sensors

Article
Design of Twin Builder-Based Digital Twin Online Monitoring
System for Crane Girders
Baogui Huang 1 , Yanbo Hui 1, *, Yonggang Liu 2 and Hongxiao Wang 1, *

1 College of Mechanical and Electrical Engineering, Henan University of Technology, Zhengzhou 450001, China;
2021920272@stu.haut.edu.cn
2 Postdoctoral Research Workstation of Weihua Group Co., Ltd., Xinxiang 453000, China;
201992296@stu.haut.edu.cn
* Correspondence: huiyb@haut.edu.cn (Y.H.); zzwhx2019@haut.edu.cn (H.W.)

Abstract: The crossbeam is frequently subjected to alternating loads during work as an essential
load-bearing part of the crane. However, due to the large volume and the limitations of detection
technology, it is impossible to realize online monitoring of the mechanical state. The ongoing
advancement of ROMing and digital twin technology plays a pivotal role in facilitating the resolution
of this particular issue. In this paper, we take the crane beam as the physical entity and combine
the Twin Builder reduced-order technology and Deployer digital twin deployment technology to
establish a digital twin of the beam. The load recognition model within the twin system exhibits a
prediction error rate of ±5%. Furthermore, the accuracy of the ROM surpasses that of conventional
machine learning models by a factor of 25. Upon deployment on the web platform, the results
are delivered within 0.5 s, representing a substantial improvement as it is merely 1/15 of the time
required for traditional 3D displays. The digital twin online monitoring system has the advantages
of high accuracy and low requirements for monitoring equipment, which can be widely used in
engineering practice to solve the problem that the mechanical state of large parts cannot be accurately
monitored online.

Keywords: crane girder; True-Load; reduced-order modelling; Twin Builder; online monitoring

1. Introduction
Citation: Huang, B.; Hui, Y.; Liu, Y.;
In the realm of industrial manufacturing, logistics, and related sectors, cranes assume
Wang, H. Design of Twin Builder-
a pivotal role as indispensable tools for driving industrial progress. Comprehensive
Based Digital Twin Online Monitoring
investigations and research have unearthed a disconcerting trend wherein the operation of
System for Crane Girders. Sensors
lifting machinery frequently results in catastrophic structural failures, particularly beam
2023, 23, 9203. https://doi.org/
fractures [1–5]. These accidents can be attributed to factors such as crane overloading,
10.3390/s23229203
microscopic metal cracks, and metal-fatigue-induced breakdowns. The recurring incidence
Received: 30 September 2023 of such grave safety mishaps underscores the imperative to maintain a vigilant focus on
Revised: 3 November 2023 the structural integrity of cranes. A meticulous analysis of the accident causes reveals that
Accepted: 7 November 2023 many of these issues are imperceptible to the naked eye, rendering the implementation of
Published: 15 November 2023
simple protective measures challenging. The prediction of real-time stress and deformation,
whether at the global or local level, in lifting machinery is critical for ensuring crane safety.
Consequently, the establishment of a real-time monitoring system for tracking stress and
Copyright: © 2023 by the authors.
deformation in lifting machinery plays a pivotal role in preventing crane safety accidents.
Licensee MDPI, Basel, Switzerland. The conventional approach of affixing strain gauges at fixed points for monitoring the
This article is an open access article condition of mechanical components is susceptible to the influence of structural intricacies,
distributed under the terms and which hinders the acquisition of precise data. Furthermore, the sheer magnitude of crane
conditions of the Creative Commons machinery renders comprehensive monitoring unattainable. In contrast, the finite element
Attribution (CC BY) license (https:// analysis method excels in accurately forecasting the overall stress and deformation of
creativecommons.org/licenses/by/ components under specific load conditions. However, its utility is limited in real-time
4.0/). prediction for larger structural elements of cranes due to computational constraints. This

Sensors 2023, 23, 9203. https://doi.org/10.3390/s23229203 https://www.mdpi.com/journal/sensors

Sensors 2023, 23, 9203 2 of 13

method is better suited for addressing the requirements of smaller components. The
ongoing development and enhancement of digital twin technology have equipped it to
effectively address the aforementioned challenges. The pivotal factor enabling digital twins
to accomplish real-time monitoring lies in the utilization of the ROM, a methodology that
systematically reduces the complexity of the original model through a sequence of model
simplification techniques. This results in the attainment of a low-order approximation that
closely approximates the original model to a remarkable degree [6–8].
In recent years, both domestic and international scholars have conducted extensive
research to investigate the practical application of ROMs and digital twin systems founded
on these models in the field of state detection. Various techniques are employed to derive
these ROMs, including Proper Orthogonal Decomposition (POD), Volterra Sequence Repre-
sentation, and the Harmonic Balance method. It is noteworthy that the selection of these
methods depends on the specific computational physics problem at hand, as each method
exhibits distinct applicability. Saddam Hijazi et al. [9] applied the classical POD–Galerkin
projection method to construct a tailored ROM designed for simulating turbulence within a
finite-volume environment. Through rigorous experimentation, it was consistently demon-
strated that the error between the ROM and the full-order model remained below 8%.
Furthermore, this approach achieved a remarkable 90% reduction in computational time
compared to the original method. D. Xiao et al. [10] integrated orthogonal decomposition
techniques with advanced machine learning methods to formulate a robust, high-fidelity,
and confident noninvasive ROM for monitoring ambient airflow states. This innovative
model shows significant promise in supplanting the conventional Gaussian smoke plume
model. David J. Lucia et al. [11] integrated the Volterra theory with tailored orthogonal
decomposition techniques to formulate an illuminative model for fluid-state surveillance,
yielding an error of under 5% when comparing the ROM to the full-order model. The
objective behind deriving ROMs from diverse physical systems using various techniques
is to lay the foundation for a digital twin system rooted in these ROMs. This digital twin
system [12–17] is then deployed to oversee the condition of physical entities, evaluate
their status through state monitoring data, and proactively implement timely preventive
measures. Han Dong et al. [18] employed an adaptive partitioning strategy in conjunction
with Galerkin methods to introduce a kinetic model designed for the rapid resolution of
fractures in large components. The empirical findings demonstrate that employing this
model substantially reduces the degrees of freedom while ensuring accuracy. Additionally,
the computational time on the CPU is drastically reduced by almost tenfold. Furthermore,
integrating this model with sensor components enables the creation of a digital twin system
for monitoring the structural integrity of large components, facilitating the early detection
and prediction of metal fracture risks. Khamlich et al. [19] have engineered a high-precision
ROM to address the advection–diffusion challenge within environmental monitoring, em-
ploying a data-driven approach that amalgamates proper orthogonal decomposition and
regression, known as POD-R. This methodology has culminated in the development of
a digital twin system designed for the real-time monitoring of airflow states in correla-
tion with environmental conditions. The system’s exceptional precision and timely data
acquisition underscore its potential benefits in the realm of environmental management,
particularly with respect to mitigating air pollution. Gianmarco Aversano et al. [20] have
innovatively crafted a ROM of the physical system by employing a novel blend of intrinsic
orthogonal decomposition and kriging interpolation. Notably, they introduce a pioneering
digital twin model tailored for furnaces operating under flameless combustion conditions.
This innovation empowers the accurate forecast of combustion data even when facing
unfamiliar combustion scenarios. Remarkably, their digital twin consistently forecasts
combustion pollutants with an impressively low margin of error, less than 5%, while simul-
taneously achieving an outstanding 80% reduction in computational time. This work serves
as a compelling showcase of the vast potential inherent in deploying digital twins based on
ROMs, particularly in the domain of condition monitoring. In the context of a jib crane,
Lai Xiaonan et al. [21] have introduced a novelty approach known as ‘Shape-Performance
Sensors 2023, 23, 9203 3 of 13

