Low-altitude environment multi-unmanned aerial vehicle efficient patrol method based on terrain elevation map

Through the multi-UAV patrol method based on topographic elevation map, using rasterized processing and dynamic infographic updates, the efficiency and information accuracy problems in forest fire patrols under complex terrains are solved, and efficient fire detection and confirmation are achieved .

CN119045509APending Publication Date: 2024-11-29BEIHANG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411190328.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-28
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

The existing technology has problems such as low search efficiency, incomplete fire information, and not timely update of fire dynamic changes in forest fires in complex terrain, especially when the forest fire spread model information is not fully utilized.

Method used

Using the efficient patrol method of low-altitude environmental multi-UAV based on topographic elevation map, the fire presence probability map and environmental uncertainty map are dynamically updated, combined with Bayesian reasoning and particle filtering, Optimize the patrol routes of drones to improve search efficiency and accuracy of infographics.

Benefits of technology

It significantly improves the efficiency of forest fire patrols and the accuracy of infographics, meets the needs of terrain follow-up and inter-rail collision avoidance, improves the discovery and confirmation rate of fire situations, reduces the misjudgment rate, and meets the requirements of real-time and feasibility .

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119045509A_ABST
    Figure CN119045509A_ABST
Patent Text Reader

Abstract

The invention discloses a low-altitude environment multi-unmanned aerial vehicle efficient patrol method based on a terrain elevation map, and belongs to the technical field of unmanned aerial vehicle navigation guidance and control. A plurality of unmanned aerial vehicles execute a forest fire patrol task, airborne sensor parameters are determined, and a to-be-patrolled area serves as a task area; and then, initializing a patrol information graph by utilizing prior information, executing a patrol task by multiple unmanned aerial vehicles according to the initialized patrol information graph at the beginning, acquiring comprehensive information of a current patrol area, estimating posterior probability distribution of potential forest fire source items, and calculating the potential forest fire source items and credibility thereof by adopting a Gaussian mixture model. And finally, dynamically updating the patrol information graph based on the potential forest fire source item, obtaining a three-dimensional route point at the next moment based on multi-unmanned aerial vehicle horizontal route planning, updating the patrol information graph based on a detection result, and carrying out patrol route planning at the next moment. The patrol efficiency is improved, the accuracy of the patrol information graph is improved, and the requirements of terrain following and inter-aircraft collision avoidance are met.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a multi-UAV high-efficiency inspection method in a low-altitude environment based on a terrain elevation map, and belongs to the technical field of UAV navigation, guidance and control. Background Art

[0002] In recent years, with climate warming and more extreme weather, the world has entered a high incidence period of forest fires, and the risk of forest fires has increased. Forest fires not only threaten the safety of forest resources, but also seriously threaten the safety of life and property of people in forest areas and firefighters. The key to solving forest fires is "early detection and early rescue." The earlier the discovery and the more timely and comprehensive the disaster information obtained, the more scientific disaster relief measures can be formulated to extinguish the fire as soon as possible and minimize the losses caused by the fire.

[0003] Traditional manual monitoring and early warning methods have problems such as poor timeliness in detecting fires, incomplete fire information, unclear fire locations, and one-sided fire development trends, which seriously restrict the improvement of forest fire early warning and handling capabilities in complex terrain conditions.

[0004] Drones have the advantages of flexibility, precision and efficiency, and strong adaptability to harsh environments and regions. They can replace personnel to carry out all-day and all-weather patrols and early warnings over forest areas, obtain clues of fires in the first time, and achieve the goal of early detection. The drone forest fire patrol mission is essentially a special target (fire) search mission for a specific area (forest). According to the different drone route planning strategies, regional search tasks can be roughly divided into two types: coverage search and heuristic search. Among them, the characteristic of coverage search is the indiscriminate full coverage search of the search area. For the situation where the search area is completely unknown, the coverage search method is the optimal strategy, but when the initial probability distribution of the target is known or the search time is limited, the search efficiency of this method is significantly lower. When the target information is partially known, the heuristic search method can be used to describe the uncertainty of the environment or the probability distribution of the target by establishing a search probability map, making full use of prior information and real-time detection information, focusing on searching areas with a high probability of target existence, and improving search efficiency.

