You are on page 1of 13

Engineering Applications of Artificial Intelligence 117 (2023) 105609

Contents lists available at ScienceDirect

Engineering Applications of Artificial Intelligence


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

Design method for polyurethane-modified asphalt by using Kriging-Particle


Swarm Optimization algorithm
Pengzhen Lu ∗, Kai Ye, Tian Jin, Yiheng Ma, Simin Huang, Chenhao Zhou
Zhejiang University of Technology, Hangzhou 310014, Zhejiang Province, China

ARTICLE INFO ABSTRACT


Keywords: The preparation process of polyurethane (PU)-modified bitumen involves numerous design parameters and
Polyurethane-modified bitumen performance response indexes. Due to the variety of polyurethane modifiers, the preparation process of the
Preparation process polyurethane-modified bitumen is not universally applicable. However, the traditional methods such as the
Machine-learning algorithm
response surface method and orthogonal design method have some problems such as low accuracy and a large
Kriging-PSO model
number of samples required in the preparation process design. Therefore, according to different application
Design method
environments, the problem of determining the process parameters of the polyurethane-modified bitumen
accurately and efficiently needs to be solved urgently. Using Kriging-Particle Swarm Optimization (PSO)
algorithm, an efficient process design method for the preparation of polyurethane modified asphalt is proposed
in this paper. Combined with the sensitivity analysis method, the relatively sensitive response indexes are
screened out to reduce the number of samples and improve the design accuracy. Among them, the dispersion
coefficient was evaluated by fluorescence microscopy test using the Christiansen coefficient method to evaluate
the uniformity of the dispersed phase of the polyurethane modifier. According to the target performance, the
main process parameters of PU modified asphalt were obtained by the Kriging-Particle Swarm Optimization
algorithm: shear time 86 min, shear speed 2450 rpm, shear temperature 148 ◦ C, and polyurethane content
18.6%. The polyurethane-modified bitumen prepared by this optimal process met the expected performance
indicators. This study achieved the expected results with a small number of samples, indicating that this method
can achieve the purpose of designing the ideal process parameters of polyurethane-modified asphalt efficiently.

1. Introduction et al., 2018). Also, the PU prepolymer and bitumen undergo a curing
reaction under the conditions of high temperature and the addition of
With vigorous development of highway traffic construction in a chain-extending crosslinking agent, forming a cross-linked network
China, the frequency of rutting, cracking, and net cracking of asphalt structure, and becoming a uniform and stable new polymer-modified
pavement has increased, which means higher requirements for bitumen bitumen (Shirzad et al., 2020). This can greatly improve the elastic
materials. At present, polymers such as styrene-butadiene-styrene properties of the basis bitumen so that the asphalt pavement has a
(SBS) (Kumar et al., 2020), styrene-butadiene-rubber (Ren et al., 2018) strong resistance to plastic deformation, under factors such as the load
(SBR) and rubber powder (Guo et al., 2020) are widely used in bitumen of vehicle and environment (Li et al., 2021). Carrera et al. (2009)
modifiers. Among them, thermoplastic polymer represented by SBS study the influence of PU prepolymer types on the rheological proper-
is the most widely used bitumen modifier with the best application ties of the PU-modified bitumen by comparing the dynamic viscosity
effect now (Chen et al., 2021a). However, many studies have shown with different molecular weights at 60 ◦ C. Cuadri et al. (2014) use
no obvious chemical reaction between SBS and bitumen. SBS is uni- biological oil instead of polyols to react with isocyanate to prepare PU
formly dispersed and adsorbed in bitumen, which shows merely the for bitumen modification. Izquierdo et al. (2012) study the rheolog-
coexistence and integration in the physical significances belonging ical properties and microstructure of a foamed PU-modified bitumen
to the thermodynamic incompatible system. Polymer segregation or obtained by water blown. Notice that some achievements have been
degradation will occur during production, thermal storage, and use of made in the research of the PU-modified bitumen. As a new material,
SBS-modified bitumen (Chen et al., 2021b). researchers have recognized the PU-modified bitumen because of its
PU elastomer has the advantages of wear resistance, aging resis- superior performance, which proves that it has broad development
tance, high strength, and good low temperature flexibility (Bazmara prospects.

∗ Corresponding author.
E-mail addresses: pengzhenlu@163.com (P. Lu), ye212180899@163.com (K. Ye), 754897129@qq.com (T. Jin), 714799052@qq.com (Y. Ma),
809901366@qq.com (S. Huang), 707584836@qq.com (C. Zhou).

https://doi.org/10.1016/j.engappai.2022.105609
Received 22 May 2022; Received in revised form 12 October 2022; Accepted 3 November 2022
Available online 19 November 2022
0952-1976/© 2022 Elsevier Ltd. All rights reserved.
P. Lu, K. Ye, T. Jin et al. Engineering Applications of Artificial Intelligence 117 (2023) 105609

