You are on page 1of 12

This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2019.2934996, IEEE
Sensors Journal

Sensors-27720-2019 1

Beetle-swarm Evolution Competitive Algorithm


for Bridge Sensor Optimal Placement in SHM
Jianhui Yang and Zhenrui Peng

 of the structure is evaluated, and important reference


Abstract—In structural heath monitoring (SHM), for the sake information is provided for the daily maintenance, safety and
of optimizing the bridge sensor placement, a beetle-swarm durability, so as to better serve the economic construction [2].
evolution competitive algorithm is proposed with the group Therefore, the quality of data collection directly affects the
evolutionary competition mechanism introducing into beetle
antennae search optimization (BAS). This paper uses high-
accuracy of SHM, and properly installing the right number of
dimensional dynamic coverage coding to initialize the beetle’s sensors at key parts of the structure to obtain the most effective
location as binary codes, which avoids that the original BAS can and valuable data is an essential component of the overall SHM.
only solve the continuous optimization problem, and ensures the In the SHM system, the sensor placement shall meet the
dispersion of effective information in individuals, so that the following requirements: (1) In the environment where all kinds
coding rules are consistent with the actual coverage density for of noise exist objectively, install as few sensors as possible in
bridge sensors. In the process of evolutionary, horizontal mutation,
crossover, phagocytosis and elimination among individuals are
the proper position, and obtain structural response data
added to improve the diversity of individual evolution, which may information as much as possible; (2) The obtained data can
avoid the algorithm falling into local optimum, and make the effectively deduce the structural feature information [3].
excellent individuals tend to be better. Lastly, the example analysis Theoretically, the more sensors installed, the more accurate the
of Ha-Qi bridge shows that beetle-swarm evolution competitive response data of the structure can be obtained. This is feasible
algorithm has fast convergence speed and global optimization for small bridges and other structures, but for large and complex
ability, and it is suitable for solving the sensor placement
optimization for large bridge.
structures, it is obviously impractical, and is not conducive to
practical engineering implementation [4]. Therefore, how to
Index Terms—Sensor placement optimization, beetle-swarm optimize the locations of sensors to get response information
evolution competitive algorithm, high-dimensional dynamic maximumly, which is the problem of sensor placement
coverage coding, phagocytosis and elimination strategy optimization, has been widely studied in recent years.
A. Criterion of Sensor Placement Optimization
I. INTRODUCTION Effective independent (EI), modal kinetic energy (MKE) and

A T present, the infrastructure construction is in a period of modal assurance criterion (MAC) are commonly used methods
rapid development with the rise of projects such as large- by many scholars to optimize sensor placement.
span bridges and ultra-high buildings in various countries. The In terms of the large-scale spatial structures, Kammer [5]
health of those projects is closely related to major safety issues proposed the well-known EI method to optimize sensor
that, once destroyed, will cause significant damages and impact placement in 1991. By gradually eliminating the degrees of
[1]. Therefore, it is extremely important to conduct effective freedom (DOF) that contributes least to the independence of the
structural health monitoring (SHM) to understand the service target mode, its spatial resolution can be maximized.
state of the structure in real time, and to issue a safety warning The principle of MKE is that in the DOF with larger modal
before the structure is damaged. strain energy, the placement of sensors will be beneficial to
In general, a SHM system mainly includes three components, parameter identification. Li et al. [6] deduced that EI was an
namely: data collection terminal, data transmission network and iterated version of the MKE, and the reduced mode shapes were
data processing center. By installing an appropriate number of ortho-normalized repeatedly during iterations of the former.
sensors at key parts of the structure, a data acquisition terminal Given the influence exerted on three directions in the process
can be established to collect response information such as of arranging the sensors, Yi et al. [7] proposed a new three-axis
displacement, velocity and acceleration of the structure. accelerometer optimization method- MKE3 by analyzing the
Through scientific numerical calculations, the health condition relationship between EI and MKE methods.

Manuscript received April 17, 2019.This work was supported in part by the Jianhui Yang is with the School of Electronic and Information Engineering,
National Natural Science Foundation of China under Grant 51768035, in part Ph.D with the School of Mechatronic Engineering, Lanzhou Jiaotong
by the Collaborative Innovation Team Project of Universities in Gansu University, Lanzhou 730070, China (e-mail: yangjh@mail.lzjtu.cn).
Province under Grant 2018C-12, in part by the Lanzhou Science and Zhenrui Peng is with the School of Mechatronic Engineering, Lanzhou
Technology Planning Projects under Grant 2018-1-51, in part by the Youth Jiaotong University, Lanzhou 730070, China (e-mail: pengzr@ mail.lzjtu.cn).
Fund of Lanzhou Jiaotong University under Grant 2015009. (Corresponding
author: Zhenrui Peng.)

1558-1748 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2019.2934996, IEEE
Sensors Journal