Integrated Digital Twin’. This method is proficient in generating diverse downscaled

models tailored to specific requirements. Furthermore, the team has engineered a dedicated
application for visualizing 3D predictive outcomes. In tandem, they have established a
comprehensive digital twin monitoring system, proficient in tracking the spatial orientation,
structural stress, and velocity characteristics of physical entities through the installation
of a variety of sensors. The acquired data serve as a foundation for making anticipatory
assessments regarding potential operational risks. An abundance of research findings
substantiates that the utilization of a reduced-order, low-order model not only upholds
remarkable accuracy but also substantially curtails computational time. Furthermore, the
digital twin system, founded upon this ROM, is proficient in effecting real-time monitoring
of the state of physical entities.
To mitigate the likelihood of critical safety incidents within lifting machinery, this
study amalgamates intrinsic orthogonal decomposition with machine learning techniques
to address the stress and deformation degradation model of crane girders. It further intro-
duces a load recognition model to furnish inputs for the degradation model, facilitating the
construction of a digital twin system for real-time monitoring of the stress and deformation
status of crane girders. To establish a comprehensive monitoring system, Deployer software
2021R1 is utilized. This system is integrated with strain sensors to continuously collect
real-time strain data, enabling the monitoring of the crane girder’s stress and deformation
status. The data are then visualized through a cloud map representing the overall stress
and deformation of the beam. Additionally, a line graph pinpoints specific stress changes,
facilitating an in-depth safety assessment. This approach ensures timely maintenance and
overhaul of the beam, effectively averting critical safety incidents such as beam fracture and
metal fatigue failure. The specific methodology is outlined as follows: initially, conducting
a finite element analysis of the crane beam under defined operational conditions; subse-
quently, employing the True-Load software 2022R1 [22–27] for strain test pre-analysis based
on the results obtained from the finite element simulations. This process includes acquiring
the positional and angular details of the strain gauge layout within a carefully selected
area, conducive for load inversion, and finally, utilizing real-world load and strain data to
construct a load recognition model in Twin Builder. Subsequently, the ROM module within
Twin Builder is employed to process the full-order finite element model. This process
leverages intrinsic orthogonal decomposition and machine learning techniques to produce
a ROM that aligns with the specified criteria. The result display window is then established,
finalizing the construction of the monitoring system. Strain gauge data are acquired via
CSV file input, and the corresponding data file input interface is preserved. An executable
SDK file is generated using Deployer software, facilitating the deployment of the digital
twin. To achieve real-time collection of stress–strain data and real-time data transmission
interface integration in the deployment file, the API secondary development interface
offered by Donghua sensor software 2021 is utilized. This culminates in the comprehensive
design of the digital twin monitoring system.

2. Load Physical Sensing System Design

The load physical sensing system is a bridge between the physical entity and the
digital twin, through the sensor to obtain the real-time strain data at the fixed point of
the beam, as the input of the digital twin to achieve the interaction between the physical
entity and the digital twin. The core of the system is to determine the sensing strain gauges’
patch position that can accurately invert the lifting load. For this problem, this paper uses
True-Load software to invert the optimal sensing strain gauges’ patch position according to
the simulation results under the actual working conditions.

2.1. Finite Element Simulation Analysis

The experiments are conducted on an overhead crane beam with a rated lifting weight
of 1 ton, a length of 5.85 m, a width of 1.5 m, and a cross-sectional width of 20 mm as a
physical entity. The beam material is Q235 structural steel with a modulus of elasticity of
Sensors 2023, 23, x FOR PEER REVIEW 4 of 13
Sensors 2023, 23, x FOR PEER REVIEW 4 of 13
Sensors 2023, 23, x FOR PEER REVIEW 4 of 13