Now, determining the modification process parameters often in- optimum in Kriging model. Using the Kriging-PSO algorithm, the re-
volves an ODM (Ren et al., 2019) or a RSM (Lv et al., 2020). Although sponse model of PU modified asphalt with good fitting accuracy can
the ODM needs few test samples, the determined process parameters be established with less sample number, so as to solve the problem of
can involve relatively vague range values, whereas the RSM can give the design of process parameters for PU modified asphalt. Furthermore,
clear process parameters; however, too many test samples are needed. use the Kriging-PSO algorithm to design reasonable process param-
When there are 4 kinds of process parameters to be determined, 30 eters, thus providing a new research idea for the development of a
groups of qualified samples are needed. It far exceeds the sample size high-performance PU-modified bitumen.
required by the orthogonal design method. Therefore, it is a hot topic in
the field of modified asphalt to find an efficient and convenient method 2. Experimental flow and design
to design the preparation process of Pu modified asphalt.
In recent years, with the rise of artificial intelligence and the de-
In this study, the process design of the PU-modified bitumen was
velopment of large-scale computers, machine learning (ML) methods
carried out based on the Kriging-PSO algorithm. The process was
with high efficiency – with high speed and high accuracy – have
as follows: Used the shearing time, PU content, shearing speed, and
shown advantages in the field of material research and development,
shearing temperature as the four main process parameters. Conducted
especially in the aspects of material performance prediction and new
the experiments with ODM in four factors and four levels, with a total
material discovery. Many research achievements have been made in ML
of 16 groups of samples. Selected a softening point, penetration at
methods (see Table 1). Isayev et al. (2017) use the Gradient Increasing
25 ◦ C, ductility at 5 ◦ C, rutting factor at 64 ◦ C, Brookfield viscosity at
Decision Tree (GBDT) algorithm to model and predict the seven target
135 ◦ C, and a dispersion coefficient as preliminary response indicators
values of compound library AFLOW, with the prediction accuracy
to comprehensively evaluate the samples. Used range and variance
above 0.90. Gujar and Vakharia (2019) use ML methods as an artificial
analysis to analyze the sensitivity of each variable parameter to the
neural network and an order-preserving regression model to design
response indicators. Selected the response index with relatively high
and predict the composition of mineral fillers in the mixture. Wang
sensitivity to improve the fitting accuracy of the trained model. Per-
et al. (2018) use three ML methods – including Support Vector Machine
formed unified quantitative analysis on the response indicators of each
(SVM), K-NearestNeighbor (KNN) and Kernel Ridge Regression (KRR)
sample. For this to correspond to the variable parameters in the design
– to carry out high-precision calculation by modeling based on 4000
sample, constructed sample set D and determined the specific form
material performance samples of inorganic solid compounds. Therefore,
of the regression model and related models of the Kriging algorithm.
machine-learning–related algorithms have powerful functions, espe-
Trained the basic models by using the sample set D. Subsequently,
cially for the optimization of material preparation. However, many
used the PSO algorithm to search the optimal parameter 𝜃 by taking
shortcomings and urgent problems still need to be solved in the field
of artificial intelligence to guide the preliminary work of material the regression error of training samples as the termination condition
preparation and to select a suitable machine-learning algorithm for of the optimization process. At this time, the completed Kriging model
material modification process design. could be considered as the response model of process parameters and
A Kriging agent model is one of the most representative agent model sample indicators of the PU-modified bitumen. Determined the ideal
algorithm in ML methods, which is based on variogram and structure sample indicators of the PU-modified bitumen, inputted them into the
analysis (Oliver and Webster, 2014). On the premise of high correlation Kriging fitting model, and determined the regression result by the
of samples, the Kriging model has the advantage of high precision and response function. Among them, the output results were the optimal
has been widely used in many fields. However, in the optimization process parameters determined by the response model. Prepared the
process of Kriging model, the fitting accuracy of the model is deter- PU-modified bitumen based on the process parameters and evaluated
mined by the relevant parameters of variogram. In recent years many the accuracy of its response to the expected indicators through ideal
researchers have introduced PSO algorithm (Ulaganathan et al., 2015) experimental indicators. To understand the process design method of
into the Kriging model to solve the optimization problem of relevant the PU-modified bitumen based on the Kriging-PSO algorithm, Fig. 1
parameters: By introducing the Kriging-PSO algorithm into structural shows the experimental design and specific processes.
reliability research, Liang et al. (2010) point out that its result is better
than the classical RSM. Lu et al. (2020) use the Kriging-PSO algorithm 2.1. Raw material performance indicators
to achieve ideal results when studying the correlation between dynamic
characteristics and static elements of bridge structures. Ma et al. (2013) The preparation materials of the PU-modified bitumen mainly in-
use the Kriging-PSO algorithm to approximate the hypersonic vehicle cluded the basis bitumen, PU prepolymer, and chain extender.
and to obtained high-precision aerodynamic data, which proved that
the PSO algorithm can perform the efficient search of related parame- 2.1.1. The basis bitumen
ters. Yi et al. (2015) point that Kriging-PSO performs much better than East China Sea AH-70# served as the basis bitumen used in this
Kriging when it exceeds the sample point range. The above researches research. Table 2 shows the basic performance indicators.
show that the PSO algorithm can perform the efficient search work of
relevant parameters well; as a result, that the Kriging-PSO algorithm
2.1.2. The polyurethane modifier
can effectively deal with multi-dimensional variable parameters. Also,
The PU used in this study consisted of a PU prepolymer and a chain
because the Kriging-PSO algorithm has the characteristics of high effi-
extender produced. The chain extender can react with the functional
ciency and high precision, it can solve the problem of small samples
groups of the polyurethane prepolymer to expand its molecular chain
and high dimensional data. For the preparation process of the PU-
modified bitumen, multiple process parameters need to be considered and increase its molecular weight. According to the manufacturer’s rec-
simultaneously. Using ODM or RSM for experimental design causes ommendation, the amount of chain extender is 15% of the polyurethane
some problems, such as large sample numbers or a difficult quantitative prepolymer. Tables 3 and 4 show the performance indexes of the two.
determination of process parameters.
In this paper, Kriging-PSO algorithm is used for process design. The 2.2. Kriging-Particle Swarm Optimization algorithm design
Kriging model has the advantage of high accuracy compared with the
polynomial model and the radial basis function model. At the same Kriging-PSO algorithm include Kriging model and PSO algorithm.
time, thanks for the multi-point search ability, PSO can effectively In the algorithm, Kriging is used to establish response surface, and PSO
improve the ability of correlation parameter 𝜃 to jump out of local is used to find the optimal parameters of Kriging model.

2
P. Lu, K. Ye, T. Jin et al. Engineering Applications of Artificial Intelligence 117 (2023) 105609

Table 1
Summary of the existing work for ML and PSO.
Author’s Methodology Measures PROS CONS
Tian et al. Multiobjective Hybrid Differential-Evolution Reduce the operational cost and improve Simple, reliable, advanced Lack discussion
(2016) Particle-Swarm-Optimization (MHDP) the efficiency of extinguishing fire
algorithm services
Dong and Zhou Adaptive particle swarm optimization with Enhance PSO’s convergence performance Simple to understand and Not mature, narrow
(2017) supervised learning and control (APSO-SLC) program range of application
Khatir et al. Experimentally measured natural frequencies Detect and locate open cracks in Simple and good Not mature
(2018) and the Particle Swarm Optimization (PSO) beam-like structures application prospect
method
Tiachacht et al. Modified Cornwell Indicator (MCI) and Estimate the severity of structural Small consumption and Lack of experimental
(2018) genetic algorithm (GA) damage quick assessment validation, not mature
Vu-Huu et al. A polytree-based adaptive polygonal finite Find an optimal shape of submerged Reliable results, accurate Not widely applied
(2018) element method for topology optimization of breakwater
fluid-submerged breakwater
interaction
Lv et al. (2019) Surrogate-assisted particle swarm optimization The Pareto optimal solutions obtained Fast and simple Lack validation
algorithm with Pareto active learning by the proposed algorithm has better
convergence and distribution
Khatir and An inverse algorithm based on Proper Identify single and multiple cracks in Fast and stable Limited application
Abdel Wahab Orthogonal Decomposition (POD) and Radial plate structures scenarios
(2019) Basis Functions (RBF)
Khatir et al. Two-dimensional Isogeometric Analysis (IGA) Do damage assessment in beam-like Good application prospect Need more test subjects
(2019) and Finite Element Method (FEM) combined structures and enough contrast
with optimization techniques.
Tran-Ngoc et al. ANN combined with CS (ANN-CS) Determine the location and severity of Rigorous comparison Lack of experimentation
(2019) structural damage
Cao et al. (2019) A comprehensive learning particle swarm Improve PSO’s convergence rate and Considers various factors, Theory is not mature
optimizer (CLPSO) embedded with local accuracy evaluation results are
search (LS). reliable
Khatir et al. Improved ANN technique combined with Jaya Used for structural crack identification High efficiency and good High requirements, not
(2020) algorithm application prospect widely applied

Table 2
AH-70# basis bitumen technical indicators.
Penetration Softening Ductility Flash Solubility/% RTFOT residue
(25 ◦ C)/0.1 mm point/◦ C (15 ◦ C)/cm point/◦ C
Quality 25 ◦ C Residual 10 ◦ C Residual
change/% penetration ratio/% ductility/cm
Detection result 66 52.3 >100 279 99.9 −0.403 70 14
Technical requirements 60∼80 ≥46 ≥100 ≥260 ≥99.5 ≤±0.8 ≥61 ≥6

Table 3
Performance index of PU prepolymer.
Appearance (25 ◦ C) Viscosity (25 ◦ C) NCO mass fraction Density
Unit NA mPa s % g cm−3
Test results Colorless and transparent 1000∼1200 10.2∼11.2 1.05∼1.10
Executive standard Visual inspection GB/T 12009.3 GB/T 12009.4 GB/T 4472

