Urban fire monitoring-oriented unmanned aerial vehicle-WSN space-time collaborative optimization method

Abstract

The invention discloses an unmanned aerial vehicle-WSN space-time collaborative optimization method for urban fire monitoring, and the method comprises the steps: carrying out the spatial self-correlation analysis based on fire data, dividing an extremely high risk region, a high risk region, a medium risk region and a low risk region, assisting in the judgment of the access priority of an unmanned aerial vehicle, and carrying out the simulation in a time dimension through employing XG-Boost, and evaluating the reasonability; secondly, clustering fire points by using a mean shift method to obtain sensor node layout coordinates, clustering a cluster head, carrying out load balancing adjustment, and generating an initial inspection path of the unmanned aerial vehicle; thirdly, path deviation caused by wind interference is modeled, an overall-to-local precision control method is introduced, and unmanned aerial vehicle flight precision control is performed according to real-time feedback to reduce the path deviation; and finally, solving a global optimal path by using a system energy consumption scheduling algorithm. Therefore, manpower and material resources are saved, effective decision support is provided for dynamic deployment of fire rescue resources, and path deviation caused by wind interference is reduced as much as possible.

Description

Technical Field

[0001] This invention relates to the field of wireless sensor network technology, and in particular to a spatiotemporal collaborative optimization method for UAV-WSN for urban fire monitoring. Background Technology

[0002] With the advancement of urbanization and the rapid development of industrial production, the threat posed by sudden fires to public safety and the environment is increasingly severe. Existing fire early warning systems generally suffer from problems such as data synchronization delays, lagging emergency response, and insufficient routine monitoring capabilities. Wireless Sensor Networks (WSNs), due to their advantages of flexible deployment, real-time monitoring, and low cost, are gradually becoming an important technical means for fire monitoring. However, traditional WSNs still face challenges such as limited sensor node energy, poor communication stability in harsh environments, and the "energy void" effect caused by multi-hop data transmission, severely restricting network lifespan and reliability. Unmanned Aerial Vehicle (UAV)-assisted data collection and energy replenishment technology, with its high mobility, rapid response capabilities, and flexible path planning characteristics, provides a new approach to solving these problems.

[0003] In recent years, optimization research on UAV-assisted WSNs has mainly focused on trajectory planning and energy management. Existing technologies have proposed a UAV trajectory optimization algorithm that supports wireless energy transfer and inter-cluster load balancing. This algorithm uses mean shift to cluster sensor nodes and decomposes data scheduling and trajectory optimization sub-problems using mixed-integer programming. Another existing technology uses the DDPG (Deep Deterministic Policy Gradient) algorithm to optimize multi-UAV trajectories, incorporating parameters such as sensor node data storage and inter-UAV distances into the Markov decision process state space, effectively reducing system energy consumption. Finally, another existing technology proposes an energy-aware UAV smooth trajectory planning method that utilizes B-spline curves and a genetic algorithm to maximize the network's average energy.

[0004] For WSN energy management, existing technologies have designed a clustering method suitable for UAV-assisted wireless sensor networks, which solves the problems of uneven node distribution and resource constraints, extends network lifetime, and reduces packet loss. This method optimizes the energy efficiency and storage management of sensor nodes within a novel clustering framework based on K-means++ clustering and fuzzy logic-based cluster head selection. Experiments have shown that it significantly outperforms traditional algorithms such as LEACH, Voronoi partitioning, and mesh partitioning in terms of node lifetime and packet loss rate. Another existing technology designs a multi-UAV-vehicle joint trajectory planning algorithm based on load-balanced region partitioning, solving the problem of large-scale non-uniformly distributed IoT data collection. This existing technology uses convex optimization (CVX tool) to determine battery replacement points, ensuring UAV power constraints.

[0005] For WSN trajectory planning, existing technologies have designed a trajectory optimization algorithm based on multi-agent reinforcement learning, which solves the problem that single-UAV solutions cannot meet the real-time data acquisition requirements of large-scale WSNs. This existing technology proposes the MAPPO (Multi-Agent Proximal Policy Optimization) algorithm based on CTDE (Centralized-Training and Distributed-Execution), which avoids collisions between UAVs, thereby achieving safer and more efficient multi-UAV collaborative communication.

[0006] However, the above studies did not consider the impact of differences in the battery capacity of sensor nodes on the scheduling strategy in multi-charging pile collaborative scenarios, did not have constraints on path deviations caused by the instability of drone flight attitude, and the sensor node network was not large enough and lacked universality. Summary of the Invention

[0007] To address the above technical problems, this invention provides a spatiotemporal collaborative optimization method for UAV-WSN (Unmanned Aerial Vehicle) fire monitoring in urban fire monitoring, comprising the following steps: S1. Based on the Moran index, a fire risk classification model is constructed to constrain the urgency of fire risk. S2. Based on the mean drift function node clustering model, fire points are clustered to determine the deployment location of sensor nodes. Then, under the constraint of maximum communication distance, the initial cluster division is determined by reducing the UAV service path size through clustering. S3. Based on wind interference and fuzzy rule base, a UAV trajectory deviation model is used to suppress errors. A precision control method from the whole to the local is adopted. First, global trajectory planning is performed using TSP and GA algorithms, and 2-opt local search is used to optimize path smoothness. Then, local attitude control is performed using a UAV precision control algorithm based on real-time feedback to solve for real-time accuracy and construct control flow. S4. Real-time charging of sensor nodes is achieved through mobile charging vehicles. The energy consumption, load, and access path of the charging vehicles of sensor nodes are modeled to minimize energy consumption, data loss, and inspection time.