Sensors-27720-2019 2

The MAC method keeps the measured modal vectors at a bridge sensor placement optimization. Firstly, combined with
large spatial angle when selecting the measuring points, so as the density information of bridge sensors, beetle swarm with a
to retain the characteristics of the original model as much as certain scale is generated randomly, and is timely transformed
possible. Carne et al. [8] considered that the MAC matrix of into binary code through using High-dimensional Dynamic
modal confidence is a fine tool for evaluating the intersection Coverage Coding (HDCC) [13]. Then the mutation, crossover,
angle of modal vector space. Yuan et al. [9] combined EI with phagocytosis and elimination are taken place among the beetles
MAC and proposed a joint algorithm for optimal placement of with the idea of DE, in order to increase their diversity of the
sensors, which effectively solved the shortcomings of a single individual evolution, avoid the individuals falling into the local
method. optimum, improve the convergence speed, and make the
excellent individual has more opportunities to lead the whole
B. Methods of Sensor Placement Optimization
beetle swarm. The global optimal solution may be finally
Sensor placement optimization is a typical combinatorial obtained in this way and the feasibility and effectiveness of the
optimization problem. If 𝑀 sensors are placed in 𝑁 positions, algorithm are verified by an engineering example.
then 𝐶 objective functions need to be calculated by the The remainder of this paper develops as follows: Section II
exhaustive method. This can be used to solve simple structures, provides the mathematical model and objective function of the
but is difficult to solve complex ones. Many scholars have bridge sensor placement optimization. Section III introduces
proposed a variety of optimization methods, which may play a the process and derivation of BECA. Section IV presents the
good guiding role in solving sensor optimization problems. experiment and the contrast analysis of the algorithm through
One of the most important optimization methods is sequence three benchmark functions. Section V verifies the algorithm by
method, whose basic principle is to minimize the maximum off- an engineering example and compares it with other intelligent
diagonal elements of MAC matrix. Sequence method contains algorithms. Finally, Section VI concludes the paper and puts
cumulative and gradual elimination method. The former one is forward the considerations for realistic work.
to continuously increase the optional position of sensors until a
satisfactory number is reached, and the latter method is just the II. THE MATHEMATICAL MODEL FOR BRIDGE SENSOR
opposite of the former. However, the sequence method usually PLACEMENT OPTIMIZATION
obtains only suboptimal solutions.
In SHM, the first step is to determine the sensor nodes of the
In recent years, many researchers have attempted to apply
bridge structure to get the response data [20], [21]. It is
modern intelligent optimization algorithm to optimal sensor
unrealistic to get all the response information of DOF because
placement. These algorithms mainly include genetic algorithm
there are so many measurable DOF of the bridge structure.
(GA), particle swarm optimization (PSO), ant colony
Therefore, it is necessary to optimize the position of
optimization (ACO), simulated annealing algorithm (SAA),
measurement nodes in order to obtain the most valuable
artificial bee colony algorithm (ABC), monkey algorithm (MA)
response data.
and so on.
As described above, the 𝑀𝐴𝐶 matrix is considered as an
Elsersy et al. [10] used GA, with mixed integer coding and
effective tool for evaluating the intersection angle of modal
mitigated complexity, to jointly optimize sensor placement,
vector space. The elements of 𝑀𝐴𝐶 matrix can be expressed as
routing, and flow allocation. They also validated the algorithm
follows:
in a 9-story building. Cha et al. [11] used multi-objective GA
to optimize the placement of actuators and sensors in large civil 𝝋𝒊 𝝋𝒋
𝑀𝐴𝐶 = (1)
structures under earthquake excitation. Rao et al. [12] applied 𝝋𝒊 𝝋𝒊 𝝋𝒋 𝝋𝒋
PSO to optimal sensor placement for structural system where 𝝋𝒊 and 𝝋𝒋 represent the 𝑖th and 𝑗th mode shape vectors
identification and health monitoring. Yang et al. [13] proposed
respectively. A certain off-diagonal element of 𝑀𝐴𝐶 (𝑖 ≠ 𝑗)
an improved ABC algorithm based on dynamic random cover
coding initialization and matching optimization evolution can reflect the recognizability of 𝑖th and 𝑗th mode shape
strategy, took a large bridge as an example to optimize the vectors. The smaller the value, the higher the recognizability of
sensor placement, and then obtained a sound result. There are the two vectors, and the better the effect of sensor placement is.
many examples of using intelligent optimization algorithms to Accordingly, the placement of the measuring nodes shall seek
optimize placement of structural sensors, which will not be to minimize the maximum off-diagonal elements of the matrix.
described in detail here. In terms of large scaled bridge structures, six modal shape
When this kind of swarm intelligence algorithm is applied to vectors can be obtained by establishing a finite element model
optimize sensor placement, different degrees of advantages (FEM) (each node contains 6 DOFs: translation 𝑈 , 𝑈 , 𝑈 , and
have been shown from following aspects: some people solve a rotation 𝑅 , 𝑅 , 𝑅 ). Furthermore, the modal vector with the
specific problem, some improve the convergence speed and the highest participatory quality results is selected to form the mode
optimization precision of the algorithm [14-17]. shape matrix ∅ × as the basic data, among which 𝐷 is the
Based on a new beetle antennae search (BAS) optimization nodes number in the FEM while 𝐿 is the selected modal order.
algorithm [18], the swarm evolution competition mechanism In all candidate nodes, 𝑀 nodes are selected as sensor
and differential evolution (DE) idea [19], this paper proposes a placement points. That is, 𝑀 rows of data are selected from
Beetle-swarm Evolution Competitive algorithm (BECA) for ∅ × to form a 𝑀𝐴𝐶 matrix so that its maximum off-diagonal

1558-1748 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2019.2934996, IEEE
Sensors Journal

Sensors-27720-2019 3

element tends to be minimum, and a relatively sound sensor


placement result may be obtained.
The objective function 𝑓𝑖𝑡 is constructed based on the data
characteristics of 𝑀𝐴𝐶 matrix and its practical significance is
as follows:
𝑓𝑖𝑡(𝑥) = 𝑚𝑎𝑥{ 𝑀𝐴𝐶 }, (𝑖, 𝑗 ∈ [1, 𝐿], 𝑖 ≠ 𝑗) (2)
where 𝑥 is the binary array containing 𝐷 elements, its number
of “1” is exactly 𝑀, “1” and “0” represent whether to select the
row data of matrix ∅ × on their corresponding position
number, or not. Specifically, a series of location number where
“1” exists can be gotten from the binary array 𝑥. According to
these location number, we can decide which row data may be
selected from matrix ∅ × to form a 𝑀𝐴𝐶 matrix according to
Eq (1). Conversely, if the element in the binary array is “0”, it
means that the row data corresponding to “0” is not selected in
the matrix ∅ × . In this way, according to the different
distribution of element “1” in binary array 𝑥, different 𝑀𝐴𝐶
matrix can be combined. 𝑥 is just the solution space of the
problem. The objective of this paper is to find a set of suitable
𝑥 to minimize the objective function value (OFV), which is
even close to 0 in ideal, on the basis of the bridge sensor
placement optimization and the shape mode data.
Fig. 1. Flow chart of BAS

III. BEETLE-SWARM EVOLUTION COMPETITIVE ALGORITHM There is only one beetle individual in the BAS, thus
A. Beetle Antennae Search compared with other group intelligent algorithms such as ACO,
GA etc., the BAS algorithm holds the advantage of simple
Inspired by the beetle foraging principle, Jiang et al. [18]
principle, less parameters, fast calculation and easy
proposed an efficient intelligent optimization algorithm named
programming especially for on-chip code. It has a very small
beetle antennae search (BAS). Its core idea is that the beetle
time and space complexity, and has obvious advantages in
would be driven to get food quickly if the food odor intensity
computing efficiency. However, the BAS algorithm can only
(equivalents to the OFV) is enough received by the two
solve the continuous optimization problem at present, and the
antennae.
ability of a single beetle to find the optimal solution space is
BAS algorithm contains two basic rules: searching and
limited, so it is difficult to solve the high-dimensional and
detecting. In each iteration, the beetle searches for food in a
complex combinatorial optimization problem such as bridge
random direction and then its left and right antennae’s location
sensor optimization. Based on the practical requirements of
are calculated with formulas (3) and (4) to detect the next
bridge sensor placement, BAS is extended to beetle group.
moving direction. Combining the BAS’ fast iteration, group evolution mechanism
𝒙𝒍 = 𝒙𝒕 − 𝑑 𝒃 (3) and DE method, the BECA is proposed.
𝒕
𝒙𝒓 = 𝒙 + 𝑑 𝒃 (4) B. Description of BECA
where 𝒙𝒍 and 𝒙𝒓 represent the beetle’s left and right antennae  Initialization and Coding of the Beetle Individuals
vectors respectively; 𝒙𝒕 represents the current position vector of In the process of searching for the optimal solution of beetle
the beetle; 𝑑 is the search space and 𝒃 is a random vector. swarm, dual-coding-mode is used to express the beetle
After integrating formulas (3) and (4), the beetle’s position may individuals. Their positions are represented by bounded
be updated by formula (5) in the iterative process: continuous variables which are randomly generated and are
𝒙𝒕 = 𝒙𝒕 𝟏
+ 𝛿 𝒃𝑠𝑖𝑔𝑛(𝑓(𝒙𝒓 ) − 𝑓(𝒙𝒍 )) (5) directly applied to the process of the beetle’s left-right antennae
coordinate generation, beetle’s mutation, crossover,
wherein 𝛿 ∈ [0,1] is the attenuation coefficient of the beetle
phagocytosis and elimination. On the other hand, during the
step; 𝑓(𝒙𝒓 ) 𝑎𝑛𝑑 𝑓(𝒙𝒍) represent the fitness values
process of selection and comparison, the binary codes of the
corresponding to the right and left antennae of the beetle and
beetle individuals generated by HDCC [13] are used as the
𝑠𝑖𝑔𝑛 is a sign function [22].
potential solutions. After solving the OFV in accordance with
The flow chart of BAS algorithm is shown in Fig. 1.
the formula (1), only the optimal beetle’s binary code is
preserved so as to reduce the computational complexity of the
algorithm and improve the efficiency.
1) Generation Mechanism of Continuous Variables for Beetle
Individuals