Table 4 of polynomial and random distribution, namely (Jerome et al., 1989):


Technical indicators of chain extender.
Appearance Molecular Melting Chlorine Amine 𝑦(𝑥) = 𝑓 (𝑥)𝑇 𝛽 + 𝑍(𝑥) (1)
(25 ◦ C) weight point content value 𝑇
𝑓 (𝑥) 𝛽 = [𝑓1 (𝑥), 𝑓2 (𝑥), … , 𝑓𝑝 (𝑥)]𝛽 = 𝑓1 (𝑥)𝛽1 + 𝑓2 (𝑥)𝛽2 + ⋯ + 𝑓𝑝 (𝑥)𝛽𝑝 (2)
Unit NA NA ◦C NA mmol g−1
Test results Light yellow 367.16 100∼109 26 7.5 In these equations, 𝑓 (𝑥) is a polynomial model that can choose zero-
order, first-order, and second-order polynomials according to needs; 𝑝
is the number of polynomials; 𝛽 is the regression coefficient; and 𝑍(𝑥)
is a random process called a variogram or a correlation model.
2.2.1. Kriging model The polynomial part of the Kriging model can be regarded as the
The Kriging model is also called spatial local estimation and spatial parametric part of the whole model, also called a regression model, with
local interpolation (Oliver and Webster, 2014). The establishment of its function to provide global approximation for the data model (Zhao
the Kriging model is based on variation function theory and structural et al., 2011). The existence of 𝑍(𝑥) makes the Kriging model reflect
analysis. The main point of this model is to use the original data and the nonlinear state of structure input–output relationship, as well as
solves the problem of accuracy of response surface based on polynomial
variance function of the regionalized variable to make linear unbiased
function. The covariance matrix of 𝑍(𝑥) is as follows:
optimal estimation of the value of the regionalized variable at the
point to be estimated (Wu et al., 2018). The Kriging model consists 𝐶𝑜𝑣(𝑍(𝑥𝑖 )𝑍(𝑥𝑗 )) = 𝜎 2 𝑅(𝜃, 𝑥𝑖 , 𝑥𝑗 ) (3)

3
P. Lu, K. Ye, T. Jin et al. Engineering Applications of Artificial Intelligence 117 (2023) 105609

Fig. 1. Flowchart of technology design method of polyurethane-modified bitumen.

In Eq. (3), 𝑅(𝜃, 𝑥𝑖 , 𝑥𝑗 ) is the spatial correlation function of any two where 𝑥𝑖𝑚 is the mth component of the vector 𝑥𝑖 . Eq. (4) has many
sample points (𝑥𝑖 and 𝑥𝑗 ) in the sample points, which plays a vital options for the kernel function used in Eq. (4) 𝑅(𝜃, 𝑥𝑖𝑚 − 𝑥𝑗𝑚 ). The
role in the precision of model fitting; whereas 𝜃 is the parameter of commonly used kernel functions mainly include exponential function,
the correlation function, whose function form can be expressed as Gaussian function, and linear function.
( ) ∏𝑛
Among the above kernel functions, Gaussian function can well meet
𝑅 𝜃, 𝑥𝑖 , 𝑥𝑗 = 𝑅(𝜃, 𝑥𝑖𝑚 − 𝑥𝑗𝑚 ) (4)
the learning requirements of complex models, and its calculation effect
𝑚=1

4
P. Lu, K. Ye, T. Jin et al. Engineering Applications of Artificial Intelligence 117 (2023) 105609

is also ideal (Jerome et al., 1989). Therefore, Gaussian function is In this equation, regret is the regression model function handle; Corr is
selected as the kernel function of the related function in the current the handle of the related model function; S and Y are sample space
research: and corresponding response values; Theta0 is the initial value of a
| |2 parameter. The DACE toolbox uses a pattern search method to find,
𝑅(𝜃, 𝑥𝑖𝑚 − 𝑥𝑗𝑚 ) = exp(−𝜃𝑚 |𝑥𝑖𝑚 − 𝑥𝑗𝑚 | ) (5)
| | which requires a given initial point theta0. Theta0 has a great influence
where 𝜃𝑚 is the relevant parameters of the model. The commonly on the accuracy of the Kriging model. Due to the improper value
used Kriging regression models are also divided into constant type, of Theta0, the parameter is often trapped in the local optimum, to
linear type, and quadratic type. After modeling comparison, the linear reduce the accuracy of the Kriging model. Therefore, a new function
regression results fit well with the sample data. So in this study, we dacefitPSO is developed in this paper to establish the Kriging-PSO
used the linear type as the Kriging regression function model in the model. The specific form of the function is as follows:
current research:
[𝑑𝑚𝑜𝑑𝑒𝑙, 𝑝𝑒𝑟𝑓 ] = 𝑑𝑎𝑐𝑒𝑓 𝑖𝑡𝑃 𝑆𝑂(𝑆, 𝑌 , 𝑟𝑒𝑔𝑟, 𝑐𝑜𝑟𝑟) (12)
𝑓1 (𝑥) = 1, 𝑓2 (𝑥) = 𝑥1 , … , 𝑓𝑛+1 (𝑥) = 𝑥𝑛 (𝑝 = 𝑛 + 1) (6)
where PSO is used to find the optimal parameters of 𝜃. Using the multi-
In the prediction process, it can be equivalent to a class of minimum point search capability of PSO is expected to overcome the limitations
optimization problems (Ulaganathan et al., 2015): of single-point search of pattern search methods and the heavy reliance
{ 1 } on the initial guess solution theta0 (Yi et al., 2015).
𝜃 = 𝑚𝑖𝑛 |𝑅| 𝑚 𝜎̂𝑧2 (7)
𝜃>0
2.3. Performance evaluation
By solving the minimum optimization problem of Eq. (7), we could
obtain the corresponding parameter 𝜃 and establish the optimal Kriging
In this study, we determined the softening point, penetration
prediction model. In this process, parameter 𝜃 needed to be optimized.
(25 ◦ C), ductility (5 ◦ C), rutting factor (64 ◦ C), Brookfield viscosity
In this study, we selected PSO algorithm (Poli R and Blackwell, 1995,
(135 ◦ C), and dispersion coefficient to evaluate the performance of the
2007) as the optimization method for parameter 𝜃.
samples.