[0008] The beneficial effects of this invention are: (1) In this invention, by introducing Moran's index to conduct spatial autocorrelation analysis on fire points, fire risk classification is carried out in various regions, and potential high-risk areas are updated in real time, providing a reference for the dispatch of fire-fighting resources. At the same time, it guides the deployment of fire inspection and monitoring equipment, realizes rapid response in high-risk areas, and greatly saves manpower and material resources. (2) In this invention, the mean drift method is used to first cluster the fire points to obtain the sensor node layout coordinates, then cluster the cluster heads, and perform load balancing adjustment to generate the initial inspection path of the UAV; further, the path deviation caused by wind interference is modeled, and a precision control method from the whole to the local is introduced. The UAV flight precision control is performed based on real-time feedback to reduce the path deviation; finally, the system energy consumption scheduling algorithm is designed to solve the global optimal path; simulation results show that under the simulation conditions of 13 regions, 29 nodes, and a total area of ​​20 square kilometers, STCO-JCS controls the inspection path to 14-17km, optimizes the charging cycle to 10-20 minutes, and has a performance improvement of 3.17%-9.66% compared with traditional algorithms such as TSP and greedy algorithms under the collaborative scheduling of multiple charging vehicles, providing effective decision support for the dynamic deployment of fire rescue resources; (3) In this invention, mean drift clustering and load balancing cluster structure adjustment algorithm are designed to solve the sensor node layout coordinates and sensor cluster centers required for UAV inspection, so that the UAV can inspect along the cluster centers. Compared with visiting them one by one, it can cover multiple sensor nodes at the same time as much as possible, which greatly improves the information interaction efficiency between the sensor network and the UAV. Further cluster structure adjustment is carried out to make the load of each cluster nearly uniform, so as to achieve energy coordination of the sensor network. Considering the interference of urban terrain and wind interference on UAV cruise, a UAV precision control algorithm based on real-time feedback is introduced to accurately track the UAV along the inspection trajectory between the sensor node cluster centers and minimize the path deviation caused by wind interference. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a schematic diagram of a scenario in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the simulation verification of the number of fires within one year using the XGBoost method in an embodiment of the present invention, specifically the hot spot chain and cold spot chain (spatial clustering structure of fire risk zone). Figure 4 This refers to the fire spatial autocorrelation analysis algorithm in this embodiment of the invention. Figure 5 This is a schematic diagram of the clustering of region boundaries and sensor nodes in an embodiment of the present invention; Figure 6 This refers to the mean-shift clustering and load-balanced cluster structure adjustment algorithm in this embodiment of the invention. Figure 7 This is a schematic diagram illustrating the optimal hovering altitude of the UAV in an embodiment of the present invention; Figure 8 This is a schematic diagram of the drone's cruise access path in an embodiment of the present invention; Figure 9 The following is a schematic diagram of the changes in UAV flight status indicators in an embodiment of the present invention, wherein (a) is a schematic diagram of attitude disturbance changes, (b) is a schematic diagram of path deviation changes, and (c) is a schematic diagram of hit accuracy changes; Figure 10 This is a comparison chart of the total path length and cruising time of the UAV in an embodiment of the present invention; Figure 11 This is a real-time feedback-based UAV precision control algorithm in an embodiment of the present invention; Figure 12 The following are performance comparison charts of different algorithms in the embodiments of the present invention: (a) is a comparison chart of path lengths obtained by different algorithms, (b) is a comparison chart of system charging capacity obtained by different algorithms, (c) is a comparison chart of inspection time obtained by different algorithms, and (d) is a comparison chart of maximum node energy consumption obtained by different algorithms. Figure 13 The following are charging path diagrams for charging vehicles of different sizes in embodiments of the present invention, wherein (a) is a schematic diagram of a single charging vehicle path, (b) is a schematic diagram of a two-charging-vehicle path, (c) is a schematic diagram of a three-charging-vehicle path, (d) is a schematic diagram of a four-charging-vehicle path, (e) is a schematic diagram of a five-charging-vehicle path, and (f) is a schematic diagram of a six-charging-vehicle path. Detailed Implementation

[0010] This embodiment provides a spatiotemporal collaborative optimization method for UAV-WSN (Unmanned Aerial Vehicle) fire monitoring in urban fire monitoring, such as... Figure 1 As shown, it includes the following steps: S1. A fire risk grading model based on Moran's I index is constructed, and fire risk urgency constraints are established. To analyze the distribution characteristics of fire alarm events in different regions in geospatial space, this embodiment constructs a simplified spatial dataset. There are n regions in total, where region The corresponding fire alarm event density value is (Unit: times / square kilometer) , (Representing the set of positive real numbers); rearrange the original data into pairs. , for the region (Spatial Dataset) (Elements in the text) Assign two-dimensional geographic coordinates , and These represent the latitude and longitude coordinates of the region. After removing missing values, each region is mapped to its corresponding two-dimensional coordinate point according to the dictionary, ultimately forming a four-element dataset containing the region, coordinate location, and fire alarm event density. .

[0011] Define threshold distance If the area and Euclidean distance Distance less than the threshold Then it is believed and Adjacent, further define the weight matrix Weights in Furthermore, to describe the global autocorrelation in space, the Moran index is introduced, defined as: (1) in, This represents the average density of fire alarm incidents, and n represents the total number of regions.

[0012] To measure the reliability of the Moran index, the expected value and standard error of the global Moran index must be calculated, and significance must be verified using the Z-score. If the weight matrix is ​​symmetric and has no self-adjacency, then... That is, the region If no self-joins exist, the expected outcome is: (2) set up , Then the standard error estimate is: (3) If standardized scores If the value is 95%, then spatial autocorrelation is considered to exist.

[0013] To further measure the similarity between a single area and its neighbors, The local Moran index is defined as follows: (4) Where j represents the adjacent region Indexing; constructing a spatially weighted correlation matrix : (5) in, Indicates the region and The weighted correlation (unnormalized). Indicates the region The corresponding fire alarm event density value; if , indicating region and Non-adjacent; the result of this matrix is ​​the nearest neighbor correlation between the divided fire zones.

[0014] In terms of time, it is assumed that fires have obvious seasonal characteristics, such as high temperatures and dryness in summer and high electricity load in winter, which may lead to an increase in fire risk. Based on the spatial distribution of fires, this embodiment uses XGBoost to simulate the number of fires in densely populated and sparsely populated areas for 12 months to verify the reliability of the fire risk level classification.

[0015] S2. Based on the mean shift (MS) function node clustering model, fire points are clustered to determine the deployment location of sensor nodes. Then, under the constraint of maximum communication distance, the initial cluster division is determined by reducing the UAV service path size through clustering.

