1. Introduction
In recent years, international shipping has rapidly expanded as a low-cost way to transport bulk commodities [
1], accounting for over 90% of global trade [
2]. According to the United Nations’ 2023 Review of Maritime Transport report, maritime trade is expected to grow by 2.4% in the post-pandemic era [
3]. However, as the shipping industry flourishes, the environmental issues associated with marine fuels are becoming more apparent [
1]. Vessel air pollution is now recognized as the third primary source of pollution, after vehicle exhaust and industrial emissions. [
4]. Vessels emit significant pollutants such as carbon dioxide (CO
2), nitrogen oxides (NO
X), sulfur oxides (SO
X), and particulate matter (PM) while navigating. Global data indicate that SO
X emissions from vessels make up 10–15% of total human activity emissions [
5]. These pollutants not only lead to ocean acidification [
1,
6] but also cause various respiratory diseases in humans. Studies show that vessel pollution has become a major source of air pollution in port cities, particularly from container vessels powered by high-sulfur fuels exceeding 3.50% sulfur content, which severely pollute nearby waters and air [
7].
Serious environmental and health concerns have compelled the International Maritime Organization (IMO) to take decisive actions. Since 2015, the IMO has established four Emission Control Areas (ECAs) in the Baltic Sea, the North Sea, North America, and the U.S. Caribbean. The organization intends to establish additional Emission Control Areas (ECAs) in various regions to minimize pollution from vessels [
8]. The International Maritime Organization (IMO) now mandates that vessels entering these ECAs utilize fuel containing less than 0.1% m/m sulfur content [
9]. In 2020, the IMO’s Marine Environment Protection Committee (MEPC) set a global limit on sulfur content in marine fuels at 0.50% m/m [
10]. In the same year, the Maritime Safety Administration under China’s Ministry of Transportation declared the “2020 Global Marine Fuel Sulfur Restriction Implementation Plan” along with the “Guidelines for the Supervision and Management of Air Pollutant Emissions from Ships”. According to these guidelines, the sulfur content of fuel used by vessels in inland waterway emission control areas must be no more than 0.1% m/m. Additionally, the sulfur content in the fuel used by ships operating within coastal air pollution control areas and on international journeys in non-emission control areas should not surpass 0.5% m/m [
11]. Despite these regulations, the extra costs of purchasing gas purification devices or using low-sulfur fuels lead some shipping companies to violate these rules to cut costs. Methods include using cheaper, high-sulfur fuels, shutting down pollution control equipment in ECAs, and falsifying vessel logs and fuel consumption records [
1]. Historical data indicate that non-compliance rates can reach as high as 12.3% [
11].
There are two main traditional methods for monitoring air pollution from vessels. The first is manual boarding, which includes taking oil samples, checking oil change records in the logbook, and using a rapid sulfur detector for marine fuel oil. The second method involves real-time tracking of air quality using land-based environmental monitoring stations. However, both methods have drawbacks. They are inefficient, susceptible to environmental influences, and offer limited monitoring coverage. This is particularly problematic during emergencies such as epidemics, when manual boarding and testing become almost impossible.
UAVs are characterized by high flexibility, no geographical restrictions, and high operability [
12]. As UAV technology continues to develop, its use in the shipping industry is becoming increasingly widespread. UAVs are now used for port infrastructure security inspections, maritime oil spill monitoring, ship piloting, maritime smuggling detection, and hazardous chemical inspections at ports [
13,
14,
15,
16,
17]. Additionally, UAVs have started to be employed in vessel air pollution monitoring. For instance, on 20 May 2020, a DJI M210 UAV, named “Rhinoceros V2” and equipped with a system for detecting air pollution, carried out exhaust gas measurements close to the Ganjiang Marine Office of the Yangzhou Maritime Bureau in China [
18]. On 5 January 2021, Zhangjiagang Maritime Bureau used a UAV-mounted gas detector to detect a vessel with excessive sulfur emissions [
19]; on 26 June 2022, Changshu Maritime Bureau used a UAV to monitor air pollution from vessels and inspected and found a vessel with an exhaust sulfur value of 3.0% m/m or more [
20]; on 20 October 2022, the Port of Rotterdam Authority attempted to apply the new Aera3 UAV for monitoring air pollution emissions from vessels [
21]. These instances show that UAVs are a feasible and promising tool for monitoring air pollution from vessels [
7].
However, the above-mentioned studies are still nascent, with several critical issues needing resolution:
- (1)
The energy constraints of UAVs [
22].
- (2)
The limitation of the size of the sea area that can be monitored [
8].
- (3)
The limitation on the number of vessels that can be monitored [
8].
To further optimize the research on monitoring air pollution from vessels using UAVs, it is essential to thoroughly explore and address these issues. This paper proposes a monitoring scheme involving a mothership and multiple UAVs working in concert. In this case, the mothership could be a port sea patrol ship or the other official ship. We have developed a path optimization model for MUCWM monitoring of air pollution from vessels in port waters, aiming to minimize the total monitoring time.