1558-1748 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2019.2934996, IEEE
Sensors Journal

Sensors-27720-2019 4

Assuming that the scale of beetle swarm is 𝑁𝑃, the position to the formula (10) to avoid such defects as low precision
vector of the 𝑖 th beetle can be initialized with continuous caused by too large operator and precocity caused by too small
variables according to formula (6): operator.
𝒙𝒊 = 𝑥𝑀𝑖𝑛 + (𝑥𝑀𝑎𝑥 − 𝑥𝑀𝑖𝑛) × 𝑟𝑎𝑛𝑑, 𝑖 = 1,2, ⋯ , 𝑁𝑃(6) 𝛼=𝑒 ,𝐹 =𝐹 ×2 (10)
where 𝑥𝑀𝑎𝑥 and 𝑥𝑀𝑖𝑛 are upper and lower bounds of the where 𝐺 is the maximum iteration of the algorithm; 𝑡 is the
beetle position coordinates; 𝑟𝑎𝑛𝑑 is a function that generates a current iteration and 𝐹 is the mutation operator. At the
random number between 0 and 1. preliminary stage, 𝐹 = 2𝐹 . With a large value, the diversity of
2) Binary Coding of Beetle Individuals the individual is maintained to avoid the algorithm falling into
Suppose that 𝐷 nodes are used to establish FEM for a certain premature. After all iterations, the value of 𝐹 gradually
bridge structure, and 𝑀 represents the number of sensors to be decreases to 𝐹 , which retains more excellent information of the
placed, then the coverage density of sensors installed on the individual and averts the destruction of the optimal solution and
bridge can be expressed as: increases the probability that the algorithm searches for the
𝜌= × 100% (7) global optimal solution.
2) Probabilistic crossover of the beetle individuals
where 𝐷 is a fixed value for a certain bridge; 𝑀 is variable. In order to increase the diversity of the algorithm parameters,
Different values correspond to different sensor coverage the individual crossover operation is carried out in probabilistic
densities 𝜌 . Dynamically combining 𝜌 to encode the binary manner via formula (11). The elements of the crossed beetle
elements of the beetle individuals, the beetles’ binary codes 𝑥 individual vector has been set as 𝑢 :
can be expressed as follows:
𝑣 , 𝑐𝑟 ≤ 𝐶0 𝑜𝑟 𝑗 = 𝑟
1, 𝑟𝑎𝑛𝑑 > 1 − 𝜌 𝑢 = , 𝑖 = 1,2, ⋯ , 𝑁𝑃; 𝑗 = 1,2, ⋯ , 𝐷
𝑥 = , 𝑖 = 1,2, ⋯ , 𝑁𝑃; 𝑤 = 1,2, ⋯ , 𝐷(8) 𝑥 , 𝑐𝑟 > 𝐶0 𝑜𝑟 𝑗 ≠ 𝑟
0, 𝑟𝑎𝑛𝑑 ≤ 1 − 𝜌
(11)
It can be found that, the larger the number 𝑀 of bridge sensor
where 𝑐𝑟 is a random number and 𝐶0 is crossover operator,
placement, the greater the sensor coverage 𝜌, and the higher the
both of them belong to [0,1]; 𝑟 ∈ [1, 𝐷] is a random integer
probability that the corresponding number in 𝑤 dimension is
which makes sure that 𝑢 gets at least one parameter from 𝑣
initialized to “1” (chosen sensor). This situation just
corresponds to the sensor coverage intensity, and vice versa. to achieve effective crossover.
Therefore, the dispersion degree of individual initialization is 3) Phagocytosis and elimination of the beetles
dynamically constrained by sensor coverage density 𝜌 of the Phagocytosis and elimination operation are the key parts of
overall bridge. BECA, replacing the eliminated individual with the optimal one,
The coding method may avoid the shortcomings of the and increasing the probability that the group is close to the
traditional threshold segmentation, for example, the effective global optimal solution. In the process of beetle evolution, the
information is easy to be centralized or the individual diversity current fitness and stagnation times of each beetle are recorded.
is insufficient. More importantly, the coding method is related In each iteration, if the beetle does not evolve effectively, add 1
to the bridge sensor density, which is more in line with the to the number of stagnation. When the times of stagnation
actual demand. exceed the pre-set 𝑙𝑖𝑚𝑖𝑡, the beetle is discarded and replaced by
 BECA the optimal one, that is to say, the optimal beetle has more
The process of BECA is as follows: On the basis of the BAS chances to reach the optimal solution first by engulfing the
method, it integrates group information sharing, competitive worst one.
elimination and DE method. By increasing the mutation To sum up, the pseudo-code of BECA for optimizing bridge
crossover, and phagocytosis among individuals, the beetle sensor placement is as follows (where HDC (𝑥01 ) means that
individuals learn from each other, inherit the excellent gene the individual is binary coded by HDCC and the individual
segments, phagocytize and eliminate the inferior individuals, OFV is calculated):
and replace them with the best individuals, so that the optimal
one has more opportunities to engulf the whole group to find
the global optimal solution.
1) Beetle Mutation
The mutation of the 𝑖th beetle can be performed according to
formula (9).
𝒗𝒊 = 𝒙𝒓𝟏 + 𝐹(𝒙𝒓𝟐 − 𝒙𝒓𝟑 ), 𝑖, 𝑟 , 𝑟 , 𝑟 = 1,2, ⋯ , 𝑁𝑃 (9)
where 𝒗𝒊 indicates the position vector of the 𝑖th beetle; 𝑟 , 𝑟 ,
𝑟 are the other 3 different beetles from which the 𝑖th beetle
learns information (obviously the mutation operation requires
that 𝑁𝑃 is not less than 4); 𝐹 is a mutation operator which
controls the amplification of deviation variables [23]. This
paper introduces the adaptive mutation operator [24] according

1558-1748 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2019.2934996, IEEE
Sensors Journal

Sensors-27720-2019 5

Algorithm Beetle-swarm Evolution Competitive Algorithm


(BECA)
01: Input parameters: 𝐷 : the freedom of bridge (the number
of candidate sensors), 𝑀 : the number of sensors to be
placed, 𝑁𝑃 ≥ 4: the beetles scale, 𝐺: the number of iterations,
𝑠𝑡𝑒𝑝: step length of the beetle, 𝑒𝑡𝑎: attenuation coefficient of
step length, 𝑑0: the distace between the two antannaes, 𝐹0:
the mutation operator, 𝐶0: the crossover operator, 𝑙𝑖𝑚𝑖𝑡: the
upper limit stagnation times of the beetle, 𝑥𝑀𝑎𝑥/𝑥𝑀𝑖𝑛 :
upper and lower bounds of the beetle location.
02: Initialize the individual position information of the beetle
group 𝑥 .
03: Calcuate the OFV 𝑓𝑖𝑡 according to HDC( 𝑥01 ) and
record the OFV of the optimal beetle 𝑓𝑖𝑡 , 𝒙𝑩 , 𝑥01 .
04: for t = 1 : 1 : G, do
05: Calculate the vertors of the beetle’s left-right antennae,
and update the location information according to HDC().
06: Beetle mutation.
07: Crossover the individuals in a probabilistic manner.
08: Individual boundary condition processing.
09: Calculate 𝑓𝑖𝑡 _ at 𝑡th iteration by HDC(). Fig. 2. The flow chart of BECA