[0005] Although scholars have achieved rich results in the field of target search, the existing technology still generally has the following problems:

[0006] (1) Most of the search objects considered in existing methods are discrete and independent point-like static or moving targets. Few studies have discussed continuous and interrelated surface dynamic diffusion targets such as forest fires. Therefore, the likelihood information and detection information contained in the forest fire spread model cannot be fully utilized, which seriously restricts the further improvement of search efficiency.

[0007] (2) The target (i.e., fire) existence probability map is of great significance for guiding UAVs to efficiently and comprehensively detect fires. Due to the dynamic changes of fire conditions, the search information map should be dynamically updated in combination with the spread characteristics of forest fires, but few studies have considered this.

[0008] (3) The search routes planned by current methods are mostly two-dimensional, that is, it is assumed that the UAV flies at a fixed altitude. However, in complex mountainous terrain, the flight altitude of the UAV needs to be adjusted according to the terrain characteristics and undulations to ensure the flight safety of the UAV and a comprehensive search of the entire mission area. Summary of the invention

[0009] In order to solve the problems of dynamic updating of search information graphs and flight altitude tracking in complex terrain, the present invention proposes an efficient patrol method for multiple UAVs in low-altitude environments based on terrain elevation maps, aiming at forest fire targets in low-altitude mountainous environments. This method significantly improves the efficiency of forest fire patrols while ensuring the flight safety of UAVs.

[0010] The method for efficiently inspecting multiple UAVs in a low-altitude environment based on a terrain elevation map specifically comprises the following steps:

[0011] Step 1: Several drones perform forest fire inspection tasks, determine the parameters of the airborne sensors, and use the area to be inspected as the task area for rasterization;

[0012] ① Airborne sensor parameters include detection range, detection probability and false alarm probability The details are as follows:

[0013] N will participate in forest fire patrol missions u The drones are numbered 1, 2, ..., N. u The optical sensor carried by the drone always points vertically downward, and its detection range is constant as one grid directly below the drone, that is, S = Δx × Δy.

[0014] It indicates the probability that the sensor detects the target when the target (i.e., fire) exists. It represents the probability that the sensor detects the target when the target does not exist.

[0015] ② Rasterization of the task area, specifically:

[0016] Assume that the task area is a rectangle and divide it evenly into L x ×L y Discrete grid G xy , the size of each grid is Δx×Δy. Any grid is located by coordinates x and y, where x=1,2,…,L x ,y=1,2,…,L y .

[0017] A forest fire is defined as an area bounded by a spreading fire line. Within the mission area, the set of grids whose centers are within the fire line envelope is taken as the target area to be searched. All grids within the target area to be searched are discrete point targets that are related to each other.

[0018] According to the elliptical fire contour model, it is assumed that the starting point of the forest fire is (x f0 ,y f0 ), the forest fire spread time is t f ≥0, the forest fire line is expressed as:

[0019]

[0020] Among them, (x f ,y f ) is the position coordinate of any point f on the line of fire, θ f ∈[0,2π] is the angle corresponding to point f, that is, the angle of the length of the opposite arc in the ellipse; a f and b f They represent the average spreading speed of forest fire along the x-axis and y-axis in the inertial system, c W Represents the average speed of the ellipse center along the wind direction; φ W is the angle between the wind direction and the positive direction of the x-axis in the inertial system. The major and minor semi-axes of the ellipse are a f t f and b f t f .

[0021] Step 2: Use prior information to initialize the inspection information map, including the fire probability map and the environmental uncertainty map;

[0022] ①Use prior information to obtain the initial fire existence probability map, and use a two-dimensional Gaussian function to describe the initial probability distribution of forest fire.

[0023] The shape of the forest fire is considered as an ellipse, and the coordinates of its geometric center are assumed to be [x c ,y c ] T , then the initial fire existence probability map is expressed as:

[0024]

[0025] Among them, [x c ,y c ] T The probability of fire existing at the peak value p c , σ x and σ yis the variance, which determines the range of the Gaussian distribution. The semi-axes of the ellipse parallel to the x-axis and y-axis in the inertial system are expressed as and When calculating the distance, the grid coordinates x and y need to be converted into the position coordinates of the grid center in the inertial system.