2.1. Finite Element Simulation Analysis

2.1. Finite Element Simulation Analysis
2.1. Finite Element Simulation
The Element
experiments Analysis on an overhead crane beam with a rated lifting
are conducted
Sensors 2023, 23, 9203 2.1. Finite
The experiments Simulation Analysis on
are conducted an overhead crane beam with a rated lifting 4 of 13
weight Theof experiments
1 ton, a length areof conducted
5.85 m, a width on an of 1.5 m, and crane
overhead a cross-sectional
beam withwidth a rated of 20 mm
lifting
weight Theofexperiments
1 ton, a length areof conducted
5.85 m, a width on an of overhead
1.5 m, and crane a cross-sectional
beam withwidth a rated of 20 mm
lifting
as a physical
weight entity.
of 1 ton, The beam
a length of 5.85material
m, a width is Q235 of 1.5structural
m, and asteel with a modulus
cross-sectional width ofofelasticity
20 mm
as a physical
weight entity.
of 1 ton, The beam
a length of 5.85material
m, a width is Q235 of 1.5structural
m, and asteel with a modulus
cross-sectional width ofofelasticity
20 mm
of a206
as Mpa and
physical Poisson’s
entity. The beam ratiomaterial
of 0.3; the mainstructural
is Q235 body is made steelof square
with steel, and
a modulus the ends
of elasticity
of a206
as
206 Mpa Mpa
physical andand Poisson’s
entity.
Poisson’sThe beam ratio
ratio of 0.3;the
material
of 0.3; the mainbody
ismain
Q235 bodyisismade
structural made
steelof of square
with
square steel, and
a modulus
steel, and the ends
of elasticity
the
are206
of fixed
Mpa byand
welding
Poisson’sand ratio
bolting. of 0.3;Usethe Solidmain Works
bodysoftware
is made of 2020 to draw
square steel,itsand equal-scale
the ends
of
are206
fixedMpa byand Poisson’s
welding and ratio
bolting.of 0.3; Usethe Solidmain bodysoftware
Works is made of 2020square draw
to draw steel,itsand equal-scale
the ends
equal-scale
3D model,
are fixed bywith weldingoutput andasbolting.
x_t general Use format.
Solid Works The modelsoftware is converted
2020 to draw into itsa shell unit in
equal-scale
3D model,
are fixed bywith welding
with output
output and as x_t
x_tgeneral
asbolting. Use format.
general Solid
format. The
WorksThemodelmodelis is
software converted
2020 to draw
converted into a shell
its
into unit
aequal-scale
shell unitin
Space
3D model,Claim withof output
Workbench,as x_t which
generalisformat. convenient The model to extract the results
is converted intofor True-Load
a shell unit in
Space
in
3D Space
model,Claim
Claim
withof ofWorkbench,
Workbench,
output which
as x_t general whichisformat.
isconvenient
convenient
The model totoextract
extract the
the results
is converted into for
results for True-Load
a shell unit in
analysis.
Space The hexahedral
Claim of Workbench, meshwhich is selected to divide to
is convenient theextract
beam; the maximum results forcell size is 4
True-Load
analysis.
Space The
The of
Claim hexahedral
hexahedral
Workbench, mesh
mesh isisselected
which selected totodivide
is convenient divideto thethe beam;
beam;the
extract themaximum
maximum
results forcellcellsize
sizeisis4
True-Load
mm, the minimum
analysis. The hexahedral is 1 mm, meshthe is number
selected of tomeshdividecellsthe is 114,081,
beam; the and the number
maximum cellof nodes
size is 4
4 mm,
mm, thethe minimum
minimum isis
1 1
mm,mm, thethe number
number of
of mesh
mesh cells
cells
analysis. The hexahedral mesh is selected to divide the beam; the maximum cell size is
is 114,081,
114,081, and
and the
the number
number of nodes
is 4
is 111,895.
mm, When analyzing,
the minimum is 1 mm, the the endnumber weldofismesh set upcells
withisbound 114,081, contact,
and the and the self-weight
number of nodes
is 111,895.
mm, the minimum is 1 mm, the
When analyzing, the endnumber weldofismesh set upcells
withisbound114,081, contact,
and the and the self-weight
number of nodes
of111,895.
is the trolley, Whenspreader,
analyzing,andthe crane endbeamsweld is is set
added
up within the way of
bound gravity,
contact, and which is 1700 N,
the self-weight
of111,895.
is the trolley,When spreader,
spreader,
analyzing,and
andthecrane
craneendbeams weld is is set
added
up within the way contact,
bound of gravity, andwhich is 1700 N,
the self-weight
a preload
of the trolley,of 1000 N is given
spreader, to thebeams
and crane bolt joints,is addedand in thethe fixed
way constraints
of gravity,are whichapplied
is 1700to the
N,
a preload
of the trolley, of 1000 N is given
spreader, to thebeams
and crane bolt joints,
is added and in thethe fixed
wayconstraints
of gravity,are applied
which is 1700to the
N,
ends
a of theofbeams.
preload 1000 N The
is loadstowere
given the segmented
bolt joints, intothe
and fivefixed
sequential
constraintsload steps
are and applied
applied to the
aends of
of the
preload ofbeams.
the 1000 NThe
beams. The loads
loadsto
is given werethesegmented
were joints,into
segmented
bolt intofive
and fivesequential
the sequential
fixed load
load
constraints steps
steps
are and and
appliedapplied
applied
to theat
at positions
ends of the B, C, The
beams. D, E, andwere
loads F, assegmented
illustratedinto in five
Figure 1. Figure
sequential 2 steps
loadmesh shows and the mesh
applied
positions
at positions
ends B, C, D, E,
B, C, The
of the beams. and
D, E, F, as illustrated
andwere
loads F, as in Figure
illustratedinto
segmented 1. Figure
in five
Figure 2 shows
1. Figure
sequential the
load2 stepsshows delineation
and the mesh
applied of
delineation
at
the positions ofB,the
crane crossbeam, C,crane
D, E,crossbeam,
and F, asFigureillustrated 3 shows in theFigurestress 1.simulation
Figure result
2 shows cloud,
the and
mesh
delineation
at positions ofB,the D,Figure
C,crane and3 shows
E,crossbeam, F, asFigurethe stress
illustrated simulation
3 shows in theFigure result
stress cloud, 2and
1. simulation
Figure Figure
result
shows 4 shows
cloud,
the and
mesh
Figure
delineation
the 4 shows
deformation of the
the deformation
crane
simulation crossbeam,
result simulation
Figure
cloud, 3 result
shows cloud,
the which
stress shows
simulation that the
result maximum
cloud, and
Figure
delineation 4 shows
of thethe deformation
crane crossbeam, Figurewhich
simulation 3 result
shows shows
cloud, that
the stresswhichthe maximum
shows that
simulation deformation
the cloud,
result maximum andof
deformation
Figure
the 4
crossbeamshows of the crossbeam
deformation is directly
simulation below
result the force
cloud, loading
which point.
shows that This
the simulation
maximum
deformation
Figure 4 shows ofisthe
directly
the belowisthe
crossbeam
deformation force loading
directly
simulation below
resultthepoint.
cloud,forceThis simulation
loading
which showspoint. result
that theissimulation
This used as a
maximum
result is used
deformation
sample as
of a sample
the of thefor
crossbeam full-order
is directly model
below for
thethe forceROM, marking
loading point. the loads
This as inputs
simulation
result isof
deformation theof
used full-order
as a sample
the model
crossbeamof theis the ROM,
full-order
directly model
belowmarking
for
thethe the
force loads
ROM, as inputs
marking
loading point. the andloads
This stresses and
as inputs
simulation
and
result
strain stresses
is used
values and
as
as a strain
sample
outputs. values
of the as outputs.
full-order model for the ROM, marking the loads as inputs
and stresses
result is usedand as astrain
sample values
of theas outputs.model for the ROM, marking the loads as inputs
full-order
and stresses and strain values as outputs.
and stresses and strain values as outputs.

Figure 1. Load
Figure Load setting.
setting.
Figure 1.
1. Load setting.
Figure 1. Load setting.
Figure 1. Load setting.

Figure 2. Finite element meshing.

Figure 2.
2. Finite element meshing.
Figure 2. Finite
Figure Finite element
element meshing.
meshing.
Figure 2. Finite element meshing.

Figure 3. Stress map.

Figure 3. Stress map.
Figure 3. Stress map.
Figure 3. Stress map.

Figure 4. Deformation map.