[0016] Let the cluster node be The mean shift is denoted as Its nuclear density is estimated as follows: (6) in, Let B represent the feature dimension, B represent the bandwidth parameter, K(∙) represent the radial kernel function, and N represent the total number of sensor nodes in the system.

[0017] Take the gradient of the kernel density function: (7) By decomposing the denominator and numerator, we obtain the mean shift function: (8) The first term in the above formula is the cluster node. The weighted average within the window, minus the magnitude of the mean shift itself.

[0018] The iterative formula is as follows: (9) Each iteration will change the cluster nodes Move to the local weighted mean and repeat until convergence to a stable point (density peak).

[0019] Mean-shift clustering requires calculating the maximum distance between clusters as a constraint, letting For signal transmission power, For transmit antenna gain, For receiving antenna gain, The propagation path loss is a function of distance. Additional losses, including connection losses, line losses, and multipath attenuation (in dB), are considered additional losses. As a lower bound for the received signal power, the link tolerance equation is expressed as: (10) The maximum positive root of the equation is obtained by using the bisection method. Let the hovering height of the UAV be H, and the maximum coverage radius on the ground, i.e., the maximum inter-cluster distance, be... .

[0020] S3. Based on wind interference and fuzzy rule base, the UAV trajectory deviation model is used to suppress errors. A precision control method from the whole to the local is adopted. First, global trajectory planning is performed using TSP and GA algorithms, and 2-opt local search is used to optimize path smoothness. Then, local attitude control is performed using a UAV precision control algorithm based on real-time feedback to solve for real-time accuracy and construct control flow.

[0021] Choose functions r(t) to represent the UAV's roll angle, p(t) to represent the pitch angle, and p(t) to represent the yaw angle over time. To measure stability over time, the attitude perturbation function is defined as follows: (11) Among them, attitude angle constraints require .

[0022] Since the higher the cruise speed, the less interference wind speed causes with path deviation; and the more severe the attitude disturbance, the greater the impact on path deviation, the path deviation as a function of time is constructed as follows: (12) in, represents the disturbance constant, taken as 0.05; p represents the path deviation constant, taken as 0.1; Let v(t) represent wind speed, and v(t) represent the cruise speed function, which is obtained from real-time speed data.

[0023] The total path deviation function is further obtained as follows: (13) The drone's patrol trajectory must prioritize responding to areas with high fire risk levels; therefore, a risk level scoring system must be established. All sensor nodes are divided into K clusters, and for each cluster i, an event density score is defined. and local Moran score as follows: (14) (15) Drones need to prioritize access to areas with high fire risk; therefore, the local spatial autocorrelation analysis index is specified as "high-high" (regional). The event density score and local Moran score of the device itself are higher than the mean, and the event density scores and local Moran scores of its surrounding neighbors are also higher than the mean. (Significance score) The significance score was 1, while all other significance scores were 0.

[0024] The risk level score of the i-th cluster, after weighting, is: (16) in, , , Indicates the weighting coefficient. The path deviation for accessing cluster head i is divided by the global maximum path deviation and weighted by the low-risk point access penalty to obtain the access cost for cluster i: (17) in, and Indicates the loss weight. The lower the access cost, the higher the risk level, and the higher the priority for access. The access costs between each pair of adjacent cluster heads are calculated and sorted in ascending order, with a threshold set at [value missing]. That is, the difference between the mean and standard deviation of the access cost. If so, then priority access will be enforced.

[0025] After establishing the error generation model obtained from the total path deviation function, error suppression is required. Based on the attitude perturbation function S(t), a global-to-local precision control method is adopted: first, global trajectory planning is performed using the TSP and GA algorithms, and 2-opt local search is used to optimize path smoothness; then, local attitude control is performed using a UAV precision control algorithm based on real-time feedback to solve for real-time accuracy and construct the control flow.

[0026] To achieve more accurate control precision, the path deviation function D(t) is transformed into a real-time control law function: (18) in, , and Indicates the gain coefficient. Indicates the deviation reference value, by limiting Amplitude assurance .

[0027] For the path deviation D(t) caused by wind interference (all path deviations in this embodiment refer to those caused by wind interference), fuzzy compensation rule bases for low speed, medium speed and high speed are designed as shown in Tables 1 to 3 respectively.

[0028] Table 1. Low-speed fuzzy rule base

[0029] Table 2 Medium-speed fuzzy rule base

[0030] Table 3 High-speed fuzzy rule base

[0031] Based on the coupling relationship between wind speed and path deviation, the aforementioned rule base constructs a hierarchical and progressive fuzzy compensation strategy, providing a systematic adaptive adjustment mechanism for path tracking control in complex wind field environments.

[0032] Speed ​​correction amount After defuzzing, we get : (19) in, The membership function is defined as follows: (20) in, The standard deviation of the speed correction is represented by the real-time speed correction generated from the flight simulation.

[0033] In practical scenarios, it is necessary to minimize equation (13), and express the objective function as: (twenty one) (21a) (21b) (21c) Equation (21a) represents the upper bound of the attitude disturbance function, and equation (21b) is used to limit the real-time control function. The upper bound is given by equation (21c), which represents the access cost of the i-th cluster. The upper boundary.

[0034] S4. The sensor nodes are charged in real time by a mobile charging vehicle. The energy consumption, load and access path of the sensor nodes are modeled based on the classic wireless communication energy consumption model and load balancing method to minimize energy consumption, data loss and inspection time.

[0035] A wireless communication energy consumption model is adopted to decompose energy consumption into transmission energy consumption and reception energy consumption, in order to quantitatively characterize the energy consumption characteristics of wireless devices during communication. This model models the energy consumption of sensor transmission and reception, assuming the transmission energy consumption of the j-th node is... Then we have: (twenty two) in, Indicates the number of bits sent. This represents the energy consumption per bit of electron. The energy consumption coefficient of the free space model is represented by [value]. This represents the energy consumption coefficient of the multipath attenuation model. In the free-space model, the signal propagates in an ideal environment with no obstruction, reflection, or scattering, and the receiver only receives a single line-of-sight (LOS) signal. In the multipath attenuation model, the signal forms multiple propagation paths through reflection, diffraction, and scattering. The superposition of different paths leads to phenomena such as fading, cancellation, and delay spread. Typical application scenarios include urban environments, indoor communication, and WSNs with high-density node deployment. When the communication distance is less than a critical value... When the free space model is used, the multipath attenuation model is used; otherwise, the multipath attenuation model is used.

