1. Introduction
The condition of the road surface has essential impacts on transportation safety. The road administration needs to regularly detect and measure the road distress to maintain the road quality and enhance its safety [
1]. Currently, road conditions are mainly monitored by manual inspection; however, such a strategy is both time-consuming and labor-costly [
2]. Given the light weight, the high mobility and the low cost of the multirotor unmanned aerial vehicles (UAVs), using a UAV swarm to inspect roads is one of the main technologies in road detection [
3]. However, current methods that use UAV swarms to inspect roads still have many limitations in practical applications, such as the lack of or difficulty in route planning, the unbalanced utilization rate of the UAV swarm and the difficulty of the site selection for a distributed droneport. Thus, route planning for the UAV swarm to inspect roads and the site selection for the distributed droneports are the keys to achieve fully autonomous UAV swarm inspection of roads.
The route planning of the UAV swarm mainly includes two categories, i.e., UAV swarm real-time route planning and UAV swarm offline route planning. The former refers to planning and updating the routes when the UAV swarm is working, and the latter refers to planning the complete routes before the UAV swarm moves. Pertaining to the UAV swarm real-time route planning, AI Duhayyim, M. et al. [
4] applied the 6G UAV communication routing planning technology to enable the UAVs to efficiently collect data and plan the routes. Lee, M.-T. et al. [
5] applied the edge computing system to provide the UAVs a higher level of autonomous control, which enabled the UAVs to be more flexible to automatically adapt to the dynamic changes of the environment. To optimize the route planning of the UAVs in the peak traffic period, Wang, K. et al. [
6] introduced the concept of the team orienteering arc routing problem with time-varying profits and then proposed a route planning method that considered the spatiotemporal variations in monitoring demand. A bio-inspired route planning algorithm was leveraged in [
7] to address the dynamic obstacle avoidance route planning problem for the UAVs when the environment maps are unknown. However, the literature [
8] points out that drones are equipped with a limited battery supply and onboard computational power, which means they are unsuitable for detecting damage on roads in real-time. Similarly, conventional road inspection route planning is unsuitable for real-time calculation. Pertaining to the UAV swarm offline route planning, Cho, S-W. et al. [
9] proposed a method that engenders a search path to cover all nodes with the minimum computation time for a fleet of heterogeneous UAVs. With the objective defined as minimizing both the number of UAVs and their total flying distance, Phalapanyakoon, K. and Siripongwutikorn, P. [
10] studied the route planning problem with the mission time constraint and the battery capacity constraint, where each UAV may have more than one trip in a complete route due to the limited battery capacity. Considering the multi-regional route planning problem and multiple UAVs with the energy constraint, Xie, J. and Chen, J. [
11] leveraged (1) a branch-and-bound based method to engender (near) optimal routes and (2) a genetic algorithm to efficiently engender routes for large-scale problems. Zhang, H. et al. [
12] introduced a hybrid differential evolution algorithm to plan high-quality routes for fixed-wing UAVs in complex three-dimensional environments. Most of the existing studies only consider minimizing the flying distance of UAVs in the route planning process and ignore other goals such as the balanced utilization rate of the UAVs, where the corresponding droneport site selection problem is also unsolved.
Droneport site selection is an optimization decision-making problem that involves multiple factors. Current studies on the droneport site selection problem mainly focus on civil droneports. Alves, CJP. et al. [
13] systematized the droneport siting criteria and proposed a site selection decision framework with a more objective decision-making process. Erkan, TE. et al. [
14] utilized the geographic information system to select proper droneport sites, which considered 23 indicators and applied the analytic hierarchy process method and the rank-order centroid method to select the best droneport locations. Aydin, N. and Seker, S. [
15] designed a guiding framework to select a hub droneport location within Turkey to satisfy demand and attract tourists. According to the bird ecological conservation data, Zhao, B. et al. [
16] evaluated the impact of the different droneport site selection schemes on bird ecology and then selected a more suitable site for the sustainable development of humans and nature. An emergency droneport site selection was proposed in [
17] based on the GeoSOT-3D global subdivision grid model, which verifies that the discrete global grid system has good suitability when performed as a spatial data structure for site selection. Liao, Y. and Bao, F. [
18] introduced the fuzzy decision-making thesis to evaluate different droneport site selection methods, which defined the dominance degree among the methods for evaluation by using the difference between the two triangular fuzzy numbers. Due to the fact that civil droneports are seriously limited by the surrounding environment, most of the current studies mainly consider the noise, the ecology, the electromagnetic radiation and other factors when making decisions on selecting the droneport sites. Compared to the civil droneport, the UAV distributed droneports have looser requirements on the surrounding environment pertaining to the site selection, and more focus on the utilization rate of the droneport cluster and the inspection cycle. To our knowledge, there are few studies on the site selection of the UAV distributed droneport.