2.2.2. Particle Swarm Optimization algorithm


2.3.1. Fluorescence microscope test
PSO is an evolutionary computing technology, first proposed by
A key factor for determining the PU-modified bitumen’s perfor-
Kennedy and Eberhart in 1995 (Ulaganathan et al., 2015). The ba-
mance is the thermodynamic compatibility between the bitumen and
sic idea of PSO comes from the inspiration of modeling and simu-
the PU modifier (Fang and Tu, 2019). From a broad engineering level,
lation results of bird group behavior. In the PSO algorithm, a par-
compatibility consists of the formation of a stable system after the
ticle represents each bird. It constantly searches near the particle
mixing of two substances, a PU modifier evenly distributed in the
nearest to the food, and finally obtains the position of the food and
bitumen, and a storage process without stratification or phase sepa-
the particle nearest to the food. The optimization of the PSO algo-
ration. Therefore, the uniformity of the PU modifier dispersion phase
rithm needs continuous evolution to update the three control indexes
is important for evaluating the compatibility of modified bitumen. At
of the particle, and through continuous comparison, the optimal so-
present, studies on the uniformity of polymer-dispersed phase distri-
lution is finally reached (Ulaganathan et al., 2015). Suppose that
bution are mainly based on subjective observation method and grid
some particles in a D-dimensional space have their positions at 𝑥𝑖 =
( )𝑇 deviation method (Zou et al., 2021). The fluorescence microscopic (FM)
𝑥𝑖1 , 𝑥𝑖2 , … , 𝑥𝑖𝑑 , … , 𝑥𝑖𝐷 (𝑖 = 1, 2, … , 𝑚; 𝑑 = 1, 2, … , 𝐷). The velocity of
( )𝑇 analysis shows that the modified bitumen produces a large number
the particle is 𝑣𝑖 = 𝑣𝑖1 , 𝑣𝑖2 , … , 𝑣𝑖𝑑 , … , 𝑣𝑖𝐷 and its corresponding
of tiny particles, which can be hard to recognized with naked eyes
fitness value is 𝑓𝑖 . The individual extreme value of the i particle is
( )𝑇 during the process of high-speed shear. Quantitatively and accurately
𝑝𝑖 = 𝑝𝑖1 , 𝑝𝑖2 , … , 𝑝𝑖𝑑 , … , 𝑝𝑖𝐷 , with the population extreme value of the
( )𝑇 evaluating the dispersion uniformity through subjective observation is
whole population at 𝑝𝑔 = 𝑝𝑔1 , 𝑝𝑔2 , … , 𝑝𝑔𝑑 , … , 𝑝𝑔𝐷 . Through continu- difficult; as a result, statistical methods more suitable for the evaluation
ous evolution, the PSO algorithm constantly updates the particle speed of dispersion phase uniformity of the PU-modified bitumen should
and position. The following equations express the speed and position be selected. Therefore, we used the Christiansen coefficient method
in D-dimension (𝑑 = 1, 2, … , 𝐷) up to the t-generation: to evaluate the dispersed phase uniformity of PU-modified particles
( ) ( ) through FM test. In applied statistics, the Christiansen coefficient is as
𝑣𝑡𝑖 = 𝑤𝑣𝑡−1𝑖 + 𝑐1 𝑟1 () 𝑝𝑖 − 𝑥𝑡−1
𝑖 + 𝑐2 𝑟2 () 𝑝𝑔 − 𝑥𝑡−1
𝑖 (8)
{ follows:
𝑚𝑎𝑥 𝑡
𝑣 , 𝑣𝑖𝑑 > 𝑣 𝑚𝑎𝑥
𝑣𝑡𝑖𝑑 = (9) 𝛥𝜃
𝑣𝑚𝑖𝑛 , 𝑣𝑡𝑖𝑑 < 𝑣𝑚𝑖𝑛 𝐶𝑢 = 1 − (13)
𝜃
𝑥𝑡𝑖𝑑 = 𝑥𝑖𝑑
𝑡−1
+ 𝑣𝑡𝑖𝑑 𝛥𝑡 (10) ∑𝑁 | |
𝑖=1 ||𝜃𝑖 − 𝜃 ||
In these equations, 𝑣𝑡𝑖
is that velocity of the ith particle; is the 𝑥𝑡𝑖 𝛥𝜃 = (14)
𝑁
position of the ith particle; 𝑣𝑡𝑖𝑑 is that velocity of the d-dimensional
In Eqs. (11) and (12), 𝐶𝑢 is coefficient of dispersion; 𝜃 is particle
of the ith particle; 𝑥𝑡𝑖𝑑 is the position of the d-dimensional of the ith
number per unit area; 𝛥𝜃 is the average value of the absolute value
particle; 𝑤 is the inertial weight; 𝑐1 and 𝑐2 are acceleration factors,
of the difference between the number of particles per unit area and
representing the amount of change in the velocity of the particles.
the number of particles per unit area in each area; 𝜃𝑖 is the number
Moving to 𝑝𝑖 and 𝑝𝑔 ; 𝑟1 () and 𝑟2 () are random numbers in the range
of particles per unit area; and 𝑁 is the number of regions (when the
of 0 to 1; 𝑣𝑚𝑎𝑥 is the maximum value of speed; 𝑣𝑚𝑖𝑛 is the minimum
number of regions is greater than 16, the results tend to be stable).
value of speed; and 𝛥𝑡 is the time interval.

2.3.2. Dynamic shear rheological test


2.2.3. Kriging-Particle Swarm Optimization algorithm
In this study, the DSR (dynamic shear rheological) test adopted the
In this paper, DACE toolbox prepared by Lophaven et al. (2002) was
method of temperature scanning. Set the test temperature from 29 ◦ C to
used to establish Kriging model. In DACE, Dacefit function is used to
79 ◦ C and the temperature interval to 5 ◦ C. Adopted the strain control
establish the basic model of Kriging. The specific form of the function
mode. The strain was 12% and the frequency was 10 rad/s. After the
is as follows:
test, obtained the rutting factor (𝐺∗ ∕𝑠𝑖𝑛𝛿) to evaluate the rheological
[𝑑𝑚𝑜𝑑𝑒𝑙, 𝑝𝑒𝑟𝑓 ] = 𝑑𝑎𝑐𝑒𝑓 𝑖𝑡(𝑆, 𝑌 , 𝑟𝑒𝑔𝑟, 𝑐𝑜𝑟𝑟, 𝑡ℎ𝑒𝑡𝑎0) (11) properties of the PU-modified bitumen.