The core focus of this paper is the path planning problem for UAVs collaborating with a mothership to monitor air pollution from vessels. This involves three main aspects: (1) optimizing the monitoring paths for multiple UAVs, (2) developing a synergy mechanism between the mothership and multiple UAVs, and (3) optimizing the travel path of the mothership. We have created a path planning model for MUCWM monitoring designed to minimize the total time needed to complete the monitoring task.
The purpose of this study is to explore how the mothership and UAVs can monitor air pollution from port vessels in a collaborative manner. Based on the self-constructed path optimization model for the joint operation of the mothership and UAVs to monitor the air pollution from port vessels and solved by using the improved adaptive differential evolution (IADE) algorithm, the optimization of the monitoring paths of the UAVs, the collaborative work between the mothership and the UAVs, and the optimization of the travel paths of the mothership are implemented. In the case study, the effectiveness of the proposed model and algorithm is verified by using the vessel position data from the Automatic Identification System (AIS) of Ningbo Zhoushan Port. The experimental results show that our algorithm can effectively plan the traveling paths of the mothership and UAVs in collaborative operations, which is of great theoretical significance and application value for the study of UAVs monitoring air pollution from vessels in port.
The contributions of this study are significant: (1) We propose a novel path planning model for UAV collaborative mothership monitoring of air pollution from vessels. This model optimizes paths for both UAVs and the mothership working together. (2) We employ the IADE algorithm to solve the model. We have verified the superiority of this algorithm through performance comparisons and sensitivity analysis. (3) We conducted numerical experiments using vessels in the sea area around Zhoushan Port in Ningbo, China. We gathered position data for 100 ships from the Automatic Identification System (AIS) to validate the feasibility of our algorithm.
The innovations of this paper are as follows: (1) Model innovation: This study proposes a path optimization model for multi-UAV collaborative with mothership (MUCWM) surveillance. The model also optimizes the collaborative monitoring routes of the mothership and multiple UAVs, and this method can effectively reduce the overall operation time and improve monitoring efficiency. Therefore, the solution strategy of path planning in this paper can provide a useful reference for other scholars engaged in related research. (2) Algorithm innovation: In this paper, we propose an improved adaptive differential evolution (IADE) algorithm based on the collaborative operation of the mothership and multiple UAVs, which successfully solves the path planning problem of UAVs collaborating with the mothership to monitor the air pollution of port vessels. After comparative analysis, the algorithm demonstrates strong search capability and convergence speed, which significantly improves the efficiency and performance of the problem, and thus better meets the requirements of practical applications. The algorithm is especially suitable for harbor environments, where fast problem-solving is required.
The remaining sections of this paper are structured as follows:
Section 2 outlines the related literature. The methodology for establishing the MUCWM model and implementing the algorithm is detailed in
Section 3 and
Section 4, respectively.
Section 5 presents the experiments and the results, with the conclusion following in
Section 6.
4. Model Solution Based on Improved Adaptive Differential Evolution Algorithm
In this paper, an improved DE algorithm is utilized to solve the model. In order to make the DE algorithm more applicable to the problem scenarios in this paper, adaptive and arithmetic improvements are made on the basis of the inherent advantages of the original DE algorithm to increase its performance and ability to solve the problems in this paper, and the improved algorithm is referred to as the improved adaptive differential evolution (IADE) algorithm.
4.1. Adaptive Parameter Settings
- (1)
Adopt the lifespan mechanism and extinction mechanism for population NP.
This mechanism regulates population size and maintains diversity. The lifespan mechanism operates as follows: Initially, the age of all individuals is set to zero, aligning with the number of iterations. In each generation, when a new individual is created, its lifespan is assigned based on the value of its fitness function. An individual is removed from the population once its age exceeds its predetermined lifespan. The formula for calculating lifespan is as follows:
Among them,
represents the average value of the fitness function value of the current population and
and
represent the minimum value and maximum value of the fitness function in the current population, respectively. The fitness function value of the algorithm in this article is the minimum value of the total time consumption of the monitoring task. That is,
.
represents the maximum value that the life span can reach, that is, the maximum number of iterations, and
represents the minimum value that the life span can reach. At the same time,
. The extinction mechanism strategy is as follows: When the population fitness function value does not change for
times in a row, it means that the population in the current iteration may fall into a local optimum. At this time, the extinction mechanism strategy can be used to increase the diversity of the population. The parameter
is the extinction coefficient. Generally, the larger the extinction coefficient, the higher the probability of an individual being eliminated or reset; the smaller the extinction coefficient, the higher the probability of an individual being retained. This article sets the extinction coefficient
to ensure the timely generation of new individuals. After an individual becomes extinct, a new population of individuals needs to be generated. Here, elite individual replication is used to generate a new population of individuals. The specific formula is as follows:
is the replication rate. This article sets the replication rate to ensure the majority of elite individuals. is the maximum value of the population size, and is the current population size.