As the supply stations and maintenance stations for UAVs, the distributed droneports play a key role in the unmanned inspection system and the route planning of UAVs. In general, the solution quality of droneport site selection and route planning directly affect the performance of the unmanned inspection system. Thus, the good cooperation between them is the key challenge and opportunity to realize the full performance potential of unmanned inspection systems. To this end, this paper proposes a route planning method for the UAV swarm that fuses the site selection of the distributed droneports. Firstly, we construct the inspection map and remove the redundant information outside the target regions of the inspection. Secondly, we build a multi-objective optimization function, and formulate both the route planning problem and the droneport site selection problem in a unified problem. Thirdly, based on the particle swarm optimization method, we redesign the encoding strategy, the updating rules and the decoding strategy of the particle to effectively solve the problem of the route planning of the UAV swarm inspection and the droneport site selection.
2. Proposed Research Structure
The research structure of this paper is illustrated in
Figure 1. The overall structure of the proposed method for solving both the route planning problem for the UAV swarm and the distributed droneport site selection problem includes four parts: the inspection map construction, the route inspection model formulation, the redesign of the particle encoding and decoding method and the experiments with comparison and analysis.
Pertaining to the inspection map construction, we first collect the map of the inspection target regions, and then separate the layers of the target regions to remove the redundant information outside those regions. Then, we set multiple candidate points (i.e., the candidate sites) for the droneport, simulate the inspection roads and divide the inspection roads and nodes into different parts, and simplify the complex roads in the target regions to provide the candidate droneport sites. Pertaining to the mathematical model for both the route planning problem and the droneport site selection problem, we build the objective function and constraints based on the practical demands of the road inspection to obtain an inspection optimization model under a specific number of UAVs. Pertaining to the algorithm, we improve the encoding method and the updating rules of the particles and design a corresponding particle decoding strategy. By doing so, the proposed algorithm enables each single particle to perform both the route planning and the droneport site selection, which further contributes to find the optimized route planning scheme and the droneport site selection scheme under a specific number of UAVs. Pertaining to the experiments and the analysis, we propose four indicators to evaluate the effectiveness of the route planning method and the droneport site selection method, and further compare our method with the ant colony optimization algorithm (ACO).
6. Experiments
We conduct all experiments on a Windows7 operating system, where the proposed algorithm is programmed using Python and all experiments are conducted in a Intel (R) Core (TM) I5-7300HQ CPU @ 2.50 GHz, 8 GB RAM.
6.1. Evaluation Indicators
6.1.1. Balanced Utilization Rate
The balanced utilization rate
is defined as the ratio of the shortest mileage to the longest mileage during a global inspection cycle as follows:
Specifically, closing to 1 indicates small differences among the UAV mileages and a highly balanced utilization rate. Similarly, much smaller than 1 indicates that the shortest mileage is much smaller than the longest one, which shows a less balanced utilization rate. Specifically, this indicator serves as a fundamental indicator of equilibrium within the domain of UAV deployment. Proximity to a value of 1 within this metric is indicative of a heightened state of equilibrium in UAV utilization, denoting a state where resources are distributed in a highly balanced manner across the fleet of drones.
6.1.2. Optimization Rate of the Total Mileage of the UAV Routes
We define the optimization rate of the total mileage of the UAV routes
as the ratio of the mileage of the inspection roads to the total mileage of the UAV routes during a global inspection cycle, which is formulated as follows:
where
is the total mileage that needs the inspection. Taking the Panyu district as an example, the length of the roads that needs to be inspected is 888.19 km. Note that
closing to 1 indicates that the total mileage of the UAV routes is relatively short in a global inspection cycle and the optimization rate of the total mileage of the UAV routes is relatively high. The essence of this indicator resides in elucidating the variability observed in the cumulative mileage of planned routes following each algorithmic iteration. A reduced degree of variability attests to the method delineated within this paper, manifesting a heightened level of congruity in the determination of total route mileage, thereby mitigating the likelihood of extreme scenarios.