5
P. Lu, K. Ye, T. Jin et al. Engineering Applications of Artificial Intelligence 117 (2023) 105609

Table 5 bitumen samples to the maximum extent and have certain representa-
ODM of PU-modified bitumen.
tiveness. Therefore, for this study, preliminarily proposed six indicators
Technological parameter Level 1 Level 2 Level 3 Level 4 – namely, softening point, penetration (25 ◦ C), ductility (5 ◦ C), rutting
Shearing time/min 40 60 80 100 factor (64 ◦ C), Brookfield viscosity (135 ◦ C), and dispersion coefficient
PU content/% 5 10 15 20
– to evaluate the prepared samples. Among them, the FM was from
Shearing speed/rpm 1500 2000 2500 3000
Shearing temperature/◦ C 125 135 145 155
NIKON (Japan) with the model as NI-U, whereas the DSR came from
Anton Paar (Austria) with the model as MCR302. Fig. 2 shows the test
instruments and equipment.
Table 6
Orthogonal design of PU-modified bitumen samples.
3.1.1. Dispersion uniformity
No. Shearing PU Shearing Shearing
time/min content/% speed/rpm temperature/◦ C
Dispersion coefficient is used to evaluate the compatibility between
polymer and asphalt. The larger the dispersion coefficient, the better
1 40 5 1500 125
2 40 10 2500 145
the dispersion uniformity of polymer in asphalt. To characterize the
3 40 15 2000 135 distribution uniformity of PU in the bitumen, the study carried out
4 40 20 3000 155 FM test on samples to collect microscopic images of the PU-modified
5 60 5 2000 145 bitumen. We prepared 16 groups of PU-modified bitumen samples, ac-
6 60 10 3000 125
cording to the process parameters shown earlier in Table 6. We placed
7 60 15 1500 155
8 60 20 2500 135 a small amount of the PU-modified bitumen at normal temperature on
9 80 5 2500 155 a slide and heated that amount until bitumen flowed. Then we spread
10 80 10 1500 135 bitumen evenly, put it under covered glass, and then placed it under a
11 80 15 3000 145 microscope after cooling. We prepared 16 samples, with each sample
12 80 20 2000 125
13 100 5 3000 135
containing two small pieces of the PU-modified bitumen, as shown in
14 100 10 2000 155 Fig. 3.
15 100 15 2500 125 This study used the image processing function of the MATLAB soft-
16 100 20 1500 145 ware to eliminate noise interference in the process of image acquisition,
enhance the contrast between dispersed phase and background, and
improve the accuracy of feature extraction of dispersed phase. First, we
2.4. Test scheme design used the K-means algorithm to filter and denoise the acquired images.
The purpose of filtering and denoising was twofold:
Because of too many technological parameters for preparation of 1. To extract the characteristics of the PU modifier as the feature
the PU-modified bitumen (considering the convenience of control and mode of image recognition
the degree of influence on sample performance), excluded from this 2. To meet the requirements of computer processing and to elimi-
study the secondary process parameters such as one-time preparation nate the noise mixed with the PU-modified bitumen microscopic
amount, diameter of disc saw-tooth dispersion disc, and container image digitalization
size. For the research objects, selected four main process parameters:
shearing time, PU content, shearing speed, and shearing temperature. Then, we enhanced the gray scale of the image to improve the quality
To control the oil bath temperature to make the shear temperature and recognizability of the microscopic image. Finally, we used the
meet the test requirements, placed the basis bitumen (300 g) in an oven OTSU algorithm for binarization pretreatment of the microscopic im-
and heated it to 120∼130 ◦ C to soften it. Used a high-speed shearing age; that way, the gray value of the pixel of the micrograph was 0
machine to shear and stir the basis bitumen; the shearing speed should or 255. The whole image presented an obvious black and white effect
meet the test requirements. Slowly added the melted MOCA heated to (Chen et al., 2010).
120 ◦ C and PU prepolymer preheated to 90 ◦ C to the asphalt with shear To make the result region stable, we cut the pre-processed micro-
agitation. Finally, placed the evenly mixed PU-modified bitumen in an graphs into 5 × 5 regions on average—that is, the number of regions
100 ◦ C oven for 2 h. According to relevant literatures and engineering 𝑁 = 25. Thus, we calculated the Christiansen coefficient (𝐶𝑢 ), with the
experience (Bazmara et al., 2018), Table 5 shows the values of process calculation results shown later in Table 7. To evaluate the correctness
parameters at various levels. At the same time, Table 6 shows the of 𝐶𝑢 , we used samples 3 and 12 as examples; Fig. 4 shows the
micrographs collected by FM test. The 𝐶𝑢 of the two samples were
sample design obtained by ODM with four levels and four factors.
0.311 and 0.893 respectively, which well reflected the dispersed phase
uniformity of particles in the micrograph.
3. Polyurethane-modified bitumen process design and analysis
3.1.2. Rheological performance
Tested the performance indexes of the samples. And determined the The rutting factor (𝐺∗ ∕𝑠𝑖𝑛𝛿, 𝑅𝐹 ) – a technical index to evaluate the
final corresponding indexes by range and variance analyses and sig- rheological properties of bitumen – is greater and stronger than the
nificance analysis. Then established the Kriging-PSO model according deformation resistance of bitumen. Fig. 6 shows that the 𝑅𝐹 of the
to the corresponding indicators. Designed the optimum process pa- PU-modified bitumen gradually decreased with the increase of tempera-
rameters according to the ideal performance index. Then prepared PU ture. However, note that 𝑅𝐹 of the PU-modified bitumen varied greatly
modified asphalt according to the optimum parameters and tested its under different processes. Sample 4 (𝑆4)’s shearing time was 40 min,
performance. At last, evaluated the model performance by comparing PU content was 20%, shearing speed was 3000 rpm, and shearing
with the ideal performance index. temperature was 155 ◦ C. When the DSR test temperature was 29 ◦ C, the
𝑅𝐹 of 𝑆4 far exceeded that of other samples, reaching a maximum of
3.1. Polyurethane-modified bitumen tests 1.25 × 106 because the combination of the PU modifier and the bitumen
interface formed a spatial network structure, which improved the high-
Prepared the PU-modified bitumen samples according to the pro- temperature performance of the PU-modified bitumen. Also, the process
cess parameters by ODM and obtained the response indicators of the of sample 1 (𝑆1) resulted in shearing time at 40 min, PU content at 5%,
finished products through relevant detection tests. The selected re- shearing speed at 1500 rpm, and shearing temperature at 125 ◦ C. Also,
sponse indicators should show the actual situation of PU-modified 𝑅𝐹 of 𝑆1 was the minimum value, 3.5 × 105 , when the test temperature

6
P. Lu, K. Ye, T. Jin et al. Engineering Applications of Artificial Intelligence 117 (2023) 105609

Fig. 2. The experimental instruments for determining response indicators.

Fig. 3. FM test samples (No. 1–16).

Fig. 4. Micrographs of samples 3 and 12.

7
P. Lu, K. Ye, T. Jin et al. Engineering Applications of Artificial Intelligence 117 (2023) 105609

Fig. 5. DSR test.

was 29 ◦ C. Comparing 𝑆4 and 𝑆1 showed that 𝑅𝐹 gently changed with Also, according to Table 9, the influence of various factors on
temperature after adding the PU modifier. The PU content, shearing softening point is as follows:
speed, and shearing temperature had great influence on the rheological
properties of the PU-modified bitumen. However, according to 𝑆4, • The PU content had the greatest influence, and the influence of
although the maximum 𝑅𝐹 value was obtained at 29 ◦ C, the anti- other three process factors were similar.
deformation ability of bitumen was sensitive to temperature changes in • The order of influence on the rutting factor from big to small was
the range of 29 ◦ C∼64 ◦ C; as a result, the preparation process needed B > C > D > A.
to be further determined (see Fig. 5). • The order of influence on Brookfield viscosity from big to small
was B > A > C > D, among which the influence of PU content far
3.1.3. Basic performance exceeded the other three factors.
The basic performance of asphalt includes softening point, pene- • The order of influence on dispersion coefficient from big to small
tration (25 ◦ C), ductility (5 ◦ C), and Brookfield viscosity (135 ◦ C). is A > C > D > B, and the four process factors had significant
Penetration can evaluate the viscosity of asphalt and its ability to influence on it.
resist deformation under external forces. The smaller the degree of
penetration, the stronger the stickiness and the stronger the ability 3.3. Response indicators analysis
to resist deformation. The softening point is to evaluate the high-
temperature stability of asphalt according to the temperature at which After significance analysis in Tables 8 and 9, used the softening
the asphalt softens. The higher the softening point, the better the high point, rutting factor, Brookfield viscosity, and dispersion coefficient
temperature stability, and vice versa. Ductility evaluates the plasticity of the PU-modified bitumen as response indicators in this study. Ac-
of asphalt according to the length of tensile failure of asphalt under cording to these four evaluation indicators, studied and analyzed the
external force. The larger the ductility, the better the plasticity of performance of the PU-modified bitumen. Fig. 6 shows the variation
asphalt, and vice versa. Brookfield viscosity can evaluate the fluidity of trend of various performance indicators of the PU-modified bitumen
asphalt. Asphalt is a viscoelastic material, which shows strong elasticity samples, with four factors.
in low-temperature environment, and its physical characteristics are Notice in Fig. 6 that with the increase of the shearing temperature
similar to solids; In high-temperature environment, it shows strong and shearing speed, the rutting factor and dispersion coefficient both
viscoplasticity, and its physical characteristics are similar to liquids. improved, but the softening point and Brookfield viscosity had little
Tested the basic performance of 𝑆1∼𝑆16. Meanwhile, statistically change. With the increase of PU content, the softening point, rutting
analyzed the dispersion coefficient and rutting factor (64 ◦ C). Table 7 factor, and Brookfield viscosity all increased. The effect increased
shows the results. Notice in Table 7 that the response indicators of significantly when the PU content exceed 15%, but the dispersion
S1∼S16 samples were quite different. To facilitate subsequent machine coefficient barely changed. With the increase of the shearing time,
learning training, performance response indicators should be further the softening point, rutting factor, and Brookfield viscosity did not
screened. change obviously, but the dispersion coefficient increased significantly.
The improvement effect slowed down after the shearing time exceeded
3.2. Response indicators determination 80 min.