10: if 𝑓𝑖𝑡 _ < 𝑓𝑖𝑡 _ C. Analysis of Algorithmic Efficiency


11: Update the position information of beetles. The time and space complexity of BAS are 𝑂(𝐷 × 𝐺) and
12: else 𝑂(𝐷), and those of DE are 𝑂(𝑁𝑃 × 𝐷 × 𝐺) and 𝑂(𝑁𝑃 × 𝐷),
13: Add the stagnation 𝑇𝑟𝑎𝑖𝑙 of the current beetle by 1. where 𝐷 is solution dimension, 𝐺 is the number of iterations
14: if 𝑇𝑟𝑎𝑖𝑙 = 𝑙𝑖𝑚𝑖𝑡 and 𝑁𝑃 is the swarm scale. It is precisely because that the
15: Eliminate the beetle and replace it with the best one. BECA algorithm proposed in this paper is improved on the
16: end if basis of BAS and DE combining competition and cooperation
17: end if strategy among individuals, its beetle scales are consistent with
18: Save the information of the optimal beetle 𝑓𝑖𝑡 , 𝒙𝑩 , that in DE, the mutation and crossover operations among
𝑥01 . beetles are also derived from the related operations in DE.
19: end for Therefore, the time and space complexity of BECA are also
20: Output the optimal beetle information 𝑓𝑖𝑡 , 𝒙𝑩 , 𝑥01 𝑂(𝑁𝑃 × 𝐷 × 𝐺) and 𝑂(𝑁𝑃 × 𝐷), which do not increase the
and its 𝑀𝐴𝐶 matrix. extra overhead of computing time and space compared with DE.
The flow chart of BECA is shown in Fig. 2. Although BECA has a lower efficiency than BAS, it may get
higher computing capability by exchange, this is because
BECA has more individuals and more evolutionary strategies,
which may lead to the algorithm a stronger one.

IV. EXPERIMENTS OF ALGORITHM VALIDATION AND ANALYSIS


A. Benchmark Functions and Test Preparation
In order to test the effectiveness of the proposed method, we select three benchmark functions for test, which are shown in
Table I, where 𝐷 denotes the solution dimension, it may be 2 or 3 in Rosenbrock, and 30 or 60 in other functions [25].
TABLE I
TEST FUNCTIONS
Function Range Formulate

Rosenbrock [−2.048, 2.048] 𝑓(𝑥) = [100(𝑥 − 𝑥 ) + (𝑥 − 1) ]

1 1
Ackley [−32.768, 32.768] 𝑓(𝑥) = −20 𝑒𝑥𝑝 ⎛−0.2 𝑥 ⎞ − 𝑒𝑥𝑝 𝑐𝑜𝑠(2𝜋𝑥 ) + 𝑒 + 20
𝑛 𝑛
⎝ ⎠
1 𝑥 − 100
Griewank [−600, 600] 𝑓(𝑥) = (𝑥 − 100) − 𝑐𝑜𝑠 +1
4000 √𝑖

1558-1748 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2019.2934996, IEEE
Sensors Journal

Sensors-27720-2019 6

In order to compare the effectiveness of different algorithms, TABLE IV


here, four Cases are selected to search the optimal solution of AVG AND SD OF THE ACKLEY FUNCTION UNDER DIFFERENT
the three benchmark functions. CASES
Case I: Using BECA to search for the optimal solution of the
D=30 D=60
three benchmark functions. Case
Avg SD Avg SD
Case II: Using beetle swarm antennae search algorithm 1 1.89E-04 8.47E-04 0.0207 0.0839
(BSAS), which combines BAS algorithm with adaptive step 2 19.7108 0.2991 20.3918 0.0976
size feedback mechanism [26-27] to search for the optimal 3 7.0882 0.9209 8.1137 0.9635
solution of the three functions. (As BAS has one single beetle, 4 9.67E-04 4.14E-04 0.4988 0.2499
it is difficult to solve such optimization problem especially for TABLE V
high-dimensional optimization problems. Here, BSAS is used AVG AND SD OF THE GRIEWANK FUNCTION UNDER DIFFERENT
for comparative analysis). CASES
Case III: Using beetle swarm optimization algorithm (BSO),
which combines BAS algorithm with particle swarm Case
D=30 D=60
optimization (PSO) method [28], to solve the optimal solution Avg SD Avg SD
1 2.00E-03 4.24E-03 4.07E-03 5.38E-03
of the three functions.
2 2.0622 0.0663 3.8961 0.1819
Case IV: Using DE to solve the optimal solution of the three 3 0.9363 0.1015 1.0201 0.0369
functions. 4 4.59E-07 5.21E-07 0.0149 0.0120
Because the solution space of the above-mentioned functions
is continuous, the initial HDCC is removed and the initial The “Avg” and “SD” in the Tables show the BECA’
continuous solution is generated randomly in four Cases. Each superiority in searching the minimum value of the three
experiment is repeated 20 times and the number of iterations is benchmark functions, except the Griewank one (𝐷=30). In most
1000 in each calculation using MATLAB 2017b software. In cases, the results of BECA are slightly better than DE,
addition, the number of swarm scale is 50 in all Cases and the especially much better than that of BSAS and BSO. Especially
parameters of BECA method are shown in Table II. the SD results of BECA show that the convergence solution is
stable and the global convergence of BECA is guaranteed. This
TABLE II result is attributed to the cooperation and competition among
PARAMETERS OF BECA METHOD FOR THE THREE BENCHMARK BECA individuals, so that beetles with strong adaptability have
FUNCTIONS more opportunities to approach the optimal solution. The
Parameter Name Value optimal results with bold font also show that the convergence
Solution dimension (𝐷) 2/3/30/60 precision decreases with the increase of dimension 𝐷.
Swarm Scale (𝑁𝑃) 50  Comparison of Convergence Speed of Different Cases
Number of iterations (𝐺 ) 1000
Step length of beetle (𝑠𝑡𝑒𝑝) 1 In order to compare the convergence speed of different Cases,
Attenuation coefficient of 𝑠𝑡𝑒𝑝 (𝑒𝑡𝑎 ∈ [0,1]) 0.95 we extract the average convergence curves of different
Distace of the beetle’s two antannaes (𝑑0) 𝑠𝑡𝑒𝑝/5 algorithms from the 20 experiments above, which are shown in
Mutation opertor (𝐹0) 0.4 Fig. 3. Meanwhile, the average number of iterations when the
Crossover opertor (𝐶0) 0.1
Upper limit of stagnation times(𝑙𝑖𝑚𝑖𝑡) 50
optimal value appears and the average computation time with
Upper and lower bounds of the beetle location 1000 iterations are shown in Tables VI-VIII.
‘Range’ in Table I
(𝑥𝑀𝑎𝑥/𝑥𝑀𝑖𝑛)

B. Experiment Result
Function Value

Function Value

 Comparison of Convergence Precision of Different Cases


Through repeated calculation of the three benchmark
functions 20 times in different situations, Tables III–V present
the average convergence results (AVG) and standard deviation
(SD) of the Rosenbrock, Ackley, and Griewank functions
respectively.
(a) D=2 (b) D=3
TABLE III
AVG AND SD OF THE ROSENBROCK FUNCTION UNDER
DIFFERENT CASES
Function Value