[0026] ② The environmental uncertainty map is represented based on the information entropy of the probability of fire existence;

[0027] The environmental uncertainty graph ξ(x,y,k)∈[0,1] is represented by the information entropy of the corresponding fire probability:

[0028]

[0029] when or , take ξ(x,y,k)=0.

[0030] Step 3: Initially, multiple drones perform patrol tasks according to the initial patrol information map and obtain comprehensive information of the current patrol area;

[0031] The comprehensive information specifically includes: the position and ground speed vector of each drone, forest fire inspection benefits, and the position, shape and size of each environmental obstacle, which are described as follows:

[0032] ① The position and ground speed vector of the drone

[0033] Decouple the three-dimensional motion of the drone into horizontal motion and vertical motion. u The state equation of the UAV at time k is expressed as:

[0034]

[0035] Among them, p i (k) = [x i (k),y i (k),z i (k)] T represents the position of the drone, V h,i is the horizontal velocity amplitude of the i-th UAV; i (k) is the track deviation angle; ω i (k) = Δχ i (k) is the turning angle rate; V z,i and a z,i are the speed and acceleration of the UAV in the vertical direction respectively; ΔT is the planning step length. i and a z,i As control input for the drone.

[0036] ② Forest fire inspection income

[0037] Taking maximizing forest fire patrol benefits as the optimization goal, a multi-UAV collaborative forest fire patrol planning model is established as follows:

[0038]

[0039] p i (k)∈F,i=1,2,...,N u (8)

[0040] In formula (5), p * N u The optimal patrol route of the UAV during the entire patrol mission execution process. N u The optimal patrol route of a UAV at time k, is the optimal planning waypoint of the i-th UAV at time k, p is the solution space, is the total number of planned steps for the inspection task, where T f is the total inspection task time; J is the collaborative inspection benefit function, including the fire detection benefit J f , environmental exploration benefit J e , potential detection benefit J p and synergy benefits J c ; d ,λ e ,λ c is the weight of the corresponding benefit; Formula (6) represents the collision avoidance constraint between aircraft, where d min is the minimum safe distance of the UAV; Formula (7) represents the terrain elevation constraint, z i (k) and z(x i ,y i ) are the flight altitude of UAV i at time k and the terrain height at the grid where it is located; Formula (8) represents the kinematic constraints and terrain constraints, where F is the set of feasible routes considering the corresponding constraints.

[0041] ③Environmental obstacles

[0042] Environmental obstacles are all enveloping with standard convex polyhedrons. After enveloping, the obstacles are uniformly expressed in the inertial system as follows:

[0043]

[0044] Among them, a, b, c and d, e, f are all constants, which determine the size and shape of the obstacle respectively; p o =[x o ,y o ,z o ] T Represents the coordinates of the obstacle center in the inertial system; Γ(x, y, z) = 1 represents the obstacle surface.

[0045] Each drone is given a different priority. Low-priority drones regard high-priority drones as spherical motion obstacles, whose radius is the minimum safe distance of the drone. The drones are called U 1 , U 2 , ..., U Nu .

[0046] Step 4: For the current sampling time, determine whether the total time of the forest fire inspection task has reached T f , if yes, the algorithm ends; otherwise, go to step five;

[0047] Step 5: Estimate the posterior probability distribution of potential forest fire source items based on the detector information of multiple UAVs during the inspection process;

[0048] The detection results of multiple UAVs at time k are recorded as d k (R(k)), the historical detection results of multiple drones before time k are d 1:k-1 ={d 1 (R(1)),d 2 (R(2)),...,d k-1 (R(k-1))}. Estimation of forest fire source term at time k From the posterior probability density function (PDF) Description, that is, when the historical detection result is d 1:k Under the condition of In the Bayesian reasoning framework, It is expressed as:

[0049]

[0050] in, is the normalization coefficient; is the prior probability density function, that is, the source term estimation result at time k-1; is the likelihood, that is, the forest fire source term is The detection result is d k probability;

[0051] Traverse the rasterized task area and combine it with particle filtering (PF) to calculate Specifically: In particle filtering, the posterior probability distribution It is represented by a set of weighted sampling points, i.e. particles. Assuming the number of particles is N, the jth particle at time k is denoted as Represents possible forest fire source items, and its weight express is the probability of the real forest fire source term, that is The posterior probability density function of the forest fire source term is approximately expressed by particle filtering as:

[0052]

[0053] Where δ(·) is the unit pulse function.