When the lifespan mechanism comes into play, the diversity of the population may decline sharply, which increases the risk of the population falling into a local extreme. Once this situation occurs, the extinction mechanism will be activated to enhance the diversity of the population by eliminating some individuals and using the replication strategy of elite individuals. This mechanism helps prevent premature convergence of the population, thereby improving the search efficiency and convergence of the algorithm. Therefore, the lifespan mechanism and the extinction mechanism collaborate with each other to realize the adaptive adjustment of at different stages.
- (2)
Introducing an adaptive parameter adjustment strategy of Beta distribution for scaling factor and crossover probability .
Beta distribution is a common continuous probability distribution in the field of probability statistics, and its value range is limited to the [0, 1] interval. The Beta distribution is often used to describe the probability distribution characteristics of random variables, such as describing the prior distribution of probability, proportion, or parameters. In many fields, such as machine learning and mathematical statistics, Beta distribution plays an important role. The specific form of the Beta distribution is given by its probability density function, which describes the probability distribution characteristics of the Beta distribution in detail. The probability density function of the Beta distribution is as follows:
where
is the value of the random variable,
and
are the two shape parameters of the Beta distribution, and
is the Beta function:
. When
, the Beta distribution degenerates into a uniform distribution. When
and
are large, the Beta distribution presents a bell-shaped shape concentrated in the middle. When
and
are small, the Beta distribution presents a long-tail shape.
This article introduces a method of dynamically adjusting
and
; that is, the dynamic adaptive adjustment method of Beta distribution is introduced to improve the performance of the algorithm. Specifically, the Beta distribution is used to generate random numbers, and then the generated random numbers are mapped to the value ranges of
and
. Through this dynamic adjustment method, the DE algorithm is able to select appropriate values of
and
at different stages. In the initial state, we set
so that the probability distribution function presents a (0, 1) uniform distribution. During the population evolution process, the parameters
and
change based on the following Beta probability distribution formula:
where
is the number of individuals with better performance in the current
(or
) (such as successful evolution or better fitness function value), and
is the number of individuals with poorer performance in the current
(or
) (such as failure to evolve successfully or making the fitness function value better, which does not make the fitness function value more optimal). We substitute the number of individuals with better performance
and the number of individuals with poor performance
in the
generation population into Equation (16), respectively, to calculate the values of
and
in generation
, and make constraint boundaries for
and
(
,
), so that
and
can change adaptively according to the iteration results, effectively improving the convergence speed of the algorithm and the diversity of the population. The calculation process of adaptive
and
is shown in
Figure 2.
4.2. Hybrid Mutation Improvement Strategy
The traditional differential evolution algorithm has limitations in mutation improvement strategies, which limits its versatility. To address this problem, this paper proposes a mutation strategy based on “DE-BB” (differential evolution bare bones) to optimize mutation operations. This strategy is an innovative combination of the backbone particle swarm algorithm and the differential evolution algorithm. Experimental results show that this strategy effectively utilizes the advantages of the backbone particle swarm optimization algorithm in the local depth search, thereby improving the convergence accuracy of the differential evolution algorithm. In the strategy, we incorporate elements of the classic differential mutation and combine them with the adaptive improvement strategy of the F-value to generate individual trial vectors. The specific mutation improvement strategy formula is as follows:
In the formula, are random integers that are different from each other and randomly selected from the population set , and is the scaling factor, where , where is also a random decimal in the interval [0, 1]. is the historical local optimal value of the individual, and is the optimal individual value of the current population.
4.3. Hybrid Cross Improvement Strategy
In traditional differential evolution algorithms, the binomial crossover method (binomial) is often used to cross the contribution vector
generated by the mutation operation with the current target vector
to generate a test vector
. Subsequently, the algorithm competes between the trial vector
and the target direction
through a greedy selection mechanism, and only the solution vector with better fitness is retained to enter the next round of iteration, while the worse solution vector is eliminated. However, when the contribution vector
is already very close to the global optimal position, the crossover operation may destroy its complete solution structure. If the test vector
can fully absorb the information of vector
, it will help increase the diversity of potential solutions and reduce the risk of the algorithm falling into a local optimum. To this end, this article quotes the crossover strategy
of the multi-selection mechanism and combines it with the adaptive improvement strategy of
to obtain the crossover improvement strategy formula as follows:
where
represents the mixed crossover probability threshold;
represents the value of the
dimension of individual
, calculated by the traditional binomial crossover method.
represents the choice probability of performing a binomial crossover or a fully absorbing crossover. According to the experiment of the proposer of this strategy, when the value of parameter
is set in the range of [0.4, 0.6], the average convergence speed and convergence accuracy of the algorithm are better than those in other intervals. Therefore, this article sets the parameter
to 0.5.
