKDD Spencer, B. F., Jr., Ruiz-Sandoval, M. E., and Kurata, N. (2004). “Smart
Explorations, 11(1), 10–18. sensing technology: Opportunities and challenges.” J. Struct. Contr.
Heydarian, A., Golparvar-Fard, M., and Niebles, J. C. (2012). “Automated Health Monitor., 11(4), 349–368.
visual recognition of construction equipment actions using spatio- STMicroelectronics. (2013). “LIS331DL datasheet.” 〈http://www.st
temporal features and multiple binary support vector machines.” Proc., .com/st-web-ui/static/active/en/resource/technical/document/datasheet/
Construction Research Congress, ASCE, Reston, VA. CD00172345.pdf〉 (May 5, 2013).
Joshua, L., and Varghese, K. (2011). “Accelerometer-based activity Zou, J., and Kim, H. (2007). “Using hue, saturation, and value color
recognition in construction.” J. Comput. Civ. Eng., 10.1061/(ASCE) space for hydraulic excavator idle time analysis.” J. Comput. Civ.
CP.1943-5487.0000097, 370–379. Eng., 10.1061/(ASCE)0887-3801(2007)21:4(238), 238–246.

© ASCE 04014042-11 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(2): 04014042

You might also like