[0054] At time k, the particle Its weight The update rule is as follows: Combined with the likelihood The unnormalized weight of the corresponding particle It is expressed as:

[0055]

[0056] After normalization, the weight is:

[0057]

[0058] Get the effective number of particles When N eff When it is less than the threshold, the particles are resampled using the polynomial method; when N eff When it is greater than the threshold, the particles are weighted summed to obtain the forest fire source estimate The formula is:

[0059]

[0060] Step 6: Use Gaussian mixture model GMM to cluster particles containing the posterior probability distribution characteristics of forest fire source items, and calculate potential forest fire source items and their credibility;

[0061] GMM is expressed as the weighted average of L Gaussian distributions, where the mean, covariance and mixing weight corresponding to the lth Gaussian distribution at time k are μ k,l ∈Μ k ={μ k,1 ,μ k,2 ,...,μ k,L},σ k,l ∈∑={σ k,1 ,σ k,2 ,...,σ k,L} and π k,l ∈Π={π k,1 ,π k,2 ,...,π k,L},and

[0062] The particles Treat as a dataset The data points in , that is, the d-dimensional data vector to be clustered (d = 3), then its probability is expressed by GMM as:

[0063]

[0064] in, for The mean and covariance are μ k,l and σ k,l The probability of a Gaussian distribution is:

[0065]

[0066] The GMM method is used for particle clustering decomposition:

[0067] Firstly, the K-means clustering algorithm is used to preliminarily divide the particles and obtain the initial values ​​of the mean, covariance and mixing weight of each Gaussian model;

[0068] Then, iterations are performed based on the maximum expectation method (EM) until the deviation between the two GMM decomposition results is less than a given tolerance.

[0069] Finally, the relationship between the weight of each Gaussian model and the weight threshold is determined. When the Gaussian model weight π in GMM k,l Less than the weight threshold π t When , the model is eliminated; otherwise, the model is retained, and the number of Gaussian models obtained is N G ≤L.

[0070] Based on the Gaussian model parameters obtained by the above iterative decomposition, we can extract the mean μ k,l As the center, π k,l is the mixed weight, corresponding to N within three standard deviations (probability and 99.7%) G particle clusters, and the center μ of each particle cluster k,l As a potential source of forest fires.

[0071] Assess the credibility of potential forest fire sources:

[0072] Let Σ k for The covariance matrix of Then Σ k It reflects the convergence degree of particle filtering. 0 =trace(Σ 0 ) and T k =trace(Σ k ) represent the uncertainty of the estimated forest fire source term at the initial time of the mission and at time k, respectively, and the lth potential forest fire source term μ at time k k,l The credibility of is defined as:

[0073]

[0074] Step 7: Dynamically update the inspection information map based on potential forest fire sources;

[0075] The potential forest fire source item μ l Substitute it into the forest fire spread model to obtain the potential forest fire contour, and then determine the grid G xy Is it within the corresponding potential fire line envelope:

[0076] set up Represents G xy There is fire within the line of fire. Represents G xy There is no fire outside the fire line. xy The probability of fire existing at time k is Combined with the credibility of forest fire source term κ k,l , assuming that other potential forest fire sources are not considered, μ l In G xy The probability increment generated at That is, in G xy When there is no fire, the potential forest fire μ l The probability of fire spreading is

[0077] Considering all potential forest fire sources, in G xy The joint probability increment at is:

[0078]

[0079] Therefore, the dynamic update mechanism of the fire probability map is expressed as:

[0080]

[0081] in, is the threshold for dynamic probability update. When the fire exists in this grid, the probability of fire is not dynamically updated, and the probability is only updated when the UAV detects this grid.