Figure 4. Deformation map.
Figure 4. Deformation map.
Figure
Figure 4.
4. Deformation
Deformation
2.2. True-Load map.
map. Analysis
Load Inverse
2.2. True-Load Load Inverse Analysis
2.2. True-Load
Import the Load Inverse
result Analysis
file Analysis
of Workbench FEA into True-Load, select the target area for the
True-Load
2.2. True-Load
Import the Load Inverse
result file of Workbench FEA into True-Load, select the target area for the
strainImport
gauges in the appropriate
the result file of Workbench position FEAof the
intomain beam in
True-Load, order
select thetotarget
avoidareatoofor
much
the
strainImport
gauges theinresult of Workbench
the appropriate
file of Workbench
position of the
FEA intomain beam in
True-Load, order
select thetotarget
avoidareatoofor
much
the
concentration
strain gauges in the the paste position,
appropriate and break
position of the up main
the target
beamarea with ato
in order spacing
avoid of too10much
mm.
strain
strain gauges in the
concentration
gauges the paste
appropriate
position,
appropriate position
and break
position of the
of the main
up main beam
the target
beam area with atospacing
in order avoid of too10much
mm.
Then, the five in
concentration load
thesteps
pasteinposition,
the finiteandelement
break analysis are uploaded,
up the target area withset as the number
a spacing of 10 mm.of
concentration in the paste position,
Then, the five load steps inposition,the finite and
and break
element up
break analysisthe target area
are uploaded,
up the target with
area with set a spacing of
as the number
a spacing 10 mm.
of 10 mm.of
strain the
Then, gauges
five 5, andsteps
load clickinthe
the Start button
finite to run
element the calculation
analysis are uploaded,to obtain
set asthe
thecoordinates
number of
Then,
strain the five 5,
gauges load
andsteps
clickinthe
theStart
finite element
button to run analysis are uploaded,
the calculation set asthe
to obtain thecoordinates
number of
of the gauges
strain five best-sensed
5, and load
click the strain
Start gauges’
button patch
to run positions
the and the
calculation to X, Y, Zthe
obtain axis deflection
coordinates
strain
strain gauges
of the gauges 5, and
and click
five best-sensed
5, click
loadthestrain
the Start gauges’
Start button to
button to run the
patch
run the calculation
positions and the
calculation to obtain
to obtain the
X, Y, Zthe coordinates
axis deflection
coordinates
angles.
of the
of The
the five interface
five best-sensed of
best-sensed load the analysis
load strain results
strain gauges’
gauges’ patchis shown
patch positionsin Figure
positions andand 5.5.
the
the The
X, Condition
Y, Z
X, Y,
Y, Z axis Number
axis deflection
deflection
angles.
of Thebest-sensed
the five interface ofload the analysis resultspatch
strain gauges’ is shown in Figure
positions and theThe
X, Condition
Z axis Number
deflection
(CN) ofThe
angles.
angles. theinterface
The analyzedof
interface ofpatch is 1.67,results
the analysis
the analysis and the
results evaluation
is shown
is shown criteria
in Figure
in Figure are Condition
5. The
5. The as follows:Number
Condition CN = 1
Number
(CN) ofThe
angles. theinterface
analyzedof patch
the is 1.67,results
analysis and the is evaluation
shown in criteria
Figure 5. are Condition
The as follows:Number
CN = 1
Perfect;
(CN)
(CN) of
of CN
the = 1~10
analyzedVery Good;
patch is CN =
1.67, 10~20
and theGood; CN
evaluation= 21~50 Usable;
criteria are CN
as > 51
follows: Unusable.
CN =1
(CN) ofthe
Perfect; theanalyzed
CN =analyzed
1~10 Verypatch is 1.67,
Good;
patch is CN and the evaluation
= 10~20
1.67, and theGood; CNcriteria
evaluation= 21~50areUsable;
criteria asare
follows:
CN
as >CN51 =Unusable.
follows: 1CN
Perfect;
=1
Perfect;
CN = 1~10 CN = 1~10 Very Good; CN = 10~20 Good; CN = 21~50 Usable; CN > 51 Unusable.
Perfect; CN Very
= 1~10 Good;
Very CN Good;= 10~20
CN = Good; CN = CN
10~20 Good; 21~50 Usable;
= 21~50 CN >CN
Usable; 51 Unusable. The
> 51 Unusable.
corresponding scaling factors of the five load steps are similar, and the scaling factors
are identical to each other, as shown in Figure 6, which is in line with the experimental
expectation, and the patch position results are shown in Figure 7.
Sensors
Sensors2023,
2023, 23,
23, xx FOR
FOR PEER REVIEW
PEER REVIEW 5 of5 13
of 13

The corresponding
The correspondingscaling
scalingfactors
factorsofofthe
thefive
fiveload
loadsteps
steps are
are similar,
similar, and
and thethe scaling
scaling factors
factors
Sensors 2023, 23, 9203 5 of 13
are identical to each other, as shown in Figure 6, which is in line with the
are identical to each other, as shown in Figure 6, which is in line with the experimental experimental
expectation,and
expectation, andthe
thepatch
patchposition
positionresults
resultsare
areshown
shown inin Figure
Figure 7. 7.

Figure 5. True-Load condition number results.

Figure 5.
Figure 5. True-Load
True-Loadcondition
conditionnumber
numberresults.
results.

Figure 6.
Figure 6. Analysis
Analysis of
of measurement point errors.
measurement point errors.
Figure 6. Analysis of measurement point errors.

Figure 7.
Figure 7. Strain
Strain gauges’
gauges’ location.
location.
Figure 7. Strain gauges’ location.
2.3. Physical Load
2.3. Physical Load Sensing
Sensing Modelling
Modelling
2.3. Physical
The
The functionLoad of
function Sensing
ofthetheloadModelling
sensing
load sensingmodel is to invert
model is to the real-time
invert lifting mass
the real-time according
lifting mass
to the
accordingvalues
The function of the
to the values sensing
of the
of the load
load strain
sensing
sensing gauges
load model pasted
strain gauges on the
is to invert crossbeam
pasted the on the during
real-time
crossbeam the
liftingcrane
duringmass
working
the craneprocess,
according working which
process,
to the values can besensing
which
of the realized
can be byrealized
load the built-in
strain by response
the
gauges pastedsurface
built-in on theROM
response module
surface
crossbeam ROM of
during
Twin
module Builder
the crane software2023R1.
ofworking
Twin Builder process, The
software2023R1.steps
which can beThe are to take the
steps are
realized change
to take
by the history
the change
built-in of the
response strain
history
surface gauge
of the
ROM
value
strain as input and the change history of the corresponding
module of Twin Builder software2023R1. The steps are to take the change history ofasthe
gauge value as input and the change history of the lifting mass
corresponding as output
lifting during
mass
multiple sets ofmultiple
experiments, import the response surface
output during
strain gauge value as input sets and
of experiments,
the change import
history theROM,
ofthe response andsurface
corresponding generate ROM,
lifting the load
and as
mass
sensing
generate model,
the load so it is necessary
sensing model, to carry out multiple
so it is necessary sets of
to carry experiments
out surface to
multipleROM, establish
sets ofand
output during multiple sets of experiments, import the response
the control relationship
experiments between the strain gaugebetween value and the load. When the crane is
generate thetoload establish
sensingthe control
model,relationship
so it is necessary the to strain
carry gauge value
out multiple andsets
the of
running,
load. When the trolley
the cranewillis move along
running, thethe beam,
trolley and
will its
move position
along will
the be
beam,changed;
and to
its achieve
position
experiments to establish the control relationship between the strain gauge value and the
the
will prediction
beWhen
changed; of the mass of thethe lifted weightof thewhen the trolley
liftedisthein a different situation,
load.
the trolley is the to
stopped
achieve
crane atisfive
running,prediction
equal the trolley
position centers.
mass
will So,
of the
move along
establish
weight
five beam,when
channels and
to
theitstrolley
carry position
out
is in
will a
bedifferent
changed; situation,
to achieve the trolley
the is
predictionstopped
of theat five
mass equal
of theposition
lifted centers.
weight So,
when establish
the trolley
multigroup lifting experiments.
five
is inchannels
a different to carry out multigroup
situation, lifting experiments.
Experiment with floodthe trolley
control is
sandbagsstoppedas theat weight,
five equal dueposition centers.
to the quality of So, establish
sandbags,
five Experiment
channels to with
carry flood
out control
multigroup sandbags
lifting as the weight, due to the quality of
experiments.
is not uniform, according to the actual weighing situation set up: 0 N, 1014.1 N, 1977.5 N,
sandbags, is not uniform, according to the actual weighing situation set up: 0 N, 1014.1 N,
3048.5Experiment
N, 4121 N, 4986.5with floodN, 5947control
N, 6890.1 sandbags
N, 7796.1asN,the 9025.1weight,
N weight due oftothethetenquality
groups of
1977.5 N, 3048.5 N, 4121 N, 4986.5 N, 5947 N, 6890.1 N, 7796.1 N, 9025.1 N weight of the
sandbags,
of is not 120
weight lifting; uniform,
mm ×according
120 mm ×to themm
140 actual
× 2 weighing
T specification situation set up:
soft pallet as0lifting
N, 1014.1
gear, N,
ten groups of weight lifting; 120 mm × 120 mm × 140 mm × 2 T specification soft pallet as
1977.5 N,test
Donghua 3048.5 N, 4121real-time
DH5908N N, 4986.5 N, 5947
strain N, 6890.1
collector as the N, data 7796.1 N, 9025.1
acquisition N weight
device, according of the
lifting gear, Donghua test DH5908N real-time strain collector as the data acquisition
tenthe
to groups of weight
coordinates lifting;
of the 120 mmsolution
True-Load × 120 mm × 140
in the mm × 2 T specification
corresponding position ofsoft the pallet
paste as
liftinggauges,
strain gear, Donghua
supporting testtheDH5908N
use of thereal-time
DHDAS dynamic strain collector as the datasystem,
strain acquisition acquisition
in
different positions were lifting the above different masses of weight, to be stable when the
stability of the repeated recording of each strain gauge value is taken for the average of
 the table are used as the strain change history, and the last column is used as the load
change history, which is imported into the response surface ROM to generate the required
load recognition ROM, i.e., the physical load sensing model.
In order to verify the accuracy of the physical load sensing model, a total of ten sets
Sensors 2023, 23, 9203
of experiments were measured starting from 1000 N and spaced from 1000 N to 10,000 6 of 13
N,
comparing the mass of the actual load with the mass output from the load recognition
ROM, and the absolute value of the error rate was calculated as shown in Figure 8b. The
results show that the difference between the prediction results of the physical load sensing
the value. The experimental field diagram is shown in Figure 8a, and the measurement
model
resultsbased on the
at position response
three surface
are shown ROM
in Table andfirst
1. The the five
actual load will
columns change
in the with
table are theasrise
used
ofthe
thestrain change history, and the last column is used as the load change history, whichwhich
lift weight, but the absolute error rate of the two is always less than 2.0%, is
indicates
importedthat intothe
thephysical
responseload sensing
surface ROM model has a high
to generate prediction
the required loadaccuracy. It can
recognition satisfy
ROM,
the
i.e.,needs of the practical
the physical scenarios.
load sensing model.