4.4. IADE Algorithm Solution Steps
4.4.1. Codec Strategy
The encoding method used in this article is natural number encoding. Each individual is composed of two chromosomes. The natural number
in the first chromosome represents the serial number of the UAV, that is,
, where
represents the number of monitored vessels; the natural number
in the second chromosome represents the serial number of the monitored vessel; that is,
also represents the number of monitored vessels. Among them, the upper and lower correspondence between chromosomes represents the vessel to be monitored by the UAV. The chromosome encoding strategy is shown in
Figure 3. The UAV with serial number 1 corresponds to the serial number of the monitored vessel {3, 8, 1, 10}, which means that the order of the vessels monitored by UAV No. 1 is {3→8→1→10}.
Decoding requires designing an appropriate decoding solution based on the characteristics of the problem and converting the encoding after the algorithm is executed into the solution to the problem. The specific steps are as follows:
Step 1: According to the order in which each UAV monitors the vessel, consider the UAV endurance time
. If, before the UAV goes to the location of the next monitored vessel,
, that is, the UAV cannot return to the supply point after monitoring the next vessel, it is necessary to insert a new supply point gene after the gene of the current vessel position, as shown in
Figure 4.
represents the remaining range of the UAV,
represents the distance between the two positions
and
,
represents the current location of the vessel being monitored by the UAV, and
represents the position of the next vessel that needs to be monitored, and
represents the location of the previous charging point of the UAV;
represents the monitoring time.
Step 2: Select a charging point that meets the conditions for the UAV based on the remaining power of the UAV. If
is satisfied, add the current charging point to the gene sequence and add the feasible charging points to the UAV monitoring sequence that is added to the UAV feasible monitoring sequence set
, as shown in
Figure 4. Finally, a recovery and supply point need to be added at the end of the genetic sequence of each UAV. After the UAV completes its monitoring mission, the mothership is used to recover it.
Step 3: According to the time window constraint in
, that is,
, select the charging point driving path that meets the mothership speed constraint, as shown in ①→②→⑤ in
Figure 5 and the ①→④→⑤ sequence.
represents the maximum time point when the UAV reaches supply point
and hovers above it,
represents the current time point, and
represents the speed of the mothership.
Step 4: Calculate the total duration of the monitoring task based on the charging point driving path of the mothership, return the fitness function value, and end.
Table 3 gives the solution form of MUCWM path optimization.
4.4.2. Algorithm Process
The process of implementing the IADE algorithm is as follows:
Step 1: Initialize parameters, including population size , extinction coefficient , elite individual replication rate , individual component value range , mutation scaling factor , shape parameters , , crossover probability , mixed crossover probability threshold , the maximum number of iterations .
Step 2: Generate an initial population, with each individual containing parameters, and the value range of the parameters is .
Step 3: Decode the initial population, multiply the first dimension of the individual by , and perform a rounding operation to determine the set of vessels monitored by the UAV. represents the number of UAVs and corresponds to the second-dimension parameters of the individual. Attach a set of sequences to the position, sort the parameters from small to large, retain the attached serial number in the sorting, and use this sorting as the order of the UAV monitoring vessel collection; means the number of vessels to be monitored.
Step 4: Let represent the first iteration.
Step 5: Let represent the first individual.
Step 6: Mutation operation: randomly select 4 individuals , , , and in the population, and generate mutated individuals according to the mixed mutation improvement strategy of Formula (19) .
Step 7: Crossover operation: according to the hybrid crossover improvement strategy of Formula (19), generate new test individuals .
Step 8: Greedy selection: calculate the fitness function value (minimum value of monitoring task completion time); if the fitness function value , update the individual of the previous generation , otherwise the individual of the previous generation is retained.
Step 9: If , let and return to step 5, otherwise go to step 10.
Step 10: Adaptive parameter adjustment. According to the fitness function value, use Formulas (15) and (16) to calculate the adjusted population; use Formula (18) to calculate the adjusted mutation scaling factor and crossover probability .
Step 11: If , , go to step 4, otherwise go to step 11.
Step 12: Output the historical optimal solution and the algorithm ends.
The pseudo-code of the improved adaptive differential evolution algorithm is shown in Algorithm 1.
Algorithm 1: Improved adaptive differential evolution algorithm |
1 Begin |
2 Initialize related parameters: , , , , , , , , , |
3 Generate initial population |
4 Decoding |
5 for in |
6 for in |
7 = The optimal individual and four random individuals generate mutant individuals through Formula (17) |
8 Boundary constraint check after mutation |
9 = Generate new test individuals through crossover using Formula (18) |
10 if then |
11 |
12 end if |
13 end for |
14 = Calculate the adjusted population by Formulas (14) and (15) |
15 , = scaling factor and crossover probability adaptively adjusted through Formula (16) |
16 end for |
17 Output the optimal solution , the optimal individual , the algorithm execution time and the algorithm convergence chart |
18 Finish |
The flow chart of the algorithm is shown in
Figure 6.
4.5. Algorithm Design for MUCWM Path Optimization Problem
This section mainly describes the basic ideas for solving the MUCWM path optimization problem in three aspects: optimization of UAV monitoring paths, collaboration between motherships and UAVs, and optimization of mothership monitoring paths.