6.1.3. Relative Droneport Utilization Rate
Given a droneport site selection scheme where the number of droneports is
times the number of the UAVs, the relative droneport utilization rate
is defined as the ratio of the total number of droneports on the routes to the total number of practical droneports, which is formulated as follows:
where
N is the total number of practical droneports and
. Note that some droneports might be used multiple times in the practical site selection process, which means there are some droneports that may share the same location among the routes. Moreover, the total number of droneports on the routes could infer the total number of droneports that are necessary without the overlapping locations. Thus, based on the ratio
, we could retrieve the utilization rate of the droneports given a specific number of UAVs where the site selection is not coincident. Specifically, the higher value of
indicates the higher utilization rate of droneports. The intrinsic importance of this indicator is rooted in its capacity to explicate the extent of repeated utilization within the framework of distributed droneports. As the frequency of this reuse surges, it unveils a concurrent decrease in construction expenditures, thereby accentuating the economic reverberations entwined with this phenomenon.
6.1.4. Optimization Rate of a Global Inspection Cycle
The optimization rate of a global inspection cycle
is defined as
In this paper, we constrain a global inspection cycle within one month and limit the UAVs to inspecting during the working day. In such a setting, we use 22 to represent the number of working days within a month. Specifically, if the is closer to 1, a global inspection cycle is shorter, and the optimization rate of a global inspection cycle is higher. The inherent significance of this indicator lies in its capacity to elucidate the interplay between the quantity of unmanned aerial vehicles (UAVs) and the formulation of route planning strategies. The envisioned outcome posits that as this indicator tends toward stabilization, it serves as an indicator of the method outlined in this paper, achieving a state of consistent and reproducible route planning outcomes.
6.2. Experimental Design
We take the Panyu district as an example to verify the feasibility and the effectiveness of the proposed method. We set the initial population to 100, and the values of learning factors and to 0.4 and 0.2, respectively.
The experiments are designed to check the feasibility (i.e., Experiment 1) and the effectiveness (i.e., Experiments 2–6) of the proposed method.
Experiment 1: We verify the feasibility of a method by checking whether it can generate reasonable routes and droneport sites. If yes, the method is deemed to be feasible.
Experiment 2: We show the correlation between the balanced utilization rate of UAVs and the total number of UAVs. The primary objective of this experiment is to meticulously examine the influence of the aggregate quantity of UAVs on the equilibrium of UAV utilization. In particular, a notable decline in the equilibrium of utilization would signify potential challenges in sustaining a uniform performance, casting a discerning light on the viability of our proposed method.
Experiment 3: We show the correlation between the optimization rate of the total mileage of the UAV routes and the total number of the UAVs. The primary objective of this experiment is to investigate whether there is a significant correlation between the total number of UAVs and the total planned route mileage. The rationale behind this investigation is that maintaining a stable total route mileage can facilitate a more consistent maintenance cycle for the UAV fleet, leading to a reduction in management costs.
Experiment 4: We show the correlation of the utilization rate of the droneports and both the number of the UAVs and the number of the droneports. The primary objective of this experiment is to investigate the impact of changes in the number of UAVs and droneports on the reusable rate of distributed droneports. The reusable rate is a crucial metric that can significantly reduce construction costs, and its increase can have a profound impact on the scalability and sustainability of distributed droneport systems.
Experiment 5: We explore the optimization rate of a global inspection cycle and show the optimization rates for different inspection solutions. The primary objective of this experiment is to investigate the impact of changes in the number of UAVs on the reusable rate of distributed droneports. The reusable rate is a crucial metric that can significantly reduce construction costs, and its increase can have a profound impact on the scalability and sustainability of distributed droneport systems.
Experiment 6: We compare our strategy with the ACO algorithm to show the effectiveness of the proposed method. The primary objective of this experiment is to rigorously evaluate the efficacy of the proposed enhancements to the particle swarm optimization algorithm in this paper by conducting a comprehensive comparative analysis with other state-of-the-art optimization algorithms. The results of this analysis will provide valuable insights into the strengths and weaknesses of the proposed approach and its potential for practical applications.
6.3. Experimental Results and Analysis
6.3.1. Experiment 1
We first retrieve the calculations carried out in the case of 5, 10 and 20 UAVs and the corresponding best particles. Then, we illustrate the route planning results and the droneport site selection results of the UAVs, which are obtained by decoding the best particles with different numbers of UAVs in
Figure 6 and
Figure 7, respectively.
In
Figure 6, the green lines refer to the routes that need the inspection, the red lines refer to the overlapped inspection routes in a global inspection cycle and the blue lines refer to the UAV routes planned by the proposed algorithm. We can find that the planned UAV routes could cover all the routes in the region, which demonstrates that the proposed method can plan reasonable routes for the global inspection. Thus, we conclude that the proposed method is feasible for route planning. In
Figure 7, the green lines refer to the routes that need the inspection, where the lines cross in blue refer to the droneports that are used only once during a global inspection cycle and the red line crossings refer to the droneports that are used multiple times during a global inspection cycle. We can observe that the selected droneport sites are all near the planned routes, and some droneports are repeatedly used for different numbers of UAVs, which reduces the number of droneports and further saves the construction costs. The observations demonstrate that the proposed method could select reasonable droneport sites and show the feasibility of our method in droneport site selection.