Function Value

D=2 D=3
Case
Avg SD Avg SD
1 1.72E-04 7.63E-04 1.76E-02 0.0116
2 8.69E-04 8.82E-04 0.1977 0.1904
3 0.0246 0.0557 0.9618 1.1553
4 1.59E-03 0.0071 0.0290 0.0711
(c) D=30 (d) D=60

1558-1748 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2019.2934996, IEEE
Sensors Journal

Sensors-27720-2019 7

BECA algorithm. When the three benchmark functions are


solved by BECA, the solution dimension 𝐷 has a great
Function Value

Function Value
influence on the solution effect of the algorithm. When the
dimension is too large, it is difficult to solve the problem for the
algorithms naturally, which is manifested in the slow process of
solving and the difficulty of finding the global optimal solution.
Generally, the solution accuracy may be provided in some
degrees by increasing the size of the swarms and the number of
(e) D=30 (f) D=60 iterations.
Fig. 3. Average convergence curves of the three benchmark functions Among the parameters of the BECA algorithm itself, the
influence are as follows:
TABLE VI
The swarm scale 𝑁𝑃 affects the global optimization ability
AVERAGE NUMBER OF ITERATIONS AND COMPUTATION TIME
of the algorithm. The larger the 𝑁𝑃 is, the richer the diversity
FOR ROSENBROCK FUNCTION of individuals is and the stronger the solving ability of the group
D=2 D=3
is, which will make it easier for the algorithm to find the optimal
Case solution at a faster speed.
Iterations Time(s) Iterations Time(s)
1 710 0.7315 525 0.7705 The number of iterations 𝐺 should be flexible with 𝑁𝑃 ,
2 622 1.9788 587 2.0474 otherwise, for too large 𝑁𝑃 with too small 𝐺, it is difficult for
3 471 1.7100 99 1.7298
each individual to maximize its potential to find the optimal
4 993 0.2169 838 0.2215
solution. Especially when solving high-dimensional problems,
TABLE VII the value of 𝐺 should not be too small.
AVERAGE NUMBER OF ITERATIONS AND COMPUTATION TIME 𝑠𝑡𝑒𝑝 , 𝑒𝑡𝑎 and 𝑑0 are size parameters of the beetle
FOR ACKLEY FUNCTION individuals, wherein 𝑠𝑡𝑒𝑝 decays with 𝑒𝑡𝑎 ( 𝑠𝑡𝑒𝑝 = 𝑠𝑡𝑒𝑝 ×
𝑒𝑡𝑎). When the solution boundary of the function to be solved
D=30 D=60
Case
Iterations Time(s) Iterations Time(s)
is large, the big height beetles (𝑠𝑡𝑒𝑝 is a larger value) can be
1 980 1.3927 991 1.6476 selected to find the optimal solution with fast changing step size
2 462 3.2398 395 4.0941 (𝑒𝑡𝑎 is smaller but not less than 1), so as to avoid the beetles
3 90 2.2152 142 2.2768 falling into local optimum and make them have a wider search
4 997 0.4639 995 0.6119
areas. On the contrary, when the solution domain is narrow, we
TABLE VIII can use small beetles (𝑠𝑡𝑒𝑝 is a smaller value) to search for the
AVERAGE NUMBER OF ITERATIONS AND COMPUTATION TIME optimal solution with a slowly changing step size (𝑒𝑡𝑎 is larger
FOR GRIEWANK FUNCTION but not greater than 1) and a higher search density, so that no
more solutions are lost.
D=30 D=60 𝐹0 and 𝐶0 are mutation and crossover factors of beetle
Case
Iterations Time(s) Iterations Time(s)
individuals. 𝐹0 controls the influence of difference variables,
1 985 1.2659 999 1.7351
2 575 2.9417 484 4.1311 and its value should not be too large, otherwise the difference
3 905 2.1808 861 2.4138 information between individuals may have a great impact on
4 997 0.4581 996 0.6912 the mutated individuals; 𝐶0 represents the crossover
As can be seen from Fig. 3 and Tables VI-VIII, that BSO probability of the beetle individuals. The larger it is, the more
algorithm has relatively better convergence speed, but it is very opportunities for individuals to cross-operate, but it may also
easy to get trapped in a local optimum. Unfortunately, BSAS destroy the original better individuals. Therefore, 𝐹0 and 𝐶0
has poor convergence speed and accuracy in solving the determine the information sharing degrees among individuals
benchmark functions, especially high-dimensional ones. BECA together.
converges slightly faster than DE but takes a little longer time 𝑙𝑖𝑚𝑖𝑡 is the elimination threshold, or the number of
to calculate 1000 iterations as shown in Tables VI-VIII. This is
opportunities for the worst individual to work in the solution.
because BECA algorithm has more beetles’ behavior and a
Obviously, it may lead to the futility of the poorer individuals
special competitive operation comparing with DE, which leads
and waste of resources if the value of 𝑙𝑖𝑚𝑖𝑡 is too large. On the
to a slightly longer computation time, but in return for a better
calculation accuracy. In addition, BECA and DE both other hand, if 𝑙𝑖𝑚𝑖𝑡 is too small, it is easy to eliminate the worst
encounter the optimal value near the end of the iteration. If the individuals too early without giving them enough chance to
number of iterations is increased, it is possible to further correct themselves, which may lead to miss the potential good
improve the solution accuracy. In short, BECA has a better individuals. Therefore, the choice of 𝑙𝑖𝑚𝑖𝑡 is the key part of the
convergence speed and acceptable computation time when elimination strategy.
solving the three benchmark functions. Finally, the upper and lower bounds 𝑥𝑀𝑎𝑥/𝑥𝑀𝑖𝑛 depend
entirely on the solution space of the problem to be solved. If
C. Sensitivity Analysis of the Parameters
they hold large values, which are equal to a very large space for
The parameters of BECA algorithm are shown in Table II. beetle activity, then the parameters of the algorithm such as
Each parameter has a certain influence on the performance of

1558-1748 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2019.2934996, IEEE
Sensors Journal