[0036] Let the i-th cluster have The total energy consumption of cluster head i for a given set of sensor nodes is expressed as: (twenty three) in, This represents the number of control information bits broadcast by the UAV, and the energy consumption of the i-th cluster head receiver is... , This indicates the communication distance between cluster head i and the UAV. This represents the transmission power consumption of cluster head i.

[0037] The energy consumption of the j-th node within the i-th cluster is: (twenty four) Wherein, the energy consumption of the j-th sensor node within the cluster is , This represents the communication distance between cluster head i and the j-th sensor node within the cluster.

[0038] Define the load of the i-th cluster as: (25) Inter-cluster balance is defined by variance. for: (26) in, The smaller the value, the more balanced the load tends to be. At that time, the load of each cluster is perfectly balanced.

[0039] The model constraints are divided into two parts: cluster cover and cluster head energy. The cluster cover constraint for the i-th cluster is: (27) (28) The energy constraint for the cluster head of the i-th cluster is: (29) Where o represents the energy consumption threshold factor, and in this embodiment o=0.85.

[0040] Sensor nodes rely on electricity to operate, and the entire system also requires a mobile charging vehicle to meet real-time charging needs. The goal is to minimize energy consumption. The planned path starts from the data center, passes through each sensor for charging, and returns, as shown in the following formula: (30) in, This represents the distance from sensor node j to node j+1 of the mobile charging vehicle. denoted by , where represents the energy consumption rate of sensor node j, and v represents the speed of the charging vehicle.

[0041] To ensure sufficient power for the sensors, the minimum charging amount for the j-th sensor node must satisfy the following condition: (31) Where p represents the charging rate (in J / s). This indicates the charging time of sensor j. This represents the redundancy amount (in J) that is not less than the threshold.

[0042] In practical scenarios, it is necessary to minimize equation (30), and express the objective function as: (32) The constraint condition states that the charging amount of the j-th sensor node must be greater than the maximum energy consumption of the nodes within the cluster and the cluster head.

[0043] Unified optimization of total energy consumption of sensor nodes Total data loss and total drone cruise time Three key performance indicators: (33) (34) (35) in, This represents the data storage capacity of cluster head i. This represents the amount of data received by sensor node j within the cluster.

[0044] The total cruise time of a drone is measured using the time cost of the drone visiting all cluster heads in one round: (36) Where d(i,i+1) represents the distance between cluster heads i and i+1. Let v(t) represent the mean value of the drone's flight between cluster heads i and i+1. This indicates the time during which the drone interacts with cluster head i.

[0045] In practical scenarios, it is necessary to minimize equations (33) to (35), and the objective function is expressed as: (37) in, See equation (31), , , See equations (33), (34), and (35) respectively; , , These represent adjustment factors, used to control the weight distribution among the three objectives.

[0046] To adapt to dynamic changes in node states within the network, such as sudden drops in cluster head energy or near-full storage, an urgency-driven dynamic path adjustment mechanism is needed. This mechanism requires that after each round of scheduling, an urgency index is calculated based on sensor node state feedback. The urgency index corresponding to the j-th sensor node is: (38) in, and These represent the factors that control the weighting of energy and storage, respectively. This represents the maximum energy consumption of all sensor nodes during this round of cruise. This represents the amount of data lost at the j-th node. This represents the maximum amount of node data lost during this round of patrol; the sorted index function is used to describe the sensor node. Sort in ascending order, there are When the system performs task scheduling, from Starting from the lowest node, according to Sensor nodes are accessed in an ascending order to reduce the risk of data loss and premature node failure.

[0047] In this embodiment, as Figure 2 As shown, within a 5km × 5km × 2km urban space, multiple fire simulations were conducted using fire dispatch data from the past two years, including event type and dispatch time, combined with population density and area of ​​each region. The simulated fire points were then categorized and filtered based on their latitude and longitude coordinates, eliminating or merging areas with extremely sparse fire distribution and very small land area (≤1 fire incident within 2 years & ≤2 hectares of land area). Following this process, fire risk level standards were established, sensor node deployment was implemented, drone flight trajectories and optimal flight control strategies were simulated, and data center site selection and multi-charging pile scheduling schemes were determined.

[0048] Approximately 1400 fire dispatch records from a certain region over two years were selected. The region set was defined as R={A,B,C,D,...,M,...,P}. After filtering and removing areas without fire incidents, and removing interference terms, the event density was obtained as follows: average 10.70 times / square kilometer, median 10.35, showing a right skewness, with data mainly distributed in the mid-to-high value range; standard deviation 2.03, indicating a certain degree of density difference between regions, with extremely high and extremely low values. Other data showed a high concentration at the mean.

[0049] Furthermore, the global Moran's I is 0.12, indicating a weak spatial clustering of fire incidents overall. The Monte Carlo p-value is 0.243. After randomly shuffling the area coordinates and resimulating, the clustering is not significant, indicating that the generation of fire points meets the randomness requirement. The local spatial correlation analysis is then performed, as shown in Table 4.

[0050] Table 4 Local Spatial Correlation Analysis

[0051] As shown in Table 4, the event densities in regions F, G, K, J, and N are all greater than the mean. Among them, F, K, and J exhibit a "high-high" spatial correlation, showing a clear clustering relationship. Region F has a p-value below 0.05, indicating high significance and is a core hotspot. In contrast, regions M, P, and A all exhibit a "low-low" spatial correlation, and their event densities are also lower than the mean. In particular, point M has a p-value below 0.05 and a low event density, indicating that point M is a significantly low-risk cold spot. The fire risk level classification is shown in Table 5.

[0052] Table 5 Fire Risk Level Classification

[0053] Based on Table 5, we can preliminarily infer the spatial clustering structure of fire points between regions as follows: 1. Hotspot Chain: The high-density zone centered on the prominent hotspot F corresponds to a high fire risk level in the adjacent areas on the map.