4.5.1. UAV Monitoring Path Optimization
First, the mothership carries multiple UAVs and sets off from the port starting point at the same time. The UAVs will take off from the mothership at the beginning to perform monitoring tasks. At this time, monitoring needs to be planned for multiple UAVs. To achieve this goal, two problems need to be solved: (1) How to determine the collection of UAVs to monitor vessels; (2) how to determine the order of UAVs to monitor vessels.
Ignoring the power constraints of the UAV, that is, treating the UAV power as infinite, we randomly assign vessel collections and monitoring sequences to the UAV. The specific strategy is as follows: First, add an extra dimension
sequence to the test individual
. Secondly, after multiplying the values of the first dimensional elements of the test individual
by
respectively, since the value range of the individual elements belongs to
, and after rounding off the elements and adding 1, the value range of the first dimension element value of
will become
,
, where
is the number of UAVs, and then we sort the second-dimensional elements of
from large to small. During the sorting process, the corresponding positions of the additional sequences remain unchanged. Through the above two strategies, the random allocation of the UAV monitoring vessel collection and monitoring sequence is achieved. The random allocation algorithm for UAV monitoring is shown in Algorithm 2.
Algorithm 2: Random allocation algorithm for UAV monitoring |
1 np.round |
2 np.arangereshape | Generate additional sequences |
3 The trial individual formed by appending to the third dimension of |
4 | Count different values in |
5 np.unique |
6 for in enumerate |
7 = Split according to the value of unique_values |
8 end for |
9 Sequential allocation strategy for UAV monitoring of vessels: |
10 for in |
11 = Sort according to the value of |
12 end for |
Figure 7 shows a random allocation strategy in a scenario with 2 UAVs and 5 monitored vessels. The sequence of UAV monitoring vessels is divided into the following:
,
.
4.5.2. Collaboration between the Mothership and UAVs
Since, in the MUCWM path optimization problem, the mothership plays the role of mobile resupplying for UAVs, this means that the mothership needs to meet the UAVs at the supply point and replenish power for the UAVs in time. Therefore, the mothership and the UAVs need to face the problem of human–machine collaboration, which will be realized based on the selection of UAV power supply points.
First, divide the monitored sea area into equal-area grids, and use the midpoints of all grids as the set of candidate power supply points for UAVs:
, where
is the divided grid quantity. Based on the value of the individual
after the random allocation strategy is executed, and based on the maximum endurance time of the UAV
, it is judged whether it is necessary to insert a power supply point for the UAV between the two vessels to be monitored during the monitoring process. Take the UAV
in
Figure 8 as an example to describe the supply point selection strategy, taking into account the UAV power constraints during the UAV monitoring of vessels.
In the initial case, set as the flag that shows that the UAV is fully charged and starts to perform monitoring tasks, and loop through each UAV in sequence; then select the supply points it needs to pass during its voyage. Let ; represents the time consumed by the current UAV to perform the task, loop through the vessel nodes that the UAV needs to monitor, and judge whether it requires the insertion of feasible supply points between two consecutive monitored vessels. The specific strategies are as follows:
(1) Determine whether the vessel node is the first node monitored by the UAV when it departs from the port, that is, whether it meets
and
at the same time. If the conditions are met, it is judged that the UAV flies to the first node. The vessel node to be monitored will return to the port
after monitoring it, that is,
will be greater than or equal to
, where
represents the UAV flight speed; if the conditions are met, the UAV will fly to the location of
and monitor it, as shown in
Figure 8, the process of UAV
flying to
. In the algorithm, it is necessary to record the remaining endurance time of the UAV
, the UAV monitoring sequence
and
and set the flag to 0, indicating that the next vessel node is not the first vessel in the UAV monitoring sequence. It is worth mentioning that since the UAV may have multiple candidate power supply points that meet the conditions in subsequent monitoring tasks, the latter node will be calculated from the previous supply point. Therefore, the solution space is an unmanned power supply point. The first monitored vessel node in the UAV monitoring path is a tree structure array with the root as the root. Therefore, the current vessel node
is assigned to
as the root node of the UAV monitoring path
.