In this study, used range and variance analyses to study the in- 3.4. Response model construction and calculation
fluence of the shearing time (A), PU content (B), shearing speed (C),
and shearing temperature (D) on the performance of the PU-modified According to scheme by ODM (in Table 6), response indicators (in
bitumen. Then analyzed their significance. At the same time, selected Table 7), and the sensitivity analysis results (in Tables 8 and 9), two
the significance level of 𝜎 = 0.05 to screen the preliminary response low correlation indicators – penetration and ductility – can be deleted,
indicators. Tables 8 and 9 show the calculation results and a four-variable/four-response training samples set containing 16
According to the data shown in Table 8, the four main process learning samples can be constructed. Fig. 7 shows the sample set
parameters screened had no obvious influence on penetration and space.
ductility of the PU-modified bitumen. Therefore, to facilitate the test According to this training sample set, we could construct the Krig-
operation, excluded these two indicators as evaluation indicators of the ing algorithm model through the deep learning toolbox in MATLAB
PU-modified bitumen in the follow-up studies. software. Determined the kernel function of the correlation model to

8
P. Lu, K. Ye, T. Jin et al. Engineering Applications of Artificial Intelligence 117 (2023) 105609

Table 7
Response indicators results of PU-modified bitumen samples.
No. Softening point/◦ C Penetration (25 ◦ C)/mm Ductility (5 ◦ C)/cm Rutting factor (64 ◦ C)/Pa Brookfield viscosity (135 ◦ C)/Pa s Coefficient of dispersion
1 84.0 3.79 40.7 6025 3809 0.286
2 80.9 3.42 20.2 14 454 60 775 0.343
3 71.7 2.96 21.2 13 710 32 695 0.311
4 91.8 2.53 11.2 45 134 85 800 0.813
5 80.5 3.66 25.1 10 938 5629 0.354
6 79.1 2.82 23.3 18 189 13 228 0.632
7 85.6 3.76 14.5 19 945 10 790 0.793
8 89.3 3.36 12.4 30 389 111 150 0.732
9 75.2 3.25 9.3 17 593 6344 0.931
10 74.7 2.97 19.9 13 575 14 008 0.922
11 81.6 3.13 8.6 26 093 10 985 0.684
12 85.3 2.66 8.9 25 950 92 950 0.893
13 78.3 3.35 11.2 15 331 10 205 0.911
14 80.5 2.71 9.6 27 310 11 089 0.821
15 81.1 2.78 18.0 24 867 14 600 0.938
16 86.5 3.54 11.2 28 113 89 700 0.863

Table 8
The ANOVA of performance test results of PU-modified bitumen.
Preliminary response indicators The process parameters
Shearing time PU content Shearing speed Shearing temperature
F 2.14 12.8 1.03 1.19
Softening point/◦ C
Significant parameters PU content

Penetration F 1 4.79 2.29 2.98


(25 ◦ C)/mm Significant parameters NA

Ductility F 3.08 2.07 1 1.84


(5 ◦ C)/cm Significant parameters NA
Variance
analysis Rutting factor F 1.42 26.12 3.27 6.24
(64 ◦ C)/Pa Significant parameters PU content

Brookfield viscosity F 5.12 188.45 6.01 6.5


(135 ◦ C)/(Pa s) Significant parameters PU content

Dispersion F 54.16 9.63 9.67 14.26


coefficient Significant parameters Shearing time, PU content, shearing speed, shearing temperature

F critical value 9.28

Table 9
Range analysis of performance test results of PU-modified bitumen.
Preliminary response indicators The process parameters
Shearing time (𝐴) PU content (𝐵) Shearing speed (𝐶) Shearing temperature (𝐷)
R 4.4 9.4 3.2 4.8
Softening point/◦ C
Order 𝐵>𝐷>𝐴>𝐶

Penetration R 0.4 0.53 0.56 0.43


(25 ◦ C)/mm Order 𝐶>𝐵>𝐷>𝐴

Ductility R 11.7 10.7 8 11.6


Range (5 ◦ C)/cm Order 𝐴>𝐷>𝐵>𝐶
analysis
Rutting factor R 4074.3 19 924.9 9272.1 9244.2
(64 ◦ C)/Pa Order 𝐵>𝐶>𝐷>𝐴

Brookfield viscosity R 14 698 88 403.3 13 508.8 4074.3


(135 ◦ C)/(Pa s) Order 𝐵>𝐴>𝐶>𝐷

Dispersion R 0.445 0.205 0.244 0.216


coefficient Order 𝐴>𝐶>𝐷>𝐵

be the Gaussian function to meet the learning requirements of this Notice in Fig. 8 that the model had high regression accuracy for the
multidimensional complex model and set the main parameters – such as training samples – that is, the trained Kriging model well reflected the
first-order polynomial correlation model and linear function regression implicit association between process parameter variables and sample
model – to construct the Kriging basic model. At the same time, set the indicators – so it could be considered a high-precision response model
initial value of parameter 𝜃 to 1 and carried out the spatial optimization between them.
search by using the PSO algorithm. The optimization stopped condition According to different practical application environments, the ideal
of parameter 𝜃 was that the regression error of each sample index in performance indicators of modified bitumen varied. Based on the ac-
the training sample set was less than 1%, until the optimization stop tual project requirements, Table 10 shows how the ideal performance
condition was met. Deemed the Kriging-fitting model as completed. indicators of the PU-modified bitumen were set. It was input into the
Fig. 8 shows the regression comparison results. Kriging fitting model, and we obtained the results by response function

9
P. Lu, K. Ye, T. Jin et al. Engineering Applications of Artificial Intelligence 117 (2023) 105609

Fig. 6. Change trend of various properties of modified bitumen at different levels.

Fig. 7. Sample set spatial diagram.

regression. Table 11 shows the final output results, regarded as the to meet the needs of practical engineering, the shearing time, shearing
optimal process parameters determined by the response model. In order speed and shearing temperature are integer values.

10
P. Lu, K. Ye, T. Jin et al. Engineering Applications of Artificial Intelligence 117 (2023) 105609

Fig. 8. Regression comparison diagram of training samples.