[0082] The dynamic update mechanism of the environmental uncertainty graph is expressed as:

[0083]

[0084] Step 8: Calculate the multi-UAV horizontal route with the highest collaborative inspection benefit based on the rolling horizon optimization and pruning strategy;

[0085] Based on the Receding Horizon Optimization (RHC) framework, the horizontal motion problem of the UAV is transformed into a series of small-scale optimization problems in a short time domain, namely:

[0086]

[0087] st(6),(7)and(8)

[0088] Among them, N r is the rolling time domain step size.

[0089] The steps of rolling horizon optimization are as follows: Assuming the current planning time is k, the greedy algorithm is used to solve the future N of each drone r The optimal patrol route within the step length, and each aircraft only executes the first step, that is, As the planned horizontal speed at the current moment; when reaching the next planned time k+1, repeat the above steps until the inspection process is completed.

[0090] The pruning method is used to solve the multi-machine collaborative planning problem. The steps are as follows:

[0091] (1) For each UAV, calculate the patrol benefits of all feasible horizontal routes in the rolling time domain. This benefit only includes the fire detection benefit J in formula (30): f 、Environmental exploration benefitsJ e and potential detection benefit J p , excluding synergy benefits J c ; Then, select n horizontal routes with the greatest benefit for each drone.

[0092] (2) Arrange and combine the selected UAV routes to obtain (N u ) n The routes of multiple UAVs are grouped and the complete collaborative inspection benefit of each route group is calculated according to formula (30); then, the route combination with the largest benefit is selected as the optimal horizontal planning route for multiple UAVs.

[0093] Step 9: Introduce the IFDS algorithm to plan and obtain feasible three-dimensional waypoints that meet the requirements of terrain following and aircraft collision avoidance at the next moment;

[0094] The process of calculating three-dimensional waypoints using the IFDS algorithm is:

[0095] First, in order to make the drone return to the preset flight altitude after passing through the undulating terrain area, the following altitude maintenance law is designed:

[0096]

[0097] Among them, η z >0 and η z >0 is a constant, and z i (k) are the expected altitude and current altitude of the i-th UAV, respectively.

[0098] Then, the drone regards obstacles such as mountains and hills and other drones with high priority as obstacles, and changes the planned horizontal movement speed to The speed maintained with altitude is used as the confluence velocity u in the IFDS algorithm i (k), namely:

[0099]

[0100] Finally, the disturbed flow velocity is generated based on the IFDS algorithm To achieve terrain following and collision avoidance, there are:

[0101]

[0102] in, and are the disturbance matrix in the IFDS algorithm, the reference speed of the obstacle, and the speed of the jth obstacle. For terrain obstacles, since the UAV needs to detect them, the UAV should climb when the current altitude is lower than the terrain elevation, that is, the directional coefficient in the IFDS algorithm is set to θ i =π / 2.

[0103] Finally, Perform kinematic constraints to obtain the actual planned speed v of the drone i (k).

[0104] The expression of the next desired waypoint of the i-th UAV is:

[0105] p i (k+1)=v i (k)ΔT (34)

[0106] Step 10: Update the inspection information map according to the detection results of the drone during the flight, return to step 3, and plan the multi-drone forest fire inspection route at the next sampling time;

[0107] During the forest fire patrol process, the drone continuously updates the probability of fire in the detected area based on the sensor detection results, combined with the sensor detection probability and false alarm probability According to the Bayesian criterion, the updating process is expressed as:

[0108]

[0109] Among them, d(x,y,k)∈{0,1} is the sensor’s observation of the grid G ​​at time k. xy The detection result of G xyIf the value is less than 1, it means that a fire is detected, otherwise it means that no fire is detected. d(x,y,k) is a random event and obeys the following probability distribution:

[0110]

[0111] in, Represents the probability of a random event.