(2) Determine whether the next vessel node is the last vessel node in the UAV monitoring sequence. If not, we need to determine whether the UAV can return to the UAV after flying to the next node and monitoring the vessel. Then we determine the last power supply point of the machine, that is, determine whether
is greater than or equal to
; if it can, the UAV will fly to the next supply point. In
Figure 8, the UAV
flies from the location of
to the location of
for monitoring. The reason for using the above strategy to determine whether the UAV needs to replenish electrical energy is that it must ensure that the UAV has a candidate supply point where it can land. Even if the UAV reaches the next vessel node and completes the monitoring task, the UAV’s remaining power is not satisfied. The recharge point with electric energy constraint can also be returned to the previous recharge point for charging. This strategy ensures that the UAV can land and charge at any point in time and will not run out of power and have nowhere to land during the monitoring mission. If it is not satisfied, it means that the remaining range of the UAV is not enough to support its subsequent monitoring tasks. At this time, all electric energy supply points
are traversed, and the unmanned machine selects feasible power supply points. If
is greater than or equal to
, then
is added to the solution of the tree structure, and branches may appear, as shown in
Figure 8. The area in
, the supply point in the area, represents the position of the supply point that can be reached from the location of the vessel
to be monitored under the constraints of UAV endurance. Then we update
and
. At this time,
=
+
, where
represents the charging time of the UAV, and we let flag = 1. Finally, the time consumption
from each time the UAV passes through the power supply point is added to the time window set at
in order to solve the path planning of the mothership. In
Figure 8, the UAV
is traveling from the location of the vessel
to the vessel
. Between the locations, the
supply point needs to be inserted. At this time, the next node is the first vessel node monitored by the UAV after it is fully charged and takes off from the mothership; that is, it satisfies flag = = 1 and j > 1. Then it is judged whether the UAV can return to the location of the current supply point after flying to the first vessel node to be monitored and monitoring it, that is,
; if satisfied, then we update
and
, and set the flag to 0. As shown in
Figure 8, the UAV
flies from
to the vessel node
to be monitored.
(3) If the next node is the last node in the UAV monitoring sequence, select an electric energy supply point for the UAV that satisfies the remaining power constraint (recovery point: the UAV will be recovered by the mothership at this node, and the mothership returns to the port together and ends the monitoring task), then is added to the solution of the tree structure, branches may appear, and then , , and are updated. At this time, .
The algorithm for selecting UAV power supply points is shown in Algorithm 3 (the first node time monitored when the UAV takes off from the port).
Algorithm 3: The algorithm for selecting UAV power supply points |
1 Begin |
2 flag = 1 |
3 for i in |
4 |
5 for j in |
6 | The first node monitored when the UAV takes off from the port |
7 if flag = =1 and j = =1 |
8 if |
9 = |
10 = + |
11 |
12 Add to solution structure |
13 end if |
14 end if |
15 end for |
16 end for |
17 Output , , |
18 Finish |
4.5.3. Mothership Traveling Path Optimization
The results in
Section 4.5.2 are used as parameters and input into the optimization algorithm of the mothership monitoring path, which is illustrated in
Figure 9. The optimization strategy of the mothership monitoring path is as follows:
(1) First, select one of the feasible monitoring paths of each UAV for calculation. Sort
in time order according to
to get
, which is the sequence in which the mothership passes through each supply point. As shown in
Figure 9, the navigation path of the mothership
is
. Then assign the supply point corresponding to the sorted
to the set
. Traverse the elements in
W, and use the speed constraint of the mothership to determine whether the mothership can reach the corresponding supply point to charge the UAV before the UAV’s power is exhausted, that is, whether
is less than
; if satisfied, add
to the mothership’s travel path
, and update
, as shown in
Figure 9, each time the mothership moves from the current supply point to the next supply point, such as
.
(2) Since the mothership and the UAV can wait for each other at the supply point, it is necessary to determine the order in which the mothership and the UAV arrive at the supply point and then update the time. Specifically, it needs to be judged whether is greater than , where represents the time when the mothership arrives at the supply point. If it is satisfied, it means that the mothership arrived after the UAV arrived at the supply point . At this time, needs to be updated, = , and the UAV waits for the mothership’s time to be added to the time consumption.
(3) After executing a set of data, assign to . represents the completion time of the monitoring task. represents the mothership passing the last supply point. This formula represents the mothership passing through the last supply point. The time when the vessel-borne UAV returns to the port is added to and then assigned to . After all the data have been traversed, the smallest is selected as the optimal solution of the algorithm. Finally, output and .
The pseudocode of the mothership traveling path optimization algorithm is shown in Algorithm 4.
Algorithm 4: The Mothership Traveling Path Optimization Algorithm |
1 Begin |
2 Input , , |
3 for i in |
4 for j in |
5 … |
6 for n in |
7 = Sort in time order according to |
8 W = corresponding supply point sequence after sorting |
9 = 0 |
10 for k in W |
11 if < |
12 |
13 = |
14 end if |
15 end for |
16 + |
17 end for |
18 … |
19 end for |
20 end for |
21 |
22 Output |
23 Finish |
5. Experiment and Analysis
In this section, in order to prove the effectiveness of the improved adaptive differential evolution algorithm, this article uses Python 3.12.0 to implement the solution of the algorithm. This experiment was run on a computer with an Intel (R) Core (TM) i5-7200U CPU at 2.50 GHz to 2.70 GHz and a memory of 8 GB.
5.1. Data Preparation and Parameter Settings
This paper conducted a numerical experiment on the vessels entering and leaving the port in a certain port area of Ningbo Zhoushan Port. At 17:40 on 3 April 2023, 100 vessels in the above port were obtained from the automatic identification system (AIS). The position data of vessels sailing in the sea area are partly shown in
Table 4. The table gives the vessel serial number, vessel name, vessel type, vessel location (including longitude and latitude), and the update time of the current position. In order to more intuitively display the position of the vessel in the port sea area, a real-time satellite image of the selected sea area is given, as shown in
Figure 10. The area within the dotted line is the monitored port sea area.