[0054] 2. Cold chain: The significant cold spot M and the peripheral cold spots P and A form the edge of the cold spot, which corresponds to the low fire risk area on the map.

[0055] 3. Marginal anomaly areas: {B, G, N}, these areas belong to the "low-high" spatial correlation type, with a local Moran index less than 0, and the adjacent areas are significant hot or cold spots, which are easily affected by surrounding changes; areas C, D, I, L belong to the "high-low" spatial correlation type, with a local Moran index greater than 0, and are potential clusters in the event of a sudden fire, requiring key prevention.

[0056] Compare the number of fires on hot and cold chains within a year over a time dimension, such as... Figure 3 As shown, in the hot spot chain area, the number of fires is mainly concentrated in May to July, followed by October to February of the following year; in the cold spot chain area, the number of fires is mainly concentrated in April to June, followed by November to December, which verifies the hypothesis proposed earlier that fires have obvious seasonal characteristics in the time dimension.

[0057] To determine the spatial correlation between fire events, this embodiment presents pseudocode for a fire event spatial autocorrelation analysis algorithm, as follows: Figure 4 The algorithm is shown.

[0058] like Figure 5 As shown, the boundaries of 13 regions are marked with boxes. The fire points are clustered using the mean-shift method. In regions with an event density greater than 9, one sensor node is placed for every 9-11 fire points, for a total of 29 sensor nodes. Further mean-shift clustering is performed on the sensor nodes under the maximum communication distance constraint, resulting in 14 cluster heads. The results after merging the regions, clusters, and sensor nodes are summarized as follows: Figure 5 As shown.

[0059] from Figure 5 As can be seen, in the hot spot chain region, the cluster center is located at the center surrounded by sensor nodes; in the cold spot chain region, due to the sparse distribution of sensors, the cluster center is the sensor node itself. The sensor deployment and clustering results conform to the characteristics of fire risk classification.

[0060] Fire severity levels can provide a reference for the number of sensor nodes to deploy. The next step is to determine the sensor node deployment coordinates and the sensor cluster centers required for UAV inspections. The first step is mean-drift clustering. To further achieve energy coordination of the sensor network, cluster structure adjustment is also necessary, aiming to make the load on each cluster approximately uniform. The pseudocode for the mean-drift clustering and load-balancing cluster structure adjustment algorithm is as follows: Figure 6 The algorithm is shown.

[0061] The initial conditions are set as shown in Table 6.

[0062] Table 6 Initial Conditions Table

[0063] Simulations were conducted for two different scenarios: one involving dense urban buildings and the other satisfying the confidence level of a (LOS / NLOS) mixed probability model. The numerical solutions obtained are shown in Table 7.

[0064] Table 7 Numerical Solution Table

[0065] Taking into account interference factors such as flight control and complex terrain limitations, the maximum communication distance was simulated. Maximum communication radius on the ground Setting the wind speed influence factor to 0.05, the recommended cruising altitude is 1411.89m, the optimal flight speed is 400km / h, the dive angle is 60°, the average dive time is 13.20s, the average climb time is 13.20s, the total communication time is 400.00s, and the total single-point interaction time is 426.41s.

[0066] The optimal hovering altitude and access path for drones when inspecting and collecting data along cluster centers are as follows: Figures 7 to 8 As shown. The optimal hovering height for the UAV to conduct data communication at each cluster head is as follows. Figure 7 As shown, the lowest hovering height is 141 meters and the highest hovering height is 345 meters, both within the maximum communication range. Based on this, the patrol sequence is as follows: Figure 8 As shown in Figures 1 to 14, hotspot chains are accessed first while ensuring the shortest path. The approach meets the urgency constraints. After passing through a medium-risk area G, it performs 5-12 steps to inspect the high-risk area as much as possible, and then quickly takes the shortest path to pass through medium- and low-risk areas such as B, A, and N. The trajectory planning satisfies the requirement of prioritizing access to extremely high-risk areas, and secondly, while ensuring priority inspection of high-risk areas, it tries to shorten the access path as much as possible, and quickly traverses the medium / low-risk areas.

[0067] Based on the above trajectory, the flight attitude of the UAV is simulated in real time: assuming sensitivity parameters The value is 0.1. Simulations are performed on attitude disturbances, path deviations, and accuracy along the inspection path. Changes in flight state indicators are shown below. Figure 9 As shown.

[0068] from Figure 9 As can be seen from (a), the average attitude perturbation is 0.024 rad, which is much smaller than 0.067 rad; from Figure 9 As can be seen from (b), the average path deviation is 4.55m, which is much smaller than the path deviation threshold of 8.19m, indicating that there is no significant trajectory deviation during the drone's cruise; from Figure 9 As can be seen from (c), the average accuracy is 0.957, which is very close to 1, indicating that the speed correction can effectively reduce the impact of wind speed and attitude disturbances.

[0069] To evaluate the performance of the method in this embodiment in terms of total path length and cruise time, the total path length and cruise time of the method in this embodiment are compared with those of the nearest neighbor method, the random search method, and the greedy insertion method, respectively. The comparison results are as follows: Figure 10 As shown.

[0070] Depend on Figure 10 It can be seen that the performance of the method in this embodiment is superior to the other three algorithms. Specifically, the total inspection path distance of the method in this embodiment is approximately 7.8 kilometers, and the time taken is 70.4 seconds. The corresponding values ​​for the other three algorithms are 8.13 kilometers and 73.2 seconds; 11.79 kilometers and 106.1 seconds; and 8.80 kilometers and 79.2 seconds, respectively. Compared with the nearest neighbor method, the random search method, and the greedy insertion method, the total path length of the method in this embodiment is reduced by 3.9%, 22.5%, and 11.2%, respectively, and the cruise time is reduced by 3.8%, 22.4%, and 11.1%, respectively.

[0071] The initial trajectory obtained after clustering does not yet take into account the interference of urban terrain and wind on the drone's cruise. To accurately determine the drone's inspection trajectory between the cluster centers of sensor nodes and minimize path deviation caused by wind interference, the accuracy function is derived from the path deviation function:

[0072] make For function The maximum value is continuously updated over time until it reaches its maximum value. The drone's flight altitude at time t is denoted as . The maximum communication height at each cluster head is .