Sensors-27720-2019 8

swarm scale, individual size and number of iterations should be Since each node of the bridge FEM has 6 DOFs, on the basis
as large as possible, so that there are enough individuals in the of analyzing the participatory quality results of mode shape, the
large space to find the optimal solution, and improve the 𝑈𝑦 direction mode shape vector has been selected as the basic
algorithm's solving ability. data to form a modal matrix ∅ × . The challenge has been
transformed to how to select 𝑀 rows data in ∅ × so that
V. ENGINEERING EXAMPLE ANALYSIS the corresponding maximum off-diagonal element of 𝑀𝐴𝐶
A. Introduction to the Engineering Example matrix (OFV) is minimized.
In order to further verify the effectiveness of the proposed B. Optimizing Bridge Sensor Placement Based on BECA
method, this paper takes the Ha-Qi railway bridge (abbreviated It is difficult to determine the exact number of sensors for
as H-Q RB) as an example to optimize the sensor placement. large bridges according to human experience. Therefore, it is
The bridge is a four-line continuous beam combination one, necessary to determine the optimal number 𝑀 of sensors firstly
located in Harbin, China. It was built in 2010 and opened to and then determine the optimal position of sensors based on 𝑀.
traffic in 2015. The main bridge length is 624.4 (77 + 3×156.8 When using BECA to optimize the sensor placement of H-Q
+ 77) meters, and the main span bridge is 156.8 m. The actual RB, according to the BECA description, the algorithm
scene of the bridge is shown in Fig. 4. parameters are set as shown in Table X.
TABLE X
PARAMETERS OF BECA FOR ENGINEERING ANALYSIS
Parameter Name Value
𝑈𝑦 DOF (𝐷) 1251
Sensor Number (𝑀) undetermined
Fig.4. View of H-Q RB Sensor coverage density (𝜌) 𝑀/𝐷
Swarm Scale (𝑁𝑃) 50
Using Midas Civil2017 software to establish a FEM to Number of iterations (𝐺 ) 1000
analyze the component with beams unit, there are 1321 nodes Step length of beetle (𝑠𝑡𝑒𝑝) 1
Attenuation coefficient of 𝑠𝑡𝑒𝑝 (𝑒𝑡𝑎 ∈ [0,1]) 0.95
and 1393 finite elements in the model. Considering that the
Distace of the beetle’s two antannaes (𝑑0) 0.2
calculation of the bridge is mainly based on the superstructure, Mutation opertor (𝐹0) 0.4
and the deformation of the bridge is mainly reflected in the Crossover opertor (𝐶0) 0.1
vertical direction, while the vertical deformation of the pier is Upper limit of stagnation times(𝑙𝑖𝑚𝑖𝑡) 50
very small, unless the foundation settlement occurs. Thus, when Upper and lower bounds of the beetle
10/-10
location (𝑥𝑀𝑎𝑥/𝑥𝑀𝑖𝑛)
optimizing the position of bridge sensor nodes, only the
superstructure is extracted for calculation. There are 1251 nodes  Determining the Optimal Sensor Number 𝑀
and 1373 finite elements in total. The three-dimensional FEM There are 1251 nodes in H-Q RB model, and based on the
of the H-Q RB is shown in Fig. 5. actual demand, the number of the placed sensors is not be too
large or small. In order to reduce the calculation, this paper
determines the value of 𝑀 only between 20 to 920 initially by
larger step size and then determines the optimal number of
(a) FEM (b) Upper FEM sensors by smaller one. Three Cases are compared and analyzed
Fig. 5. Three-dimensional FEM of the H-Q RB in MATLAB 2017b:
For the modal analysis of the upper structure, considering Case I: Using BECA to search for the number of sensors
that the lower order modes have larger mode participation, and placed in H-Q RB.
the first ten order mode shapes are extracted and the frequencies Case II: Using BSAS to search for sensor number of H-Q RB.
are shown in Table IX. Case III: Using DE to search for the number of sensors placed
in H-Q RB.
TABLE IX
1) Searching 𝑀 value with larger step size
THE FIRST TEN ORDER MODAL FREQUENCY OF UPPER
STRUCTURE OF H-Q RB Here, the value of 𝑀 is searched preliminarily from 20 to 920
with step size of 50. A single calculation is not sufficient to
Frequency determine the optimal sensor number, the selected 𝑀 has been
Mode No. Cycle(s)
(rad/s) (cycle/s) calculated for 10 times to get the average value of the objective
1 1.683166 0.267884 3.732956
2 1.906926 0.303497 3.294929
function. The curves of the average value in three cases are
3 1.926835 0.306665 3.260884 shown in Fig. 6.
4 2.498795 0.397696 2.514487
5 4.009812 0.638181 1.566953
6 4.706711 0.749096 1.334942
7 5.347616 0.851099 1.174951
8 5.394406 0.858546 1.164759
9 7.161397 1.139772 0.877369
10 9.235729 1.469912 0.680313

1558-1748 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2019.2934996, IEEE
Sensors Journal

Sensors-27720-2019 9

0.5
The trend of the results in Fig. 7 (a) is similar to that in Fig.
Case I: BECA 6, which shows that the range of the sensors number in Fig. 7
Case II: BSAS
Case III: DE
(a) is still relatively wide, while that in Fig. 7 (b) is relatively
0.4
stable. So, we can try to get the optimal sensors number from
Fig. 7 (b). Combining with the average value and the economic
0.3
consideration, the number of selected sensors of this bridge is
determined to be 52.
0.2
 Determining the placement of sensors
In order to analyze the performance of the algorithm
0.1
proposed in this paper, the 52 sensor nodes are further
calculated and compared in four Cases:
0
100 200 300 400 500 600 700 800 900 Case I: BECA optimizes the sensor placement of H-Q RB.
Sensor Number(Larger step size) Case II: BSAS optimizes the sensor placement of H-Q RB.
Fig. 6. Average curve of the OFV in Case I to III Case III: BSO optimizes the sensor placement of H-Q RB.
Case IV: DE optimizes the sensor placement of H-Q RB.
As can be seen from Fig. 6, the OFV decreases first and then In the above four Cases, the binary coding of beetles is
increases with the increase of sensor number. This is because, carried out by HDCC. Fig. 8 is the average curve showing the
at the preliminary stage of the OFV curve, there are only few OFV changes as the number of iterations by calculating 20
sensors to be placement, that is to say, the number of “1” in 𝑥 times, and Fig. 9 shows the optimal 𝑀𝐴𝐶 matrix’s bar graph of
is very small in Eq (2). At this time, it is not enough to obtain the best results among 20 calculations in each Case.
more structural modal information to form a good 𝑀𝐴𝐶 matrix,
and the calculated OFV is larger. As the sensor number
increases, more structural modal information may be extracted
and the calculated OFV decreases gradually. However, with the
sensor number increases further, more data will be extracted
from matrix ∅ × and the OFV will increase in another
direction. Thus, it is necessary to achieve a trade-off between
OFV

the number of sensor placement and the OFV through


optimization method. By calculation, the number of sensors is
locked in the vicinity of 70 in Case I and its corresponding OFV
is obviously superior to Case II and III. Therefore, when
searching for 𝑀 value with smaller and micro step size, only
Case I is used for calculation.
2) Determining 𝑀 value with smaller step size
More broadly, research is also needed to determine the Fig. 8. Average curve of the OFV under four Cases
optimum 𝑀 value with smaller and micro step size. The
BECA: 0.019722 BSAS: 0.097157
average value curve of the OFV with the sensor number from
20 to 120 (take 10 as a step size) is given in Fig. 7. (a). Based 1 1

on the results of Fig. 7 (a), Fig. 7 (b) shows the average value 0.5 0.5
curve corresponding to the optimal OFV when the number of 0 0
sensor arrangements is between 40 and 70 (in step size of 1). 10
5
10
10
5
10
Mo 5 Mo 5
de 0 0 e de 0 0 e
Mod Mod

0.15 BSO: 0.052309 DE: 0.046741


Case I: BECA
0.1
1 1
0.05
0.5 0.5
20 40 60 80 100 120
0 0
Sensor Number(Smaller step size) 10 10
10 10
(a) Mo
5 5 Mo
5 5
de 0 0 e de 0 0 e
Mod Mod
0.06 Case I: BECA
Fig. 9. MAC matrix of the best iteration in each Case
0.04
By comparing the average curves of Case I and II/III in Fig.
40 45 50 55 60 65 70 8, it can be seen that the BECA algorithm proposed in this paper
Sensor Number(Micro step size)
(b) improves the convergence speed and results in varying degrees
compared with the other improved BAS algorithms. The reason
Fig. 7. Average change curve of the OFV in Case I
is that the horizontal learning, information sharing, and
evolutionary renewal mechanisms among the beetles greatly

1558-1748 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2019.2934996, IEEE
Sensors Journal