Table 10 Levenberg–Marquardt (LM) algorithm to study the deterioration law of


Index table of ideal sample of the PU-modified bitumen.
pavement performance after ultraviolet (UV) aging of asphalt mixture.
Softening Rutting factor Brookfield viscosity Dispersion
The maximum errors between the calculated and tested values are
point/◦ C (64 ◦ C)/Pa (135 ◦ C)/Pa s coefficient
0.35%, 7.64% (AC-13 asphalt mixture) and −0.81%, 11.45% (AC-16
90 30 000 80 000 0.92
asphalt mixture), respectively. Schneider et al. (2021) investigated
the rheological properties of asphalt using linear discriminant analysis
Table 11 (LDA) and random forest (RF) machine learning algorithms. The suc-
Output results of Kriging-fitting model. cess rates of LDA and RF were 87% and 83%, respectively. Considering
Shearing PU content/% Shearing Shearing the limited number of different bitumen samples tested, this can be
time/min speed/rpm temperature/◦ C
regarded as a very good result. Therefore, the fitting model constructed
86 18.6 2450 148
by the Kriging-PSO algorithm had good response function, and it was
feasible and efficient to design and guide the preparation of the PU-
modified bitumen by Kriging-PSO algorithm.
3.5. Response model verification
4. Conclusion
Prepared the PU-modified bitumen sample finished product ac-
cording to the optimal process parameters and obtained the response In this study, a Kriging PSO method was proposed for the prepara-
indicator of the sample finished product through the relevant perfor- tion process design of PU modified asphalt. Firstly, a small number of
mance testing test. Table 12 shows the comparison between the actual test samples was made by using the ODM, and the response indicators
sample response indicator and the ideal sample response index. with high correlation degree were selected by sensitivity analysis.
Notice in Table 12 that preparing the PU-modified bitumen based Then, based on the ideal response indicators, used the Kriging-PSO
on the output result of the Kriging fitting model is feasible, with algorithm for the fitting modeling and obtained the required optimal
its response indicators all close to the ideal response indicators. The process parameter solution. By using the process parameters, prepared
response indicator with the smallest error value was the dispersion the PU-modified bitumen and verified the error value between its
coefficient with an error of only 1.1%, and the largest one was the performance indicators and ideal indicators. The following conclusions
rutting factor with an error of 8.0%. Tong et al. (2016) used the can be drawn:

11
P. Lu, K. Ye, T. Jin et al. Engineering Applications of Artificial Intelligence 117 (2023) 105609

Table 12
Comparison between actual and ideal sample response indicators.
Softening point/◦ C Rutting factor/Pa Brookfield viscosity/Pa s Dispersion coefficient
Ideal sample response indicators 90 30 000 80 000 0.92
Actual sample response indicators 91.7 27 600 74 500 0.93
Error value 1.9% 8.0% 6.9% 1.1%

Based on the fluorescence microscope test, with the help of the Carrera, V., Partal, P., Garcia-Morales, M., Gallegos, C., Paez, A., 2009. Influence of
MATLAB image processing function, used the Christiansen coefficient bitumen colloidal nature on the design of isocyanate-based bituminous products
with enhanced rheological properties. Ind. Eng. Chem. Res. 48, 8464–8470.
method to quantitatively analyze the dispersion coefficient of the PU-
Chen, B., Dong, F., Yu, X., Zheng, C., 2021a. Evaluation of properties and micro-
modified bitumen. The results showed that the Christiansen coeffi- characteristics of waste polyurethane/styrene-butadiene-styrene composite modified
cient method well characterized the distribution uniformity of the asphalt. Polymers (Basel) 13.
PU-modified bitumen. Using the analysis of variance and range, the in- Chen, M., Geng, J., Xia, C., He, L., Liu, Z., 2021b. A review of phase structure of
dex with insignificant influence was excluded. It reduces the number of SBS modified asphalt: Affecting factors, analytical methods, phase models and
improvements. Constr. Build. Mater. 294.
responses, which improves the efficiency and accuracy of the prediction
Chen, Q., Xu, X., Sun, Q., Xia, D., 2010. A solution to the deficiencies of image
model without increasing the samples. enhancement. Signal Process. 90, 44–56.
Determined the main technological parameters of PU modified as- Cuadri, A.A., García-Morales, M., Navarro, F.J., Partal, P., 2014. Processing of bitumens
phalt through the Kriging-PSO algorithm. The performance test of the modified by a bio-oil-derived polyurethane. Fuel 118, 83–90.
final prepared ideal PU-modified bitumen shown that all indicators Dong, W., Zhou, M., 2017. A supervised learning and control method to improve
particle swarm optimization algorithms. IEEE Trans. Syst., Man, Cybernet.: Syst.
basically met the ideal indicators. And the error accuracy met the 47, 1135–1148.
requirements. This indicated that the fitting model constructed by using Fang, J., Tu, J., 2019. Effect of ultraviolet (UV) aging on rheology properties and
the Kriging-PSO algorithm had a good response function and could microstructure of polyurethane (PU) modified asphalt. Mater. Res. Express 6.
guide the determination of process parameters of the PU-modified Gujar, R., Vakharia, V., 2019. Prediction and validation of alternative fillers used
in micro surfacing mix-design using machine learning techniques. Constr. Build.
bitumen.
Mater. 207, 519–527.
Achieved the purpose of accurate design of process parameters by Guo, F., Zhang, J., Pei, J., Ma, W., Hu, Z., Guan, Y., 2020. Evaluation of the
using a small number of samples, which proved the feasibility and compatibility between rubber and asphalt based on molecular dynamics simulation.
efficiency of the Kriging-PSO algorithm. It can replace or combine the Front. Struct. Civ. Eng. 14, 435–445.
conventional design methods such as orthogonal design method and Isayev, O., Oses, C., Toher, C., Gossett, E., Curtarolo, S., Tropsha, A., 2017. Uni-
versal fragment descriptors for predicting properties of inorganic crystals. Nature
response surface method for the preparation process design of modified Commun. 8 (15679).
asphalt. The method can provide new ideas for the process design Izquierdo, M.A., Navarro, F.J., Martínez-Boza, F.J., Gallegos, C., 2012. Bituminous
of multi-variable/multi-response engineering material preparation. The polyurethane foams for building applications: Influence of bitumen hardness.
model performs well when there are no more than 4 variables and Constr. Build. Mater. 30, 706–713.
Jerome, S., W, W.J., Toby, J.M., Henry, P.W., 1989. Design and analysis of computer
responses. However, when the dimensions are further improved, the
experiments. Statist. Sci. 4, 409–423.
demand for the number of samples will increases significantly, and Khatir, S., Abdel Wahab, M., 2019. Fast simulations for solving fracture mechanics
the prediction accuracy will also decreases. The next research plan inverse problems using POD-RBF XIGA and Jaya algorithm. Eng. Fract. Mech. 205,
is to further optimize the algorithm to enhance its performance in 285–300.
high-dimensional situations. Khatir, S., Abdel Wahab, M., Boutchicha, D., Khatir, T., 2019. Structural health
monitoring using modal strain energy damage indicator coupled with teaching-
learning-based optimization algorithm and isogoemetric analysis. J. Sound Vib. 448,
Declaration of competing interest 230–246.
Khatir, S., Boutchicha, D., Le Thanh, C., Tran-Ngoc, H., Nguyen, T.N., Abdel-Wahab, M.,
The authors declare that they have no known competing finan- 2020. Improved ANN technique combined with Jaya algorithm for crack identifi-
cation in plates using XIGA and experimental analysis. Theor. Appl. Fract. Mech.
cial interests or personal relationships that could have appeared to 107.
influence the work reported in this paper. Khatir, S., Dekemele, K., Loccufier, M., Khatir, T., Abdel Wahab, M., 2018. Crack
identification method in beam-like structures using changes in experimentally
Data availability measured frequencies and Particle Swarm Optimization. Compt. R. Méc. 346,
110–120.
Kumar, Y., Singh, S.K., Oberoi, D., Kumar, P., Mohanty, P., Ravindranath, S.S., 2020.
No data was used for the research described in the article. Effect of molecular structure and concentration of styrene-butadiene polymer on
upper service temperature rheological properties of modified binders. Construct.
Build. Mater. 249.
Acknowledgments
Li, T.S., Gomez, N.H.C., Lu, G.Y., Liang, D., Wang, D.W., Oeser, M., 2021. Use of
polyurethane precursor-based modifier as an eco-friendly approach to improve
The authors gratefully acknowledge the financial support provided performance of asphalt. J. Transp. Eng. Part B-Pavements 147.
by the Science Foundation of China Postdoctor (Grant No. Liang, Y., Cheng, X.Q., Li, Z.N., Xiang, J.W., 2010. Multi-objective robust airfoil
2016M600352), the Science and Technology Agency of Zhejiang optimization based on non-uniform rational B-spline (NURBS) representation. Sci.
China-Technol. Sci. 53, 2708–2717.
Province (Grant No. 2015C33222 LGF19E080012), the Science and Lophaven, S.N., Nielsen, H.B., Søndergaard, J., 2002, DACE-A Matlab Kriging toolbox,
Technology Project of Zhejiang Provincial Department of Transporta- version 2.0.
tion (Grant Nos. 2019H14 and 2018010), and the Jiaxing Science and Lu, P., Xu, Z., Chen, Y., Zhou, Y., 2020. Prediction method of bridge static load test
Technology Bureau of China (Grant No. 2021AY10043). results based on Kriging model. Eng. Struct. 214.
Lv, Z., Wang, L., Han, Z., Zhao, J., Wang, W., 2019. Surrogate-assisted particle swarm
optimization algorithm with Pareto active learning for expensive multi-objective
References optimization. IEEE/CAA J. Autom. Sin. 6, 838–849.
Lv, S., Yuan, J., Peng, X., Borges Cabrera, M., Guo, S., Luo, X., Gao, J., 2020.
Bazmara, B., Tahersima, M., Behravan, A., 2018. Influence of thermoplastic Performance and optimization of bio-oil/Buton rock asphalt composite modified
polyurethane and synthesized polyurethane additive in performance of asphalt asphalt. Constr. Build. Mater. 264.
pavements. Constr. Build. Mater. 166, 1–11. Ma, S., Du, B., Chen, W., 2013. Collaborative optimization of hypersonic vehicle based
Cao, Y., Zhang, H., Li, W., Zhou, M., Zhang, Y., Chaovalitwongse, W.A., 2019. on surrogate model. J. Beijing Univ. Aeronaut. Astronaut. (China) 39, 1042–1047.
Comprehensive learning particle swarm optimization algorithm with local search Oliver, M.A., Webster, R., 2014. A tutorial guide to geostatistics: Computing and
for multimodal functions. IEEE Trans. Evol. Comput. 23, 718–731. modelling variograms and kriging. Catena 113, 56–69.