[0073] Suppose that the UAV is at the i-th cluster head at time t, and generates i-1 cruise trajectories. Let the attitude perturbation function between the i-th cluster head and the (i+1)-th cluster head be denoted as... The path deviation function is denoted as Define respectively and The mean is:

[0074]

[0075] The pseudocode for a drone precision control algorithm based on real-time feedback is as follows: Figure 11 The algorithm is shown.

[0076] Based on the sensor node clustering results, and with the goal of responding to fires in each area as quickly as possible, the latitude and longitude coordinates of the data center were determined to be [120.697262, 36.375535]. Combined with the dynamic energy consumption adjustment algorithm proposed by Kissing, parameters were set... , , (Multipath model amplification factor). Table 8 shows a comparison of energy consumption and scheduling for a single-charge vehicle system.

[0077] Table 8 Comparison of Energy Consumption and Dispatch for Single-Charge Vehicle System

[0078] As shown in Table 8, in a single-charger scenario, the total energy consumption of the method in this embodiment is 7.8% lower than that of the TSP & heuristic algorithm in one charging cycle. With all-weather coverage, it can reduce the number of charging cycles by 23, and the average charging time per round is significantly reduced by 44.8%. Compared to the traditional heuristic algorithm, the method in this embodiment abandons the inefficient method of charging all sensor nodes once per round. It adopts the dynamic reconstruction mechanism proposed above to dynamically adjust the number of sensors charged in each cycle, accelerating the charging frequency of high-energy-consuming nodes and appropriately reducing the charging frequency of low-energy-consuming nodes, thus significantly improving the scheduling efficiency of the charging vehicle.

[0079] Next, the performance of the method in this embodiment will be compared and analyzed. The method in this embodiment will be compared with TSP, greedy strategy, nearest neighbor, and 2-opt, respectively. The path length, system charging amount, inspection time, and maximum node energy consumption obtained by different algorithms will be compared as follows: Figure 12 Figures (a), (b), (c), and (d) are shown in the diagram.

[0080] Depend on Figure 12As shown in (a), the method in this embodiment is not optimal in terms of path length when the number of mobile charging vehicles is small. However, as the number of mobile charging vehicles increases from 2 to 6, the total path length increases from 13.81 km to 17.25 km. Compared with TSP, greedy strategy, nearest neighbor and 2-opt, the growth rate of the method in this embodiment is the slowest, and the total path length is shortened by 4.67%, 2.89%, 2.89% and 0.55% respectively. This shows that as the scale of charging vehicles increases, the advantage of the short total inspection path required by the method in this embodiment becomes more and more obvious.

[0081] Depend on Figure 12 As shown in (b) above, within the same charging cycle, as the number of charging vehicles increases, the charging frequency of the sensor system accelerates, and the charging amount required by the sensor system continuously decreases. When the number of mobile charging vehicles increases from 2 to 6, the charging amount required by the system using the method of this embodiment decreases from 10161.82J to 3980.45J, which is 6.42%, 6.47%, 6.47%, and -1.47% lower than the other four algorithms, respectively. This is because the method of this embodiment uses a dynamic reconstruction mechanism to adjust the number of sensors charged in each cycle, abandoning the inefficient method of charging sensor nodes once per cycle in the traditional method, thereby accelerating the charging frequency of high-energy-consuming nodes and reducing the charging frequency of low-energy-consuming nodes. The advantages of the method of this embodiment are more significant, especially when the number of mobile charging vehicles is small.

[0082] Depend on Figure 12 As shown in (c), when the number of mobile charging vehicles increases from 2 to 6, the method in this embodiment performs well under various mobile charging vehicle scales, reducing the inspection time from 1286 seconds to 505.6 seconds. Compared with the other four algorithms, the total inspection time is reduced by 12.61%, 11.72%, 11.72%, and 4.23%, respectively. Furthermore, from Figure 12 Analysis (d) shows that when the number of mobile charging vehicles increases from 2 to 6, the power threshold of each sensor node in one charging cycle of the method in this embodiment is 604.14J, 567.16J, 608J, 343.36J and 298.5J respectively. Compared with the other four algorithms, these are reduced by 2.89%, 10.25%, 7.47% and 1.42% respectively. This is because the method in this embodiment adopts a dynamic reconstruction mechanism, which reduces the capacity threshold of the sensor node battery, thereby making the whole system more energy-efficient.

[0083] Considering four metrics—path length, required charging capacity of the sensor system, total inspection time, and sensor node battery capacity threshold—this embodiment achieves performance advantages of 6.61%, 9.66%, 9.02%, and 3.17% respectively compared to four other algorithms. The corresponding charging vehicle paths for different numbers of charging vehicles are as follows: Figure 13 The figures (a)-(f) are shown in the table.

[0084] from Figure 13 As the number of charging vehicles increases, the sensor node network around the data center is divided into more smaller areas, the path size of the charging vehicles continues to shrink, and the trajectory becomes more compact. This is more conducive to the flexible peak-shifting scheduling of charging vehicles. At the same time, the system can better serve the high-energy-consuming sensor nodes in high-risk fire areas, providing a strong guarantee for the system to have a faster response speed to fires.

[0085] This embodiment discloses a spatio-temporal collaborative optimization for joint control and scheduling (STCO-JCS) algorithm for UAVs and wireless sensor networks (WSN). It relates to the field of sensor network technology, can save human and material resources, provide effective decision support for the dynamic deployment of fire and rescue resources, and minimize path deviations caused by wind interference.

[0086] In addition to the embodiments described above, the present invention may have other implementations. All technical solutions formed by equivalent substitution or equivalent transformation fall within the protection scope claimed by the present invention.