(a) (b)
Figure
Figure8.8.Test
Testsite
site and
and results analysis.(a)
results analysis. (a)Strain
Straintest
testsite;
site;(b)
(b)Load
Load error
error analysis.
analysis.

Table
Table1.1.Strain
Strainand
and Load Measurements.
Load Measurements.

Strain
StrainGauge
Gauge1/με
1/µε Strain
StrainGauge
Gauge2/με
2/µε StrainGauge
Strain Gauge 3/µε
3/με Strain
Strain Gauge
Gauge 4/με Strain
4/µε Strain Gauge
Gauge 5/µε5/με MassMass of Weight/kg
of Weight/kg
1.1 3.7 4.3 1.3 1.033 40.01
2.41.1 3.7
6.4667 4.3
7.6333 1.3
2.7333 1.033
2.3333 40.0170.21
2.4
5.1333 6.4667
13.7 7.6333
15.2333 2.7333
3.7333 2.33334.59 70.21
119.71
5.1333
7.1 13.7
19.1333 15.2333
23.3333 3.7333
8.8667 4.59
7.3333 119.71
192.69
7.1 19.1333 23.3333 8.8667 7.3333 192.69
12.0333 26.5 32.4667 12.8333 10.633 277.19
12.0333 26.5 32.4667 12.8333 10.633 277.19
13.8667 31.5667 39.4667 15.4667 12.111 333.94
13.8667 31.5667 39.4667 15.4667 12.111 333.94
16.2 36.0667 45.8667 18.2667 15.05 393.14
16.2 36.0667 45.8667 18.2667 15.05 393.14
17.5 41.5 52.633 19.8 16.89 452.8
17.5 41.5 52.633 19.8 16.89 452.8
20.833 46.8 59.333 23.5333 19.4 513.55
20.833 46.8 59.333 23.5333 19.4 513.55
22.4667 50.8 63.9 26.533 21.1833 550.75
22.4667 50.8 63.9 26.533 21.1833 550.75
24.2
24.2 54.6
54.6 69.433
69.433 27.133
27.133 23.333
23.333 601.75
601.75
25.2
25.2 57.3
57.3 73.133
73.133 27.9333
27.9333 24 24 632.4632.4

In order to verify the accuracy of the physical load sensing model, a total of ten sets
of experiments were measured starting from 1000 N and spaced from 1000 N to 10,000 N,
comparing the mass of the actual load with the mass output from the load recognition
ROM, and the absolute value of the error rate was calculated as shown in Figure 8b. The
results show that the difference between the prediction results of the physical load sensing
model based on the response surface ROM and the actual load will change with the rise
of the lift weight, but the absolute error rate of the two is always less than 2.0%, which
indicates that the physical load sensing model has a high prediction accuracy. It can satisfy
the needs of the practical scenarios.

3. Reduced-Order Model Construction

The ROM creation method provided by Twin Builder software carries out the intrinsic
orthogonal decomposition of the finite element full-order model to obtain a set of orthog-
onal basis functions, which are used as the subspace for the ROM, and then construct a
polynomial response surface approximation model using the data difference technique
Sensors 2023, 23, 9203 7 of 13

integrated with ANSYS. The ROM is obtained, which can be used for real-time mechanical
analysis under different input parameters. To validate the accuracy of the ROM, the predic-
tion model obtained from the training of the undecomposed full-order model data using a
BP neural network [19–22] is used as a reference group to carry out a comparative analysis.

3.1. Generating a ROM

Creating a ROM requires multiple full-order models as sample data, each representing
different load conditions within the effective lifting load range. The design of experiments
(DOE) function provided by the Workbench software2022R1 can randomly set different
input sample points according to the experimental requirements in combination with the
lifting weight range. There are several methods for random sampling design, including
intermediate composite design, optimal space-filling design, Box–Behnken design, and
Latin hypercube sampling design. In order to ensure the practicability and extensiveness
of the model sample data, this paper adopts user-defined settings. It adds 20% overload
simulation experiments, setting up a total of 24 groups of sample experiments from 500 N
to 12,000 N loads, with a gradient of 500 N, and saving the simulation results corresponding
to the different load conditions as two folders of Stress and Deformation as the downscaled
model. At the same time, the simulation results were saved in a separate data format as the
training dataset for the full-order model of the BP neural network in the reference group.
The generated Stress and Deformation results folder is imported into the Static ROM
Builder module in Twin Builder, and a certain proportion of data is selected as the Snapshot
matrix to generate the model order error analysis curve with the descending model order
as the horizontal coordinate and the error as the vertical coordinate. The blue curve is the
size of the error between the ROM and the full-order model for the same point, and the
green curve is the trend of the blue curve with the change in the order of the ROM. The
generated curve needs to meet the following two conditions: first, after the intersection of
the blue curve and the green curve, the green curve needs to be on top of the blue curve;
second, after the intersection of the green curve, the trend of the green curve changes
should be smooth, and there is no cliff change; this is achieved by constantly adjusting the
ratio of the data. The correct case is shown in Figure 9a, and the error is shown in Figure 9b.
The highest order of the ROM that conforms to the model analysis curve is significantly
reduced, and the operation amount is suddenly reduced. The machine learning method is
Sensors 2023, 23, x FOR PEER REVIEW 8 of 13
used to learn the sample data of the ROM, and the twin models that can predict the inputs
of the unknown load, the stress ROM and the deformation ROM, can be obtained.