[0112] The update formula of the environmental uncertainty graph ξ(x,y,k) is:

[0113]

[0114] The advantages of the present invention are:

[0115] (1) The present invention provides a multi-UAV high-efficiency patrol method for low-altitude environments based on terrain elevation maps, which can fully combine the likelihood information of forest fire spread models and UAV detection information, thereby greatly improving patrol efficiency;

[0116] (2) The present invention provides a multi-UAV efficient patrol method in a low-altitude environment based on a terrain elevation map, and designs a dynamic update mechanism for a fire patrol information map, which can significantly improve the accuracy of the patrol information map;

[0117] (3) The present invention provides an efficient multi-UAV patrol method in a low-altitude environment based on a terrain elevation map, which can meet the needs of terrain following and collision avoidance between aircraft and is closer to actual application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0118] Figure 1 It is an overall flow chart of the high-efficiency inspection method of multiple UAVs in a low-altitude environment based on a terrain elevation map of the present invention;

[0119] Figure 2 It is a schematic diagram of a multi-UAV forest fire efficient inspection planning in a low-altitude environment multi-UAV efficient inspection method based on a terrain elevation map of the present invention;

[0120] Figure 3 It is a schematic diagram of the discrete movement directions of drones in the horizontal plane in the high-efficiency inspection method of multiple drones in a low-altitude environment based on a terrain elevation map of the present invention;

[0121] Figure 4 It is a schematic diagram of terrain following of drones in the high-efficiency inspection method of multiple drones in a low-altitude environment based on a terrain elevation map of the present invention;

[0122] Figure 5 is a collaborative forest fire patrol planning result diagram based on the raster patrol method in an embodiment of the present invention (t=7000s); wherein, Figure 5 (a) is the probability map of fire existence; Figure 5 (b) is the environmental uncertainty diagram; Figure 5(c) is a three-dimensional route; Figure 5 (d) is the projection of the route on the horizontal plane; Figure 5 (e) is a magnified image of the forest fire area; Figure 5 (f) is the fire situation determination result;

[0123] Figure 6 is a collaborative forest fire patrol planning result diagram based on the standard probabilistic heuristic patrol method in an embodiment of the present invention (t=7000s); wherein, Figure 6 (a) is the probability map of fire existence; Figure 6 (b) is the environmental uncertainty diagram; Figure 6 (c) is a three-dimensional route; Figure 6 (d) is the projection of the route on the horizontal plane; Figure 6 (e) is a magnified image of the forest fire area; Figure 6 (f) is the fire situation determination result;

[0124] Figure 7 is the forest fire patrol planning result based on the method proposed in the embodiment of the present invention; Figure 7 (a) is the result of the coordinated forest fire patrol planning when the fire was first confirmed (t = 2720s); Figure 7 (b) is the result of the coordinated forest fire inspection planning when the fire confirmation rate reaches 50.63% (t = 3740s); Figure 7 (c) is the collaborative forest fire patrol planning result diagram at the end of the mission (t = 7000s);

[0125] Figure 8 is a distance diagram between each UAV and the nearest obstacle surface when the proposed method is adopted in the embodiment of the present invention;

[0126] Fig. 9 is a time distance diagram when the proposed method is used in an embodiment of the present invention;

[0127] Fig.10 is a vertical motion information diagram of each UAV when the proposed method is adopted in the embodiment of the present invention; wherein, Fig.10 (a) is the flight altitude; Fig.10 (b) is the vertical velocity; Fig.10 (c) is the vertical acceleration;

[0128] Fig.11 is a comparison chart of average fire confirmation rates of different methods in the embodiments of the present invention;

[0129] Fig.12 It is a comparison chart of the average single-step calculation time consumption of different methods in the embodiments of the present invention. DETAILED DESCRIPTION

[0130] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0131] The present invention proposes an efficient multi-UAV patrol method in a low-altitude environment based on a terrain elevation map. The forest fire patrol information map is dynamically updated based on the online estimation of the unknown fire situation, and a three-dimensional collaborative forest fire patrol planning method is designed, so that the flight altitude of the UAV can change with the terrain undulations, which significantly improves the forest fire patrol efficiency while ensuring the flight safety of the UAV.

[0132] An efficient multi-UAV inspection method for low-altitude environments based on terrain elevation maps. The overall process is as follows: Figure 1 As shown, the specific steps include:

[0133] Step 1: rasterize the task area;

[0134] Assuming that the task area is a rectangle, in the target search problem, in order to simplify the solution space, the task area is usually rasterized, that is, it is evenly divided into L x ×L y Discrete grid G xy , the size of each grid is Δx×Δy. Generally speaking, the size of the divided grid depends on the detection range S of the airborne sensor. Any grid can be located by the coordinates x and y, where x = 1, 2, ..., L x ,y=1,2,...,L y .