Sensors-27720-2019 10

increase the diversity of individuals, which makes BECA easier TABLE XII
to approach the global optimal solution. Looking closely at the IMPROVEMENT PERCENTAGE OF CASE I COMPARED WITH THE
previous iteration of Fig. 8, we find that BECA converges very OTHERS
fast, which is attributed to the characteristics of BAS and the Case 1 2 3 4
group competition and elimination strategy of BECA. In 1 0 79.70% 62.30% 57.81%
addition, observing the curves of Case I and IV, it can be found
that the search accuracy of BECA is better than that of the From Tables XI-XII, the average and the best results both
classical DE algorithm. The reason lies on the phagocytosis and show that the results of Case I are better than that of the other
elimination strategy which abandon the unimproved beetle Cases. Among the four Cases, the solution stability of Case 1 is
individuals and replace them with the better ones. The also the best. The results quality of the BECA is 79.70% better
phagocytosis and elimination strategy increase the possibility than that of BSAS and 57.81% better than that of DE. Its fine
that the optimal individual reaches the optimal solution, makes solution results are attributed to the cooperation and
the algorithm avoid falling into the local optimum, and competition mechanism among beetle individuals and the
increases the global search ability. appropriate phagocytosis strategy, which makes the excellent
On the other hand, the maximum off-diagonal elements of individuals reach the optimal solution with a higher probability.
the four 𝑀𝐴𝐶 matrices in Fig. 9 are too small to be analyzed. In All the above data show that the algorithm proposed in this
paper is more effective in solving such high-dimensional
order to compare the results of calculation under four Cases
more intuitively, a comparison of the maximum off-diagonal optimization problem and optimizing the bridge sensor
elements in each modal column vector is given below as shown placement.
in Fig. 10. C. Sensor placement of HQ-RB
Finally, the optimal nodes number of the sensor placement of
0.12
Case I: BECA
HQ-RB and its structure description are shown in Table XIII
0.1
Case II: BSAS based on the BECA proposed in this paper. More intuitively,
Case III: BSO
Case IV: DE the specific sensor placement in FEM of the bridge is shown in
0.08 Fig. 11 (the solid dot indicates the location of the optimal sensor
installation).
0.06
TABLE XIII
0.04 No. Lo Str No. Lo Str
1 1 SSB 27 549 SSA
0.02
2 5 SB 28 580 SSA
3 13 SB 29 617 SSA
4 20 SB 30 671 SSA
0
1 2 3 4 5 6 7 8 9 10 5 30 SB 31 679 SSA
Mode 6 36 SMB 32 686 SSA
Fig. 10. Maximum off-diagonal elements in every mode column of MAC 7 40 SM 33 689 SSA
8 61 SM 34 691 SSA
As is shown in Fig. 10, the maximum off-diagonal element 9 69 SM 35 697 SSA
values of each mode column vector in Case I are smaller than 10 77 SM 36 717 SSA
11 251 MCP 37 739 SSA
those in the other three Cases, and the minimum value is 12 252 MS 38 741 SSA
0.019722, which is very close to the ideal value 0. The average 13 265 MPB 39 743 SSA
value and standard deviation of OFV under the four cases by 14 270 SM 40 744 SSA
calculating 20 times are shown in Table XI. At the same time, 15 300 SM 41 746 SSA
16 336 SM 42 748 SSA
the best results of 20 times calculation under each Case, which 17 356 SM 43 751 SSA
just reflects the solving ability of the corresponding Case, are 18 366 SM 44 758 SSA
also shown in Table XI. Further, among the best results for each 19 382 SS 45 761 SSA
Case, the improved percentages of Case I, compared with the 20 386 SS 46 840 MSA
21 393 SS 47 957 MSA
other three Cases, are shown in Table XII, which reflects the 22 397 SS 48 992 MSA
improvement of the BECA algorithm. 23 402 SS 49 1009 MSA
24 410 SSA 50 1114 SMS
TABLE XI
25 509 SSA 51 1128 SMS
STATISTICAL RESULTS AND THE BEST CONVERGENCE RESULTS 26 531 SSA 52 1142 SMS
OF THE FOUR CASES The acronyms in Table XIII are as follows. Lo: Node location
number of the sensor placement in FEM; Str: Structure description;
Case Avg SD The Best Result SSB: Side span bearings; SS: Side span; SMB: Sub-midspan bearings;
1 0.034138 0.007980 0.019722 SM: Sub-midspan; MCP: Main span crossover position; MS: Main
2 0.157302 0.024820 0.097157 span; MPB: Main pier bearings; SCP: Sub-midspan crossover position;
3 0.115689 0.029050 0.052309 SSA: Sub-midspan side arch ribs; MSA: Main span side arch ribs;
4 0.069975 0.015854 0.046741 SMS: Sub-midspan Suspender; MSS: Main span suspender.

1558-1748 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2019.2934996, IEEE
Sensors Journal