(a) (b)
Figure 9.
Figure 9. Error
Error curve
curve setting:
setting: (a) correct form; (b) wrong
wrong form.
form.

The reference group is the extracted data imported into MATLABR2023a software in
the format according to the load, node X, Y, and Z coordinates, and mechanical state data,
and the training data are learned using the constructed single-hidden-layer BP neural
network oriented to the Delta rule. After continuous attempts, when the number of hidden
layers is 7, the proportion of training data is 80%, the proportion of validation data is 10%,
and the proportion of test data is 10%, the model training effect is the best, as shown in
Sensors 2023, 23, 9203 (a) (b) 8 of 13
Figure 9. Error curve setting: (a) correct form; (b) wrong form.

The reference
The reference group
group isis the
theextracted
extracteddatadataimported
importedinto
intoMATLABR2023a
MATLABR2023a software
softwarein
the format according to the load, node X, Y, and Z coordinates, and mechanical
in the format according to the load, node X, Y, and Z coordinates, and mechanical state state data,
and the
data, andtraining datadata
the training are are
learned
learnedusing
usingthetheconstructed
constructedsingle-hidden-layer
single-hidden-layer BP neural
BP neural
networkoriented
network orientedto
tothe
theDelta
Deltarule.
rule. After
After continuous
continuous attempts,
attempts, when
when thethe number
number of of hidden
hidden
layers is
layers is 7,
7, the
the proportion
proportion ofof training
training data
data is
is 80%,
80%, the
the proportion
proportion of validation
validation data is 10%,
and the
and the proportion
proportion ofof test
test data
data is
is 10%,
10%, the
the model
model training
training effect
effect is
is the
the best,
best, as
as shown
shown inin
Figure 10.
Figure 10.

Figure 10. BP
Figure 10. BP neural
neural network
network training
training results.
results.

The load recognition model and the step-down model are visualized using Twin
The load recognition model and the step-down model are visualized using Twin
Builder software2022 individually. Additionally, the output ‘F’ ports of the load recognition
Builder software2022 individually. Additionally, the output ‘F’ ports of the load
model are interconnected with the ‘Force_Magnitude’ ports of the two step-down models.
recognition model are interconnected with the ‘Force_Magnitude’ ports of the two step-
This configuration facilitates the utilization of the load recognition model to furnish load
down models. This configuration facilitates the utilization of the load recognition model
information to the step-down models. Furthermore, the system configuration includes the
to furnish load information to the step-down models. Furthermore, the system
creation of an overall stress and deformation cloud map for the beam, along with stress
configuration includes the creation of an overall stress and deformation cloud map for the
or deformation line graphs at specific points. This comprehensive simulation system is
beam, along with stress or deformation line graphs at specific points. This comprehensive
presented in Figure 11. Figure 11 depicts the simulation system, with the cart positioned
simulation system is presented in Figure 11. Figure 11 depicts the simulation system, with
at beam marking location 3. Similar simulation systems have been constructed for the
remaining four positions following the same operational procedure. This results in a
comprehensive simulation system comprising five positions, as illustrated in Figure 12. The
simulation system operates by assessing the magnitudes of five strain gauge values and
assigning a value of 2 to the location with the highest strain measurement while assigning
a value of 1 to the remaining positions. These values are then stored in a table denoted as
‘C’. Data in ‘C’ are compared with a default value of 1 within the switch function. When
it exceeds 1, the switch is engaged to execute the simulation at this position, and vice
versa. In this context, ‘Y’ denotes the strain gauge numerical interface, ‘Z’ signifies the load
recognition model for each position, ‘S’ represents the stress (stress) reduction model for
each location, ‘D’ stands for the deformation (deformation) reduction model for the specific
position, and ‘plus’ is employed to configure the pulse function, determining the frequency
of the cloud output.
 measurement
then stored in while
a tableassigning
denoted aasvalue of 1 in
‘C’. Data to ‘C’
theare
remaining
compared positions. These values
with a default value ofare1
then stored in a table denoted as ‘C’. Data in ‘C’ are compared with a
within the switch function. When it exceeds 1, the switch is engaged to execute the default value of 1
within the switch function. When it exceeds 1, the switch is engaged to
simulation at this position, and vice versa. In this context, ‘Y’ denotes the strain gauge execute the
simulation
numerical at this position,
interface, and vicethe
‘Z’ signifies versa.
loadInrecognition
this context,model
‘Y’ denotes the strain
for each gauge
position, ‘S’
numerical
represents interface,
the stress ‘Z’ signifies
(stress) the load
reduction modelrecognition
for eachmodel for ‘D’
location, eachstands
position, ‘S’
for the
Sensors 2023, 23, 9203 represents 9 ofthe
13
deformationthe(deformation)
stress (stress)reduction
reductionmodel
modelforforthe each location,
specific ‘D’ stands
position, for
and ‘plus’ is
deformation (deformation)
employed to configure reduction
the pulse function,model for thethe
determining specific position,
frequency of the and
cloud‘plus’
output. is
employed to configure the pulse function, determining the frequency of the cloud output.

Figure11.
Figure 11.Digital
Digitaltwin
twinsimulation
simulationsystem
systemfor
foraasingle
singlelocation.
location.
Figure 11. Digital twin simulation system for a single location.

Figure 12. The entire digital twin simulation system.

Figure12.
Figure 12.The
Theentire
entiredigital
digitaltwin
twinsimulation
simulationsystem.
system.

3.2. Reduced-Order Model Error Analysis

The maximum difference between the stress and deformation prediction results of
the down-order model and the full-order BP neural network learning model and the finite
element simulation results are calculated, respectively, under the same loading conditions
as shown in Figure 13. The error between the stress and deformation prediction results of
the ROM and the simulation results is minimal, which is one-twenty-fifth of that of the
full-order BP neural network learning model. Meanwhile, due to the excessive amount
of data processed by the full-order BP neural network learning model, the time used is
much higher than that of ROM, and the comparison data are shown in Table 2. The results
reveal that the BP neural network necessitates 200 times more time for computation than
the ROM. Furthermore, the aforementioned findings underscore the remarkable accuracy
of the ROM achieved within a significantly shorter time frame.
 the ROM and the simulation results is minimal, which is one-twenty-fifth of that of the
full-order BP neural network learning model. Meanwhile, due to the excessive amount of
data processed by the full-order BP neural network learning model, the time used is much
higher than that of ROM, and the comparison data are shown in Table 2. The results reveal
that the BP neural network necessitates 200 times more time for computation than the
Sensors 2023, 23, 9203 10 of 13
ROM. Furthermore, the aforementioned findings underscore the remarkable accuracy of
the ROM achieved within a significantly shorter time frame.