Before parameter setting, we consider some more realistic factors, such as whether the sea state and weather environment have any influence on the path optimization problem of air pollution monitoring of port vessels by shipboard UAVs.
First of all, the performance and stability of the current DJI UAV are very high, and as an example, the wind resistance of the DJI Warp M300 RTK model UAV can reach 12 m/s, i.e., level 6. According to the survey, in some open sea areas, such as the central part of the ocean or some particularly windy sea areas (such as the roaring westerly wind belt in the South Atlantic Ocean), the average wind speed may be between 7 m/s (about 25 km/h) and 10 m/s (about 36 km/h). Wind speeds are somewhat lower in harbor-based waters, so the monitoring mission is largely unaffected by winds at sea. Secondly, wave heights average a few centimeters to 20 m and can only reach 30 m in very extreme natural conditions [
49]. However, when monitoring vessels, UAVs need to fly into the vessel’s plume to do so, and the average vessel’s air height (i.e., the height of a vessel in the ocean between sea level and the very top of the vessel) can vary greatly depending on the type and use of the vessel, and the following are the approximate ranges of air heights for some common types of vessels: container vessels: larger container vessels, usually in the range of 40 to 60 m; bulk carriers: medium-sized bulk carriers may be between 20 and 30 m; tankers: roughly between 30 and 50 m; cruise vessels: usually over 50 m; and medium-sized passenger vessels: these are generally between 20 and 30 m. The vessel’s plume is at a position even higher than the highest point of the vessel, so under normal circumstances, the plume height is higher than the waves, and therefore, the UAV monitoring process will not be affected by the waves as long as it maintains the flight height at the plume position.
The default values of the parameters involved in the experiment are shown in
Table 5.
The description of the main parameters in
Table 5 is as follows:
Number of UAVs: in the scenario of the vessel exhaust gas measurement trial carried out by the Hanjiang Maritime Safety Department of the Yangzhou Maritime Safety Administration, the number of UAVs was established at three.
Number of equal area divisions in the sea area (number of candidate supply points): In the sea area, it is not advisable to select too many candidate supply points, because too many supply points will greatly increase the running time of the algorithm; at the same time, it is not advisable to select too few candidate supply points, because there may be no solution if there are too few supply points. Therefore, this article divides the sea area into nine equal areas; that is, there are nine candidate supply points.
Port starting point coordinates: set according to the coordinates of a central pier in the port.
Maximum UAV flight speed and maximum UAV endurance time: Based on the UAV produced by DJI that can be used in water conservancy: M300 RTK flight parameters: 55 min maximum flight time, 23 m/s maximum flight speed, and 12 m/s wind resistance. Taking into account the impact of the environment and special circumstances, the maximum flight speed of the UAV is set to 18 m/s, and the maximum endurance time of the UAV is set to 45 min.
Average monitoring time of each vessel: based on previous monitoring experience, the time for UAV monitoring of each vessel is set to 3 min [
50].
UAV charging time: In order to save the time required for the total monitoring task, the battery is directly replaced on the mothership to replenish the UAV’s power. Therefore, the battery replacement time is set to 2 min.
The maximum sailing speed of the mothership: Calculated based on the previous average speed of China Maritime Safety Administration port patrol vessels of 20.13 knots. After converting the unit into m/min, it is 621 m/min.
Table 6 shows the parameter settings in the improved adaptive differential evolution algorithm.
5.2. Solution to MUCWM Path Optimization Algorithm Based on IADE Algorithm
According to the actual geographical location of the port, this article scales the monitored sea area to the
plane rectangular coordinate system of 200 × 200 to run the algorithm of this article. Following scaling, the initial positions of the 100 vessels to be monitored, one mothership, and three UAVs were determined, represented by (x, y) coordinates. These data are partially displayed in
Table 7. Initially, the three UAVs and the mothership share the same location.
Figure 11 visually illustrates the initial locations of the vessels to be monitored, the mothership, and the UAVs.
The data are substituted into the improved adaptive differential evolution algorithm to solve the problem.
Figure 12 shows the MUCWM path optimization results obtained in the above scenario. Different colors represent different UAV flight paths, and the dotted line is when the UAV goes to the supply point for charging. Since the figure involves many line segments, the path of the mothership will be displayed in the table.
Table 8 shows the specific results of path optimization for the mothership and three UAVs.
5.3. Analysis of Results
5.3.1. Algorithm Performance Comparison
In order to verify the accuracy of the algorithm in small-scale calculation examples, the results of IADE solving this problem in small-scale cases are compared with the Gurobi running results. We set the problem scales,
, to conduct 12 independent experiments for each scale, and we recorded the GAP between the algorithm and Gurobi in each experiment, where GAP = (f(IADE) − f(Gurobi))/f(Gurobi) × 100%. It is stipulated that if Gurobi runs for more than 30 min and still does not obtain a result, it is considered that Gurobi cannot obtain the solution to the calculation example. The IADE algorithm records the results and running time of each generation and selects the running time of the generation with the smallest difference from the Gurobi result. The comparison results are shown in
Table 9.