[0135] In the present invention, as an unknown target to be searched, forest fire can be defined as an area bounded by a spreading fire line. After rasterizing the mission area, the forest fire can be regarded as a collection of grids with fire conditions, i.e., grids with centers located within the fire line envelope. These grids can be regarded as discrete point targets that are related to each other. Figure 2 shown.

[0136] According to the elliptical fire contour model, it is assumed that the starting point of the forest fire is (x f0 ,y f0 ), the forest fire spread time is t f ≥0, the forest fire line can be expressed as:

[0137]

[0138] Among them, (x f0 ,y f0 ) is the position coordinate of any point f on the line of fire, θ f ∈[0,2π] is the angle corresponding to point f, that is, the angle of the length of the opposite arc in the ellipse; a f and b f They represent the average spreading speed of forest fire along the x-axis and y-axis in the inertial system, cW Represents the average speed of the ellipse center along the wind direction; φ W is the angle between the wind direction and the positive direction of the x-axis in the inertial system. It can be seen that the major and minor semi-axes of the ellipse are a and f t f and b f t f The rate at which a forest fire spreads can usually be calculated based on the uniform availability of fuel, terrain, and wind conditions in the patrol area.

[0139] Step 2: N who will participate in the forest fire patrol mission u The drones are numbered 1, 2, ..., N. u , and determine the parameters of the onboard sensors;

[0140] In this embodiment, N u =4. Regardless of the drone's attitude, the optical sensor's sight line is always pointing vertically downward due to the stable platform. Assuming that the sensor's detection range is constant at 1 grid directly below the drone, then S = Δx × Δy. In addition, since the detection results are affected by factors such as sensor errors and environmental occlusion, the key parameters of the sensor also include the detection probability and false alarm probability in, It indicates the probability that the sensor detects the target when the target (i.e., fire) exists. It represents the probability that the sensor detects the target when the target does not exist.

[0141] Step 3: Initialize the inspection information map;

[0142] In order to achieve efficient collaborative patrol planning, it is necessary to extract key information in the task area, such as the probability and uncertainty of fire. To this end, the present invention introduces a patrol information map including a fire probability map and an environmental uncertainty map, that is, a mapping relationship between each grid and the information required for patrol planning is established, and it is continuously updated as the detection information continues to accumulate during the patrol process.

[0143] Assume that at time k, the grid G xy The probability of fire existing at in Represents G xy There is no fire at this time. According to the prior information of the fire department or forestry department on the distribution of forest fires in the area to be inspected, the initial fire probability map can be obtained. The shape of a forest fire can be approximated as an ellipse, so a two-dimensional Gaussian function can be used to describe its initial probability distribution. Assume that the coordinates of the geometric center of the ellipse are [x c ,y c ]T , then the probability diagram of fire existence at the initial moment can be expressed as:

[0144]

[0145] Among them, [x c ,y c ] T The probability of fire existing at the peak value p c , σ x and σ y is the variance, which determines the range of the Gaussian distribution. The semi-axes of the ellipse parallel to the x-axis and y-axis in the inertial system can be expressed as and It should be noted that when calculating the distance, the grid coordinates x and y need to be converted into the position coordinates of the grid center in the inertial system.

[0146] The environmental uncertainty graph ξ(x,y,k)∈[0,1] reflects the degree of the drone’s understanding of information in the unknown environment. The smaller ξ(x,y,k), the greater the uncertainty of G. xy The smaller the uncertainty at the location, the more accurate the drone’s understanding of its information. The environmental uncertainty map can be represented by the information entropy of the corresponding fire probability, that is:

[0147]

[0148] when or , take ξ(x,y,k)=0.

[0149] Step 4: When multiple UAVs are performing forest fire inspection tasks, obtain comprehensive information about the current inspection area;

[0150] The comprehensive information specifically includes: the position and ground speed vector of each UAV, forest fire inspection benefits, and the position, shape and size of each environmental obstacle. The mathematical description of the comprehensive information is as follows.