(a) (b)
Figure13.
Figure 13.Error
Error analysis
analysis for for
ROM:ROM: (a) Comparative
(a) Comparative analysis
analysis oferrors;
of stress stress (b)
errors; (b) Comparative
Comparative analysis
analysis of deformation
of deformation errors. errors.

Table2.2.Comparison
Table Comparisonofofthe
thetime
timeofofuse
useofofthe
thetwo
twomethods.
methods.

Back-Propagation (BP) Neural Network/s

Loads/N ROM/s Back-Propagation (BP) Neural Network/s Multiple of Both
Loads/N ROM/s Computation Time/s Cloud Map Display Time/s Total Time/s Multiple of Both
1000 0.152 Computation
1.5Time/s Cloud Map Display
35 Time/s Total Time/s
36.5 240.2
2000
1000 0.149
0.152 1.51.4 35 35 36.4
36.5 244.3
240.2
3000
2000 0.155
0.149 1.41.5 35 35 36.5
36.4 235.5
244.3
3000
4000 0.155
0.146 1.51.5 35 35 36.5
36.4 235.5
249.3
4000
5000 0.146
0.151 1.51.5 35 35 36.4
36.5 249.3
241.8
5000
6000 0.151
0.148 1.51.5 35 35 36.5
36.5 241.8
246.6
6000
7000 0.148
0.153 1.51.5 35 35 36.5
36.5 246.6
238.6
7000 0.153 1.5 35 36.5 238.6
8000 0.149 1.4 35 36.4 244.3
8000 0.149 1.4 35 36.4 244.3
9000 0.148 1.4 35 36.4 245.9
9000 0.148 1.4 35 36.4 245.9
10,000
10,000 0.152
0.152 1.51.5 35 35 36.5
36.5 240.1
240.1

4. Deployment and Application of Digital Twin Models

4. Deployment and Application of Digital Twin Models
The twin that meets the accuracy requirements is compiled in the Twin Builder
The twin that meets the accuracy requirements is compiled in the Twin Builder envi-
environment to generate a twin format file, reserving the input interface for the sensed
ronment to generate a twin format file, reserving the input interface for the sensed load
load strain gauges and the output interface for the cloud diagrams and stress change
strain gauges and the output interface for the cloud diagrams and stress change curves
curves that need to be displayed. Then, in the Deployer software, drag in the CSV-input
that need to be displayed. Then, in the Deployer software, drag in the CSV-input data
data input module, import the CSV file of the measured strain gauge data change history,
input module, import the CSV file of the measured strain gauge data change history, and
and display each strain gauge data output interface; drag in the Twin Model module that
display each strain gauge data output interface; drag in the Twin Model module that can
can encapsulate the twins, import the generated digital twins of the twin format, and
encapsulate the twins, import the generated digital twins of the twin format, and display
display the reserved input and output interfaces; and set the strain gauge data output
the reserved input and output interfaces; and set the strain gauge data output interfaces.
interfaces. Connect the strain gauge data output interface with the twin one-to-one data
Connect the strain gauge data output interface with the twin one-to-one data input inter-
face, and the twin data output interface does not need to be connected. Use the system
default settings to output the SDK deployment file. Deployer provides three forms of
deployable SDK file generation methods: Linux system web deployment, windows system
web deployment, and a small black window deployment. After experimental comparison,
web deployment can not only obtain the current moment to stress cloud map, it can also be
set up through the button to call out the specific point of the stress change curve, a more
intuitive image.
The above digital twin system uses the historical data collected by the sensed load
strain gauges, and the real-time strain gauge measurement data still need to complete the
connection with the twin system. To address this problem, this paper writes a program
through the API secondary development interface provided by Donghua test software to
connect the real-time collected sensed load strain gauge channel of the sensor with the
strain data interface in the generated SDK deployment file and to judge the five values,
 curve, a more intuitive image.
The above digital twin system uses the historical data collected by the sensed load
strain gauges, and the real-time strain gauge measurement data still need to complete the
connection with the twin system. To address this problem, this paper writes a program
Sensors 2023, 23, 9203 through the API secondary development interface provided by Donghua test software 11 of 13 to
connect the real-time collected sensed load strain gauge channel of the sensor with the
strain data interface in the generated SDK deployment file and to judge the five values,
and
andgive
give the
the maximum valuechannel
maximum value channeloutputs
outputsthe
the value
value ofof size
size forfor
thethe role
role of the
of the switching
switching
controller,
controller, to judge whether to output the corresponding position of the ROM andand
whether to output the corresponding position of the ROM to to
achieve
achieve the
the digital twin.
twin. The
Thedeployment
deploymentofofthe
the system
system application
application is shown
is shown in Figure
in Figure 14. 14.

Figure
Figure14.
14. Applications of digital
Applications of digitaltwins.
twins.

5.5.Conclusions
Conclusions
To mitigate
To mitigatethethe
riskrisk
of safety incidents
of safety such as beam
incidents such breakage
as beamduring crane during
breakage operations,
crane
this study introduces a digital twin condition monitoring system centered on the reduced-
operations, this study introduces a digital twin condition monitoring system centered on
order model of beams. This system is tasked with continuous real-time monitoring of
the reduced-order model of beams. This system is tasked with continuous real-time
both global and local stress and deformation levels within the beam. It then transmits the
monitoring of both global and local stress and deformation levels within the beam. It then
collected data to equipment maintenance engineers, enabling them to evaluate the safety
transmits
condition the collected
of the beam. When data engineers
to equipment
detectmaintenance engineers,deviations
stress or displacement enabling in
them
the to
evaluate the safety condition of the beam. When engineers detect stress or displacement
beams beyond permissible limits, they promptly initiate maintenance at the relevant beam
locations to ensure their continued normal operation. Upon comparing the output results
of the ROM with those of the full-order finite element model under identical input loads,
it is evident that the maximum error between the ROM employed in this study and the
simulation results from the full-order finite element model is well below 1.0 × 10−15 . This
compellingly illustrates the exceptional fidelity of the ROM. The ROM’s computational
time is merely 1/200 of the finite element analysis time under identical input conditions,
unequivocally affirming its exceptional computational efficiency. Utilizing the Deployer
software, we compile the entire monitoring system into an executable SDK file, enabling
independent execution apart from the Twin Builder software. This approach minimizes
the computational demands on the host system, facilitating a seamless connection to the
DHDAS strain monitoring system via the data interface for real-time monitoring. The
Sensors 2023, 23, 9203 12 of 13

digital twin-based online beam monitoring system presented in this study holds significant
engineering applicability. It not only enables swift data acquisition but also offers the
advantage of minimizing computer hardware demands, thereby conserving valuable time
and resources. To enhance its capabilities, this monitoring framework will be extended
through the incorporation of diverse sensor types. Furthermore, the historical data amassed
will be meticulously collated and subjected to rigorous analysis for predictive insights into
beam life span.