6.3.2. Experiment 2
We first explore the optimization problem in the case of 1 to 20 UAVs and obtain the corresponding best particles, and then plan the routes by decoding those particles. Furthermore, we illustrate the balanced utilization rate of UAVs under different numbers of the UAVs based on the planned routes in
Figure 8.
From
Figure 8, we could observe that the UAV utilization rate gradually decreases as the number of UAVs increases. Specifically, the utilization rate finally fluctuates around
and tends to be stable with a minimum of 73.74%. We can conclude that based on the proposed method, the route planning schemes under different numbers of UAVs achieve satisfactory balanced utilization rates. This result demonstrates that as the number of the drones increases, the balanced utilization rate of UAVs exhibits no significant decrease. This resilience in maintaining a robust level of optimization underscores the effectiveness and durability of our proposed method.
6.3.3. Experiment 3
Based on 20 types of the planned routes obtained from Experiment 2, we compute and collect the optimization rate of the total mileage of the routes under different number of UAVs. Then we illustrate the optimization rate under different numbers of UAVs in
Figure 9.
Figure 9 shows that the optimization rate of the total mileage of the UAV routes slightly decreases as the number of UAVs increases, where the optimization rate fluctuates around 76%. We could conclude that under different numbers of UAVs, the proposed method plans the routes with a consistent and promising optimization effect on the total mileage of the routes. This result elucidates that, when employing our route planning method, the cumulative mileage of the routes maintains a remarkable level of stability even as the number of drones escalates. This steadfastness in route mileage holds the potential to foster a consistent maintenance cycle for the UAV fleet, consequently yielding a reduction in overall management costs.
6.3.4. Experiment 4
Based on the 20 best particles found in Experiment 2, we decode these particles and obtain the selected droneport sites under different numbers of UAVs. We further compute the relative droneport utilization rate under different numbers of UAVs based on the selected droneport sites. Then, we illustrate the relative droneport utilization rate under different numbers of UAVs in
Figure 10.
From
Figure 10, we observe that the relative droneport utilization rate gradually increases as the number of UAVs increases and the former is always larger than 1. This indicates that the proposed method could select the droneport sites with a high relative droneport utilization rate and high profits. Moreover, there are always some repeatedly used droneports that are used multiple times during a global inspection cycle regardless of the number of UAVs, and the number of the repeatedly used droneports increases as the number of UAVs increases. Specifically, the highest relative droneport utilization rate is 1.52, which means that more than half of the droneports are repeatedly used. Thus, we can conclude that the proposed method could select the droneport sites with a relatively high droneport utilization rate. This result delineates a discernible correlation wherein the proliferation of UAVs precipitates a concomitant expansion in droneport facilities, notably marked by an augmented utilization of distributed airports. This notable trend underscores the potential for substantial reductions in construction expenses, thus exemplifying a compelling cost-saving facet within the realm of UAV infrastructure development.
6.3.5. Experiment 5
Based on 20 route planning schemes obtained from Experiment 2, we compute the optimization rate of a global inspection cycle under different numbers of UAVs. Then, we illustrate the optimization rate of a global inspection cycle under different numbers of UAVs in
Figure 11.
From
Figure 11, we can observe that the optimization rate of a global inspection cycle increases with a maximum of 93.18% as the number of UAVs increases. Specifically, when the number of the UAVs exceeds 4, the optimization rate is already above 75%, which shows a relatively high optimization rate. With more UAVs, the increasing speed of the optimization rate significantly slows down. When the number of the UAVs exceeds 9, the optimization rate achieves 88.64%, where the optimization rate after that is only slightly increased. Thus, we can conclude that the proposed method could plan the routes with a satisfactory optimization rate for a global inspection cycle. This result illustrates that as the number of drones increases, the solution space for flight path planning expands, potentially leading to better flight path planning outcomes. However, beyond 9 drones, further increasing the number of drones does not significantly enhance the effectiveness of flight path planning. This observation underscores the nuanced dynamics that govern the interplay between UAV quantity and route optimization, revealing a point of diminishing returns in the pursuit of route planning refinement.