It can be seen from the table that, in the independent experiment with , the GAP value between IADE and the accurate result is less than 5%, and the solution speed has been greatly improved. Although Gurobi can obtain accurate results, it takes a long time to solve, so the comparison with the Gurobi results proves the effectiveness of this algorithm in solving the collaborative routing optimization problem of UAVs and motherships, and it has practical application significance in real-life scenarios.
When solving the model, in order to verify the superiority of the improved adaptive differential evolution algorithm compared to the traditional differential evolution algorithm, this paper compares the parameters of the adaptive improved differential evolution algorithm with the basic differential evolution algorithm. The group size and the number of iterations are set to
= 100 and
= 300, respectively. The resulting operation iteration diagram is shown in
Figure 13.
From
Figure 14, it can be seen that the DE algorithm, after the adaptive improvement strategy, shows a trend of being better than the traditional DE algorithm in terms of convergence speed and optimal solution performance.
On the basis of the adaptive parameter differential evolution algorithm, after adding the hybrid mutation improvement strategy and the hybrid crossover improvement strategy, the test example is also run when
= 100 and
= 300, and the resulting convergence diagram is shown in
Figure 14. It can be seen from the figure that although the convergence speed has dropped slightly, the improved algorithm has strong search capabilities and performs well in manifesting the optimal solution. Therefore, the overall solution performance of the improved adaptive differential evolution algorithm is higher than that of the adaptive differential evolution algorithm.
We took the test case data from Shen et al. [
12], as shown in
Figure 15, and adjusted the parameters to the values in the article, as shown in
Table 10. Since the vessel distribution is relatively concentrated, the number of candidate supply points can be increased to improve the efficiency of the algorithm. Since Shen et al. [
12] did not consider the UAV charging time, the UAV charging time was set to 0. Bringing the adjusted data into the algorithm of this article, the result is 116.76982 min, which is smaller than the optimal solution of 126.69466 min in the article. The optimization performance of the improved adaptive differential evolution algorithm is further demonstrated.
Figure 16 shows the MUCWM path optimization results obtained by using the algorithm of this article to execute the test case data in Shen et al. [
12].
5.3.2. Sensitivity Analysis
Figure 17a illustrates that with an increasing number of UAVs, the overall time for MUCWM monitoring decreases, although the rate of reduction in total monitoring time gradually diminishes. Thus, beyond a certain number of UAVs, further increases will not significantly decrease the total duration of the MUCWM monitoring task.
Figure 17b reveals that as the number of UAVs grows, the total execution time of the algorithm consistently increases, suggesting a roughly positive correlation between the number of UAVs and the algorithm’s running time.
Figure 18a demonstrates that as the number of grid divisions rises, there is a decrease in the total MUCWM monitoring time, particularly noticeable when the grid count increases from 1 to 9. However, with further increases in grid divisions, the rate of reduction in the total monitoring task time starts to taper off. Hence, beyond a certain number of grid divisions, the advantages of further increases in reducing the overall time of the MUCWM monitoring task become marginal.
Figure 18b shows that as the number of grid divisions grows, the total execution time of the algorithm also increases, with the rate of increase accelerating, suggesting that the number of grid divisions significantly affects the algorithm’s running time.
Figure 19 indicates that extending the maximum endurance time of the UAV has a minimal effect on the total duration of the MUCWM monitoring task. This minimal impact is largely due to the introduction of the mothership serving as a mobile power supply for the UAVs, which significantly mitigates the power limitation issues of the UAV.
Figure 20 additionally reveals that, while increasing the sailing speed of the mothership reduces the overall monitoring time of MUCWM, this reduction tends to plateau. Thus, after the mothership’s speed reaches a certain level, further increases in speed yield diminishing returns in reducing the total monitoring time.
Figure 21 illustrates that, with varying numbers of UAVs, when the number of vessels to be monitored rises from 30 to 150, the algorithm’s running time increases by 2487.962 s. Notably, when the count of vessels monitored in the port surpasses 120, there is a significant spike in the algorithm’s running time. Consequently, the algorithm exhibits improved computational efficiency when the number of vessels monitored remains below 120.
In particular, we discussed the impact on the total monitoring time in the case of leakage of 5–10% of the monitored vessels, and 10 sets of experiments were conducted for each value of leakage, and the results obtained after averaging are shown in
Figure 22. From the figure, it can be observed that the total monitoring time tends to decrease as the number of missed vessels increases, and we find that the impact on the total monitoring task is more significant when it is reduced to 10%. This provides a compromise for the port; the port can tradeoff between the time of the total monitoring task and the coverage of the monitored vessels and can choose to give up a small portion of the vessels to obtain a shorter monitoring time, and this analysis provides a certain reference value for the monitoring program of the vessels in the port’s sea area.