Sensors-27720-2019 11

[2] D. J. Thomson, D. Card, and G. E. Bridges, “RF cavity passive wireless


sensors with time-domain gating-based interrogation for SHM of civil
structures,” IEEE Sensors Journal., vol. 9, no. 11, pp.1430-1438, 2009.
[3] T. H. Yi, H. N. Li, and M. Gu, “Optimal sensor placement for structural
health monitoring based on multiple optimization strategies,” Structural
Design of Tall and Special Buildings., vol. 20, no. 7, pp. 881-900, 2011.
[4] C. Yang, Z. Lu, and Z. Yang, “Robust optimal sensor placement for
Fig. 11. Specific sensor placement in FEM of H-Q RB uncertain structures with interval parameters,” IEEE Sensors Journal., vol.
The 52 sensor nodes shown in Table XIII and Fig. 11 18, no. 5, pp. 2031-2041, 2018.
[5] D. C. Kammer, “Sensor placement for on-orbit modal identification and
basically cover the key parts of the H-Q RB, including main correlation of large space structures,” Journal of Guidance, Control
span, side span, sub-midspan bearings, mid-span position, main Dynamics., vol. 14, no. 2, pp. 251-259, 1991.
pier bearings, side arch ribs, suspender, etc., which are [6] D.S. Li, H. N. Li, and C.P. Fritzen. “The connection between effective
successfully applied into the sensitive point and economic independence and modal kinetic energy methods for sensor placement,”
Journal of Sound and Vibration., vol. 305, pp. 945-955, 2007.
selection principle for bridge health monitoring. More [7] T. H. Yi, X. Wang, and H. N. Li, “Optimal placement of triaxial
importantly, compared with the literature [13], a more accelerometers using modal kinetic energy method,” Applied Mechanics
optimized sensor placement result is obtained, which illustrates and Materials., vol. 166-169, pp. 1583-1586, 2012.
the practical effectiveness of the proposed algorithm. [8] T. G. Carne, and C. R. Dohmann, “A modal test design strategy for model
correlation,” Proceedings of SPIE - The International Society for Optical
Engineering., vol. 2460, p. 927, 1994.
VI. CONCLUSION [9] A. Yuan, H. Dai, and D. Sun, “Optimal sensor placement of cable-stayed
bridge using mixed algorithm based on effective independence and modal
Based on the practical problem of optimal bridge sensor assurance criterion methods,” Journal of Vibration Measurement &
placement, this paper integrates the swarm evolution Diagnosis., vol. 29, no. 1, pp. 55-59, 2009.
competitive mechanism and the differential evolution idea into [10] M. Elsersy, T. M. Elfouly, and M. H. Ahmed, “Joint optimal placement,
the basic BAS algorithm, and proposes a new bridge sensor routing, and flow assignment in wireless sensor networks for structural
health monitoring,” IEEE Sensors Journal., vol. 16, no. 12, pp. 5095-5106,
optimal placement method namely BECA. According to the 2016.
experimental results of HQ-RB, BECA has a faster [11] Y. J. Cha, A. K. Agrawal, Y. Kim, and A. M. Raich, “Multi-objective
convergence speed and better accuracy than the other genetic algorithms for cost-effective distributions of actuators and sensors
algorithms. The specific conclusions are as follows: in large structures,” Expert Systems with Applications., vol. 39, no. 9, pp.
7822-7833, 2012.
A. Combining the density information of large bridge sensors, [12] A. R. M. Rao, and G. Anandakumar. “Optimal placement of sensors for
HDCC is used to initialize the swarm individuals. The structural system identification and health monitoring using a hybrid
advantage of this method is that the continuous variables can be swarm intelligence technique,” Smart Materials and Structures., vol. 16,
discretized effectively and the coding scheme consistent with no. 6, pp. 2658–2672, 2007.
[13] J. H. Yang, and Z. R. Peng, “Improved ABC algorithm optimizing the
the distribution density of bridge sensors can be obtained, bridge sensor placement,” Sensors., vol. 18, no. 7, p. 2240, 2018.
which can avoid the centralized distribution of the beetle’s [14] Z. R. Peng, H. Yin, A. Pan and B. R. Peng, “Chaotic monkey algorithm
effective information, increase the diversity of the potential based optimal sensor placement,” International Journal of Control &
solutions of the swarm, and facilitate BECA to search for the Automation., vol. 9, no. 1, pp. 423-434, 2016.
[15] M. R. Senouci, D. Bouguettouche, F. Souilah, and A. Mellouk, “Static
global optimal solution. wireless sensor networks deployment using an improved binary PSO,”
B. In order to solve the high-dimensional discrete optimal International Journal of Communication Systems., vol. 29, no. 5, pp.
problem such as sensors placement optimization, based on the 1026-1041, 2016.
BAS algorithm, the BECA is obtained by extending the single [16] S. K. Gharghan, R. Nordin, M. Ismail M, and J. A. Ali, “Accurate
Wireless Sensor Localization Technique Based on Hybrid PSO-ANN
beetle to a swarm which allows mutation, crossover and Algorithm for Indoor and Outdoor Track Cycling,” IEEE Sensors Journal.,
elimination to occur among individuals. The new algorithm can vol. 16, no. 2, pp. 529-541, 2016.
increase the information sharing among beetle swarm and [17] Y. Sun, W. Dong, and Y. Chen, “An Improved Routing Algorithm Based
strengthen the swallowing ability of the elite one. The search on Ant Colony Optimization in Wireless Sensor Networks,” IEEE
Communications Letters., vol. 21, no. 6, pp. 1317-1320, 2017.
speed and precision of the new algorithm are higher than those [18] X. Y. Jiang, and S. Li, “BAS: Beetle Antennae Search Algorithm for
of BSAS, BSO and DE. It shows that BECA has a strong search Optimization Problems,” International Journal of Robotics and Control.,
ability for the optimization of large dimension array data. vol. 1, no. 1, pp. 1-5, 2018.
C. According to the sensor placement results solved by BECA, [19] R. Storn, and K. Price, “Differential Evolution – A Simple and Efficient
Heuristic for global Optimization over Continuous Spaces,” Journal of
its solution accuracy and specific placement scheme have been Global Optimization., vol. 11, no. 4, pp. 341-359, 1997.
greatly improved compared with reference [13], and can [20] Y. Q. Ni, H. F. Zhou, K. C. Chan, and J. M. Ko, “Modal Flexibility
effectively cover sensitive monitoring points of the bridges. It Analysis of Cable‐Stayed Ting Kau Bridge for Damage Identification,”
shows that BECA can be used for practical reference to solve Computer‐aided Civil & Infrastructure Engineering., vol. 23, no. 3, pp.
223-236, 2010.
the problems like the optimal placement of large bridge sensors. [21] J. M. W. Brownjohn, F. Magalhaes, E. Caetano, and A. Cunha, “Ambient
vibration re-testing and operational modal analysis of the Humber Bridge,”
REFERENCES Engineering Structures., vol. 32, no. 8, pp. 2003-2018, 2010.
[1] Y. Zhou and L. Sun, “Effects of high winds on a long-span sea-crossing [22] Z. Zhu, Z. Zhang, and W. Man, “A new beetle antennae search algorithm
bridge based on structural health monitoring,” Journal of Wind for multi-objective energy management in microgrid,” IEEE Conference
Engineering and Industrial Aerodynamics., vol. 174, pp. 260-268, 2018. on Industrial Electronics and Applications., pp. 1599-1603, 2018.
[23] A. K. Qin, V. L. Huang, and P. N. Suganthan, “Differential evolution
algorithm with strategy adaptation for global numerical optimization,”

1558-1748 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2019.2934996, IEEE
Sensors Journal

Sensors-27720-2019 12

IEEE Transactions on Evolutionary Computation., vol. 13, no. 2, pp. 398- Jianhui Yang received the bachelor’s degree from the
417, 2009. School of Science, Wuhan Textile University, Wuhan,
[24] D. K. Tasoulis, N. G. Pavlidis, V. P. Plagianakos, and M. N. Vrahatis, China, in 2010. He received the master’s degree from
“Parallel differential evolution,” Evolutionary Computation., vol. 2, pp. the School of Electronic & Information Engineering,
2023-2029, 2004. Lanzhou Jiaotong University, Lanzhou, China, in 2013,
[25] Y. Jiang, Z. Q. He, Y. H. Li, Z. Y. Xu, and J. M. Wei, “Weighted Global where he is currently pursuing the Ph.D. degree in
Artificial Bee Colony Algorithm Makes Gas Sensor Deployment Mechanical and Electronic Engineering.
Efficient,” Sensors., vol. 16, no. 6, p. 888, 2016. He has published about five research papers in the
[26] J. Y. Wang, and H. X. Chen, “BSAS: Beetle Swarm Antennae Search journals particularly in the area of optimal sensor
Algorithm for Optimization Problems,” arXiv., vol. 1807, p. 10470, 2018. placement and its applications. His current research
[27] J. Y. Wang, H. X. Chen, Y. Yuan, and Y. Huang, “A novel efficient covers structural health monitoring (SHM), optimal
optimization algorithm for parameter estimation of building thermal sensor placement and signal processing.
dynamic models,” Building and Environment., vol. 153, pp. 233-240,
Zhenrui Peng received the bachelor’s and Ph.D.
2019.
degrees in mechanical engineering, Lanzhou Jiaotong
[28] T. T. Wang, L. Yang L, and Q. Liu, “Beetle Swarm Optimization
University, Lanzhou, China, and control science and
Algorithm: Theory and Application,” arXiv., 1808.00206, 2018. engineering, Zhejiang University, Hangzhou, China, in
1995 and 2007, respectively.
His major research fields include fault diagnosis of
mechanical and electrical equipment, optimal sensor
placement and modal analysis. He presided over two
projects of the National Natural Science Foundation of
China and published over 30 papers up to now, and
over 10 papers have been indexed by SCI and EI.

1558-1748 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

You might also like