Claims

1. A spatiotemporal collaborative optimization method for UAV-WSN for urban fire monitoring, characterized in that: Includes the following steps: S1. Based on the Moran index, a fire risk classification model is constructed to constrain the urgency of fire risk. S2. Based on the mean drift function node clustering model, fire points are clustered to determine the deployment location of sensor nodes. Then, under the constraint of maximum communication distance, the initial cluster division is determined by reducing the UAV service path size through clustering. S3. Based on wind interference and fuzzy rule base, a UAV trajectory deviation model is used to suppress errors. A precision control method from the whole to the local is adopted. First, global trajectory planning is performed using TSP and GA algorithms, and 2-opt local search is used to optimize path smoothness. Then, local attitude control is performed using a UAV precision control algorithm based on real-time feedback to solve for real-time accuracy and construct control flow. S4. Real-time charging of sensor nodes is achieved through mobile charging vehicles. The energy consumption, load, and access path of the charging vehicles of sensor nodes are modeled to minimize energy consumption, data loss, and inspection time.

2. The UAV-WSN spatiotemporal collaborative optimization method for urban fire monitoring according to claim 1, characterized in that: In step S1, a spatial dataset is constructed. There are n regions in total, where region The corresponding fire alarm event density value is , , Represent the set of positive real numbers; organize the original data into pairs. , for the region Assigning two-dimensional geographic coordinates , and These represent the latitude and longitude coordinates of the region, respectively. After removing missing values, each region is mapped to its corresponding two-dimensional coordinate point according to the dictionary, ultimately forming a four-element dataset containing region, coordinate location, and fire alarm event density. ; Define threshold distance If the area and Euclidean distance Distance less than the threshold Then it is believed and Adjacent, further define the weight matrix Weights in To characterize the global autocorrelation of space, the Moran index is introduced and defined as: (1) in, This represents the average density of fire alarm incidents, and n represents the total number of regions. Calculate the expected value and standard error of the global Moran's index, and verify its significance using Z-score. If the weight matrix is ​​symmetric and has no self-adjacency, then... That is, the region If no self-joins exist, the expected outcome is: (2) set up , Then the standard error estimate is: (3) If standardized scores If the value is 95%, then spatial autocorrelation is considered to exist. right The local Moran index is defined as follows: (4) Where j represents the adjacent region Indexing; constructing a spatially weighted correlation matrix : (5) in, Indicates the area and Weighted correlation, Indicates the area The corresponding fire alarm event density value; if , indicating region and Non-adjacent; the result of this matrix is ​​the nearest neighbor correlation between the divided fire zones.

3. The UAV-WSN spatiotemporal collaborative optimization method for urban fire monitoring according to claim 1, characterized in that: In step S2, let the cluster node be... The mean shift is denoted as Its nuclear density is estimated as follows: (6) in, Let B represent the feature dimension, B represent the bandwidth parameter, K(∙) represent the radial kernel function, and N represent the total number of sensor nodes in the system; take the gradient with respect to the kernel density function: (7) By decomposing the denominator and numerator, we obtain the mean shift function: (8) The first term in the above formula is the cluster node. The weighted average within the window, minus the mean itself, represents the magnitude of the mean shift; the iterative formula is as follows: (9) Each iteration will change the cluster nodes Move to the local weighted mean and repeat until convergence to a stable point, i.e., the density peak; Constraining mean-shift clustering by calculating the maximum distance between clusters, let For signal transmission power, For transmit antenna gain, For receiving antenna gain, The propagation path loss is a function of distance. Additional losses, including connection losses, line losses, and multipath attenuation, make As a lower bound for the received signal power, the link tolerance equation is expressed as: (10) The maximum positive root of the equation is obtained by using the bisection method. Let the hovering height of the UAV be H, and the maximum coverage radius on the ground, i.e., the maximum inter-cluster distance, be... .

4. The UAV-WSN spatiotemporal collaborative optimization method for urban fire monitoring according to claim 1, characterized in that: In step S3, the functions r(t) representing the UAV's roll angle as a function of time, p(t) representing the pitch angle as a function of time, and the functions p(t) representing the yaw angle as a function of time are selected. To measure stability over time, the attitude perturbation function is defined as follows: (11) Among them, attitude angle constraints require ; The path deviation over time is described by the following function: (12) in, Let p represent the disturbance constant and p represent the path deviation constant. Let v(t) represent wind speed, and v(t) represent the cruising speed function, which is obtained from real-time speed data. The total path deviation function is further obtained as follows: (13) A risk level scoring system is established, and drone patrol trajectories prioritize responses to areas with high fire risk levels; all sensor nodes are divided into K clusters, and an event density score is defined for each cluster i. and local Moran score as follows: (14) (15) The local spatial autocorrelation analysis index is defined as a high-high significance score. The significance score was 1, and all other significance scores were 0. "High-high" specifically refers to the region... The event density score and local Moran score of the i-th cluster are both higher than the mean, and the event density scores and local Moran scores of its surrounding neighbors are also higher than the mean. The weighted risk level score for the i-th cluster is: (16) in, , , Indicates the weighting coefficient. The path deviation for accessing cluster head i is divided by the global maximum path deviation and weighted by the low-risk point access penalty to obtain the access cost for cluster i: (17) in, and Indicates the loss weight. The lower the access cost, the higher the risk level, and the higher the priority for access. The access costs between each pair of adjacent cluster heads are calculated and sorted in ascending order, with a threshold set at [value missing]. That is, the difference between the mean and standard deviation of the access cost. If so, then priority access will be enforced; After establishing the error generation model obtained from the total path deviation function, the error is suppressed; based on the attitude disturbance function S(t), a precision control method from global to local is adopted; the path deviation function D(t) is transformed into a real-time control law function. (18) in, , and Indicates the gain coefficient. Indicates the deviation reference value, by limiting Amplitude assurance .