12
P. Lu, K. Ye, T. Jin et al. Engineering Applications of Artificial Intelligence 117 (2023) 105609

Poli R, K.J., Blackwell, T., 1995. Particle swarm optimization. Swarm Intell-US 1, Tran-Ngoc, H., Khatir, S., De Roeck, G., Bui-Tien, T., Abdel Wahab, M., 2019. An
33–57. efficient artificial neural network for damage detection in bridges and beam-like
Poli R, K.J., Blackwell, T., 2007. Particle swarm optimization. Swarm Intell-US 1, structures by improving training parameters using cuckoo search algorithm. Eng.
33–57. Struct. 199.
Ren, S., Liang, M., Fan, W., Zhang, Y., Qian, C., He, Y., Shi, J., 2018. Investigating the Ulaganathan, S., Couckuyt, I., Deschrijver, D., Laermans, E., Dhaene, T., 2015. A matlab
effects of SBR on the properties of gilsonite modified asphalt. Constr. Build. Mater. toolbox for kriging metamodelling. Procedia Comput. Sci. 51, 2708–2713.
190, 1103–1116. Vu-Huu, T., Phung-Van, P., Nguyen-Xuan, H., Abdel Wahab, M., 2018. A polytree-
Ren, J., Zang, G., Xu, Y., 2019. Formula and pavement properties of a composite based adaptive polygonal finite element method for topology optimization of
modified bioasphalt binder considering performance and economy. J. Mater. Civ. fluid-submerged breakwater interaction. Comput. Math. Appl. 76, 1198–1218.
Eng. 31. Wang, F., Ma, S., Wang, H., Li, Y., Qin, Z., Zhang, J., 2018. A hybrid model
Schneider, S., Schmidtke, C., Putzien, S., Dietzsch, M., Emig, S., Meyer, L., Nestle, N., integrating improved flower pollination algorithm-based feature selection and
Šandor, A., Kriegenhofer, K., Hafner, K., Šandor, M., 2021. Chemometrical per- improved random forest for NO X emission estimation of coal-fired power plants.
formance and composition prediction of bituminous binders based on MIR and Measurement 125, 303–312.
TD-NMR analyses. Road Mater. Pavement Des. 22, S310–S327. Wu, J., Yan, Q., Huang, S., Zou, C., Zhong, J., Wang, W., 2018. Finite element model
Shirzad, S., Hassan, M., Mohammad, L.N., 2020. Rheological and mechanical evaluation updating in bridge structures using kriging model and latin hypercube sampling
of polyurethane prepolymer-modified asphalt mixture with self-healing abilities. J. method. Adv. Civ. Eng. 2018, 1–11.
Mater. Civ. Eng. 32. Yi, P., Wei, K., Kong, X., Zhu, Z., 2015. Cumulative PSO-Kriging model for slope
Tiachacht, S., Bouazzouni, A., Khatir, S., Abdel Wahab, M., Behtani, A., Capozucca, R., reliability analysis. Probab. Eng. Mech. 39, 39–45.
2018. Damage assessment in structures using combination of a modified cornwell Zhao, L., Choi, K.K., Lee, I., 2011. Metamodeling method using dynamic kriging for
indicator and genetic algorithm. Eng. Struct. 177, 421–430. design optimization. AIAA J. 49, 2034–2046.
Tian, G., Ren, Y., Zhou, M., 2016. Dual-objective scheduling of rescue vehicles to Zou, G.L., Zhang, J.J., Li, Y.Y., Lin, Z.P., 2021. Quantitative characterize binder
distinguish forest fires via differential evolution and particle swarm optimization blending and diffusion in recycled asphalt mixture: An environmental-friendly
combined algorithm. IEEE Trans. Intell. Transp. Syst. 17, 3009–3021. solution using wooden cube and 3D fl uorescence image technology. J. Clean.
Tong, S., Xie, X., Zhao, D., Jiang, H., 2016. Nonlinear prediction of road performance Prod. 293.
of asphalt mixture after ultraviolet aging. J. Build. Mater. 19, 105–110.

13

You might also like