Author Contributions: Data curation, B.H. and H.W.; creating simulation systems and analyses, B.H.;
writing—original draft, B.H.; writing—review and editing, B.H. and H.W.; supervision, Y.H. and Y.L.
All authors have read and agreed to the published version of the manuscript.
Funding: Postdoctoral Research Initiation Program in Henan Province (Grant No. 202101052), This
research was funded by the Science and Technology Research Project of Henan Province, China
(Grant No. 222102220080).
Data Availability Statement: Data are contained within the article.
Conflicts of Interest: Author Yonggang Liu was employed by the company Postdoctoral Research
Workstation of Weihua Group Co., Ltd., Xinxiang 453000, China. The remaining authors declare that
the research was conducted in the absence of any commercial or financial relationships that could be
construed as a potential conflict of interest.

References
1. Beavers, J.E.; Moore, J.R.; Rinehart, R.; Schriver, W.R. Crane-related fatalities in the construction industry. J. Constr. Eng. Manag.
2006, 132, 901–910. [CrossRef]
2. Lingard, H.; Cooke, T.; Zelic, G.; Harley, J. A qualitative analysis of crane safety incident causation in the Australian construction
industry. Saf. Sci. 2021, 133, 105028. [CrossRef]
3. Neitzel, R.L.; Seixas, N.S.; Ren, K.K. A review of crane safety in the construction industry. Appl. Occup. Environ. Hyg. 2001, 16,
1106–1117. [CrossRef]
4. Raviv, G.; Shapira, A.; Fishbain, B. AHP-based analysis of the risk potential of safety incidents: Case study of cranes in the
construction industry. Saf. Sci. 2017, 91, 298–309. [CrossRef]
5. Milazzo, M.F.; Ancione, G.; Brkic, V.S.; Vališ, D. Investigation of crane operation safety by analysing main accident causes. In Risk,
Reliability and Safety: Innovating Theory and Practice; Taylor & Francis Group: Abingdon, UK, 2016; pp. 74–80.
6. Lucia, D.J.; Beran, P.S.; Silva, W.A. Reduced-order modeling: New approaches for computational physics. Prog. Aerosp. Sci. 2004,
40, 51–117. [CrossRef]
7. Bai, Z.; Dewilde, P.M.; Freund, R.W. Reduced-Order Modeling. Handb. Numer. Anal. 2005, 13, 825–895.
8. Polifke, W. Black-box system identification for reduced order model construction. Ann. Nucl. Energy 2014, 67, 109–128. [CrossRef]
9. Hijazi, S.; Stabile, G.; Mola, A.; Rozza, G. Data-driven POD-Galerkin reduced order model for turbulent flows. J. Comput. Phys.
2020, 416, 109513. [CrossRef]
10. Xiao, D.; Heaney, C.E.; Mottet, L.; Fang, F.; Lin, W.; Navon, I.M.; Guo, Y.; Matar, O.K.; Robins, A.G.; Pain, C.C. A reduced order
model for turbulent flows in the urban environment using machine learning. Build. Environ. 2019, 148, 323–337. [CrossRef]
11. Lucia, D.J.; Beran, P.S. Reduced-order model development using proper orthogonal decomposition and Volterra theory. AIAA J.
2004, 42, 1181–1190. [CrossRef]
12. Moi, T.; Cibicik, A.; Rølvåg, T. Digital twin based condition monitoring of a knuckle boom crane: An experimental study. Eng.
Fail. Anal. 2020, 112, 104517. [CrossRef]
13. Zhao, N.; Fu, Z.; Sun, Y.; Pu, X.; Luo, L. Digital-twin driven energy-efficient multi-crane scheduling and crane number selection
in workshops. J. Clean. Prod. 2022, 336, 130175. [CrossRef]
14. Dong, Q.; He, B.; Qi, Q.; Xu, G. Real-time prediction method of fatigue life of bridge crane structure based on digital twin. Fatigue
Fract. Eng. Mater. Struct. 2021, 44, 2280–2306. [CrossRef]
15. Moslått, G.A.; Padovani, D.; Hansen, M.R. A digital twin for lift planning with offshore heave compensated cranes. J. Offshore
Mech. Arct. Eng. 2021, 143, 031402. [CrossRef]
16. Kapteyn, M.G.; Knezevic, D.J.; Huynh, D.B.P.; Tran, M.; Willcox, K.E. Data-driven physics-based digital twins via a library of
component-based reduced-order models. Int. J. Numer. Methods Eng. 2022, 123, 2986–3003. [CrossRef]
17. Szpytko, J.; Duarte, Y.S. Digital twins model for cranes operating in container terminal. IFAC-Pap. 2019, 52, 25–30. [CrossRef]
18. Dong, H.; Wang, H.; Jiang, G.; Cai, Z.; Wang, W.; Liu, Y. An adaptive partitioned reduced order model of peridynamics for
efficient static fracture simulation. Eng. Anal. Bound. Elem. 2023, 157, 191–206. [CrossRef]
19. Khamlich, M.; Stabile, G.; Rozza, G.; Környei, L.; Horváth, Z. A physics-based reduced order model for urban air pollution
prediction. arXiv 2023, arXiv:2305.04575. [CrossRef]
Sensors 2023, 23, 9203 13 of 13

20. Aversano, G.; Ferrarotti, M.; Parente, A. Digital twin of a combustion furnace operating in flameless conditions: ROM development
from CFD simulations. Proc. Combust. Inst. 2021, 38, 5373–5381. [CrossRef]
21. Lai, X.; Wang, S.; Guo, Z.; Zhang, C.; Sun, W.; Song, X. Designing a shape–performance integrated digital twin based on multiple
models and dynamic data: A boom crane example. J. Mech. Des. 2021, 143, 071703. [CrossRef]
22. Augustine, P.; Hunter, T.; Sievers, N.; Guo, X. Load Identification of a Suspension Assembly Using True-Load Self Transducer Generation;
No. 2016-01-0429. SAE Technical Paper; SAE International: Warrendale, PA, USA, 2016.
23. Slowinski, D. Front-End Loader Linkage Durability Analysis Using Load Input from True-Load; Wolf Star Technologies: Milwaukee, WI,
USA, 2017.
24. Li, Y.; Ye, M.; Wang, Q.; Wei, M.; Lian, G. State of charge estimation of lithium-ion batteries using improved BP neural network
and filtering techniques. J. Phys. Conf. Ser. 2023, 2591, 012052. [CrossRef]
25. Barman, S.; Bandyopadhyaya, R. Modelling crash severity outcomes for low speed urban roads using back propagation–Artificial
neural network (BP–ANN)—A case study in Indian context. IATSS Res. 2023, 47, 382–400. [CrossRef]
26. Huang, Z.R.; Ge, M.; Pang, X.R.; Song, P.; Wang, C. The spatial distribution of interleukin-4 (IL-4) reference values in China based
on a back propagation (BP) neural network. Geospat. Health 2023, 18. [CrossRef] [PubMed]
27. Li, C.; Jia, T.; Han, X.; Jiang, X. Study on parameter optimization of laser cladding Fe60 based on GA-BP neural network. J. Adhes.
Sci. Technol. 2023, 37, 2556–2586. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.