5. The UAV-WSN spatiotemporal collaborative optimization method for urban fire monitoring according to claim 1, characterized in that: In step S3, for the path deviation D(t) caused by wind interference, fuzzy compensation rule bases for low speed, medium speed, and high speed are designed respectively. When the wind speed is 1 to 5 m / s, the low speed fuzzy compensation rule base is used. The low speed fuzzy compensation rule base stipulates that: when the path deviation is between -5 and -2 m, a positive speed correction of 1.5 to 2 m / s is applied; when the path deviation is between -2 and -1 m, the correction amount is 0.5 to 0.9 m / s; when the path deviation is between -1 and 1 m, the correction amount is -0.2 to 0.2 m / s. When the path deviation is between 1 and 2 m, a negative correction of -0.9 to -0.5 m / s is applied; when the path deviation is between 2 and 5 m, the correction is -2 to -1.5 m / s. When the wind speed is 6–10 m / s, the medium-speed fuzzy compensation rule library is used. The medium-speed fuzzy compensation rule library stipulates that: when the path deviation is -5 to -2 m, the correction amount is 2–2.5 m / s; when the path deviation is -2 to -1 m, the correction amount is 0.9–1.3 m / s; when the path deviation is -1 to 1 m, the correction amount is -0.4–0.4 m / s; when the path deviation is 1 to 2 m, the correction amount is -1.3 to -0.9 m / s; and when the path deviation is 2 to 5 m, the correction amount is -2.5 to -2 m / s. When the wind speed is 11–15 m / s, a high-speed fuzzy compensation rule library is used. The high-speed fuzzy compensation rule library stipulates that: when the path deviation is -5 to -2 m, the correction amount is 2.5–3 m / s; when the path deviation is -2 to -1 m, the correction amount is 1.3–1.5 m / s; when the path deviation is -1 to 1 m, the correction amount is -0.5–0.5 m / s; when the path deviation is 1 to 2 m, the correction amount is -1.5 to -1.3 m / s; and when the path deviation is 2 to 5 m, the correction amount is -3 to -2.5 m / s. Speed ​​correction amount After defuzzing, we get : (19) in, The membership function is defined as follows: (20) in, The standard deviation represents the speed correction amount; Minimize equation (13), and express the objective function as: (21) (21a) (21b) (21c) Equation (21a) represents the upper bound of the attitude disturbance function, and equation (21b) is used to limit the real-time control function. The upper bound is given by equation (21c), which represents the access cost of the i-th cluster. The upper boundary.

6. The UAV-WSN spatiotemporal collaborative optimization method for urban fire monitoring according to claim 1, characterized in that: In step S4, a wireless communication energy consumption model is used to decompose energy consumption into transmission energy consumption and reception energy consumption, quantitatively characterizing the energy consumption features of wireless devices during communication. This model models the energy consumption of sensor transmission and reception, assuming the transmission energy consumption of the j-th node is... Then we have: (22) in, Indicates the number of bits sent. This represents the energy consumption per bit of electron. The energy consumption coefficient of the free space model is represented by [value]. This represents the energy consumption coefficient of the multipath attenuation model; when the communication distance is less than the critical value. When the condition is met, the free space model is used; otherwise, the multipath attenuation model is used. Let the i-th cluster have The total energy consumption of cluster head i for a given set of sensor nodes is expressed as: (23) in, This represents the number of control information bits broadcast by the UAV, and the energy consumption of the i-th cluster head receiver is... , This indicates the communication distance between cluster head i and the UAV. This represents the transmission power consumption of cluster head i; The energy consumption of the j-th node within the i-th cluster is: (24) Wherein, the energy consumption of the j-th sensor node within the cluster is , This represents the communication distance between cluster head i and the j-th sensor node within the cluster.

7. The UAV-WSN spatiotemporal collaborative optimization method for urban fire monitoring according to claim 1, characterized in that: In step S4, the load of the i-th cluster is defined as: (25) Inter-cluster balance is defined by variance. for: (26) in, The smaller the value, the more balanced the load tends to be. At that time, the load on each cluster is perfectly balanced; The model constraints include cluster cover constraints and cluster head energy constraints. The cluster cover constraint for the i-th cluster is: (27) (28) The energy constraint for the cluster head of the i-th cluster is: (29) Where o represents the energy consumption threshold factor.

8. The UAV-WSN spatiotemporal collaborative optimization method for urban fire monitoring according to claim 1, characterized in that: In step S4, the sensor nodes are charged in real time using a mobile charging vehicle, with the goal of minimizing energy consumption. The planned path starts from the data center, passes through each sensor in turn to charge it, and returns, as shown in the following formula: (30) in, This represents the distance from sensor node j to node j+1 of the mobile charging vehicle. represents the energy consumption rate of sensor node j, and v represents the speed of the charging vehicle. The minimum charging amount of the j-th sensor node satisfies the following condition: (31) Where p represents the charging rate. This indicates the charging time of sensor j. This indicates a redundancy amount that is not less than the threshold. Minimize equation (30), and express the objective function as: (32) The constraint condition states that the charging amount of the j-th sensor node must be greater than the maximum energy consumption of the nodes within the cluster and the cluster head.

9. The UAV-WSN spatiotemporal collaborative optimization method for urban fire monitoring according to claim 1, characterized in that: In step S4, the total energy consumption of the sensor nodes is optimized uniformly. Total data loss and total drone cruise time Three performance indicators: (33) (34) (35) in, This represents the data storage capacity of cluster head i. This represents the amount of data received by sensor node j within the cluster; The total cruise time of a drone is measured using the time cost of the drone visiting all cluster heads in one round: (36) Where d(i,i+1) represents the distance between cluster heads i and i+1. Let v(t) represent the mean value of the drone's flight between cluster heads i and i+1. This indicates the time during which the drone interacts with cluster head i. Minimize equations (33) to (35), and express the objective function as: (37) in, , , These represent adjustment factors, used to control the weight distribution among the three objectives.

10. The UAV-WSN spatiotemporal collaborative optimization method for urban fire monitoring according to claim 1, characterized in that: In step S4, an urgency-driven dynamic path adjustment mechanism is set up, requiring that after each round of scheduling, an urgency index is calculated based on the sensor node status feedback. The urgency index corresponding to the j-th sensor node is: (38) in, and These represent the factors that control the weighting of energy and storage, respectively. This represents the maximum energy consumption of all sensor nodes during this round of cruise. This represents the amount of data lost at the j-th node. This represents the maximum amount of node data lost during this round of patrol; the sorted index function is used to describe the sensor node. Sort in ascending order, there are When the system performs task scheduling, from Starting from the lowest node, according to The sensor nodes are accessed in ascending order.

Images