6.3.6. Experiment 6
For ACO, the initial population is set to 100 and the pheromone weight is set to 1. The visibility weight is set to 2 and and the pheromone volatilization rate is set to 0.5. We apply ACO for solving the optimization problem with 13 UAVs, and obtain the route planning scheme, droneport site selection scheme and the optimization time. Based on the planned routes and selected droneport sites, the balanced utilization rate of UAVs, the optimization rate of the total mileage of the routes, the relative droneport utilization rate and the optimization rate of a global inspection cycle are computed. We summarize the results of ACO and our method in
Table 2.
From
Table 2, we can observe that pertaining to the optimization time, our method is faster than ACO, where the computation time of our method is about
shorter than the ACO. Pertaining to the balanced utilization rate of UAVs, the rate of our method is more than 90%, while the rate of ACO is less than 70%, which is much lower than ours. Pertaining to the optimization rate of the total mileage of the routes, our method is superior to ACO, where the optimization rate of our method is more than 12% higher than that of ACO. Pertaining to the relative droneport utilization rate and the optimization rate of a global inspection cycle, our method also slightly outperforms ACO, which indicates that our method achieves a higher optimization rate of the road inspection and the droneport site selection than ACO. Based on the observations, we could conclude that our method outperforms ACO on all perspectives. Additionally, the time consumed by offline route planning before take-off of a swarm of delivery drones in the literature [
22] ranges from 20 to 2173 s. By comparison, the time consumed by the method proposed in this paper is acceptable.
6.4. Discussion
Using a UAV swarm to perform inspection tasks is one of the future research trends. Following the delineation of targeted inspection areas, the foremost task at hand is to determine the locations and quantities of distributed droneports. Subsequently, leveraging this foundational information and considering the prevailing road network infrastructure, we embark on the meticulous planning of daily inspection routes for our fleet of UAVs. While extant literature has presented a plethora of studies concerning real-time route planning for UAV fleets [
4,
5,
6,
7] and offline planning methodologies [
8,
9,
10,
11], these endeavors, regrettably, have failed to incorporate the judicious selection of distributed droneport sites as an optimization objective. Moreover, they have largely neglected the critical issue of achieving equilibrium in UAV utilization. This prevailing oversight not only obfuscates the determination of optimal fleet sizes and the spatial distribution of distributed droneports but also exacerbates the problem of imbalanced drone utilization. Consequently, this paper posits the imperative of investigating the optimization goals encompassing the selection of distributed droneport locations alongside the equitable utilization of UAV resources. The experimental results in this paper show the following: (1) From
Figure 8 and
Figure 9, we can find that when the number of UAVs exceeds 14, the balanced utilization rate of UAVs is stable around 75%, which means the increase in the number of UAVs does not make much difference in the utilization rate. The corresponding optimization rate of the total mileage of the routes maintains at a stable state of around 76%, which means the increase in the number of UAVs does not make much difference in the total mileage of the inspection routes. (2) From
Figure 10, we can find that the increase in the number of UAVs will lead to the increase in the number of droneports; however, there is no linear positive correlation between them. Moreover, when the number of UAVs is less than 14,
is no less than 3.6, and when the number of UAVs exceeds 18,
is in the range of 2.6 to 3. As the number of UAVs increases, although the multiplier
decreases, the absolute total number of UAVs and distributed droneport increases, which increases the total cost. (3) From
Figure 11, the optimization rate of a global inspection cycle increases as the number of UAVs increases. The optimization rate already achieves 75% when the number of the UAVs is only 4, which means the shortest time to complete a global inspection cycle is 5.5 days with 4 UAVs continuously running. (4) We synthesize the multiple indicators of the Panyu district with 888.19 km of inspection roads under different numbers of UAVs in
Table 3, where we find that using 4 UAVs has the best cost performance.
Based on above observations, our method can effectively plan the inspection routes, balance the utilization of the UAVs and select the sites for the distributed droneports, which shows great significance for the fully autonomous UAV swarm inspection system for road inspection. Note that the proposed method in this paper remains subject to certain constraints. Specifically, one notable assumption is predicated on the presupposition that each UAV commences its flight with a fully charged battery. As stipulated within the confines of this study, each flight is constrained to a distance ranging from 27 to 30 km. However, it is essential to highlight that if a UAV takes off with a battery charge below full capacity, it may not attain the prescribed flight range of 27 to 30 km. In such instances, the flight route planning must be recalibrated in accordance with the estimated flight range based on the pre-flight battery status. Furthermore, while this paper extensively addresses the intricacies of drone route planning, it regrettably omits the consideration of UAV scheduling. This critical aspect shall be the focal point of our forthcoming research endeavors.