[0151] ① Combined with the task requirements of forest fire inspection, the three-dimensional motion of the UAV can be decoupled into horizontal motion and vertical motion. u The state equation of the UAV at time k can be expressed as:

[0152]

[0153] Among them, p i (k) = [x i (k),y i (k),z i (k)] T represents the position of the drone, V h,iis the horizontal velocity amplitude of the i-th UAV; i (k) is the track deviation angle; ω i (k) = Δχ i (k) is the turning angle rate; V z,i and a z,i are the speed and acceleration of the UAV in the vertical direction respectively; ΔT is the planning step length. i and a z,i As the control input of the UAV, the horizontal and vertical movements of the UAV will be discussed separately below.

[0154] Assume that at each sampling moment, the drone is located directly above the center of a grid, and can only move to the adjacent grid of the current grid within a sampling period. Therefore, the horizontal motion space of the drone can be discretized, that is, without considering the kinematic constraints, the drone has 8 optional horizontal motion directions at time k, with χ i (k)∈{1,2,3,4,5,6,7,8}, the horizontal movement direction corresponding to each number is as follows Figure 3 To ensure the feasibility of the planned route, assuming that the feasible heading angle deflection range of the UAV within a planning step is ±45°, then χ i (k+1)∈{χ i (k)-1,χ i (k),χ i (k)+1}mod 8, where mod is the modulus operator. In the present invention, when the horizontal motion direction of the drone is parallel to the grid diagonal, the direction is called oblique; otherwise, it is called positive. Assume that the average velocity amplitudes of oblique motion and positive motion are V s and V f It should be noted that in order to reach the corresponding adjacent grid centers within the same planning step, the ratio of the average velocity amplitude of the oblique motion and the forward motion should be Right now

[0155] To ensure flight safety and comprehensive inspection, the flight altitude of the drone should be higher than the terrain height corresponding to the current grid, that is, the terrain following requirement should be considered during route planning. Since the terrain height in the mission environment is arbitrary, in order to achieve terrain following, the vertical motion space of the drone should be continuous and satisfy the constraint V z,i ∈[V zmin ,V zmax ] and a z,i ∈[a zmin ,a zmax ].

[0156] ② The collaborative forest fire patrol planning problem can essentially be reduced to an optimization problem. Considering the kinematic constraints and flight safety of UAVs, and taking maximizing the benefits of forest fire patrols as the optimization goal, a multi-UAV collaborative forest fire patrol planning model is established as follows:

[0157]

[0158]

[0159] p i (k)∈F,i=1,2,…,N u (8)

[0160] In formula (5), p * N u The optimal patrol route of the UAV during the entire patrol mission execution process. N u The optimal patrol route of a UAV at time k, is the optimal planning waypoint of the i-th UAV at time k, p is the solution space, is the total number of planned steps for the inspection task, where T f and ΔT are the total patrol task time and planning step length respectively; J is the collaborative patrol benefit function, including the fire detection benefit J f , environmental exploration benefit J e , potential detection benefit J p and synergy benefits J c ; d ,λ e ,λ c is the weight of the corresponding benefit; Formula (6) represents the collision avoidance constraint between aircraft, where d min is the minimum safe distance of the UAV; Formula (7) represents the terrain elevation constraint, z i (k) and z(x i ,y i ) are the flight altitude of UAV i at time k and the terrain height at the grid where it is located, which is approximated by the grid center height; Formula (8) represents the kinematic constraints and terrain constraints, where F is the set of feasible routes considering the corresponding constraints.

[0161] In the inspection profit function, the specific definitions of each sub-profit function are as follows:

[0162] Fire detection benefits f It represents the weighted sum of the probability of fire existing in the area detected by the drone during the inspection process, reflecting the possibility of detecting undetermined fires, and is used to guide the drone to focus on detecting areas with a high probability of undetermined fires. f The expression is:

[0163]

[0164] Among them, R i (k) is the detection range of the i-th UAV at time k, i.e., the UAV waypoint φ i (k) The grid directly below; is the grid G xy The probability of fire existing at time k; ζ(x, y, k) is the ... xy The fire condition judgment signs at time k are:

[0165]

[0166] Among them, ζ(x, y, k) = 1 means that it has been determined that in G xy There is a fire at this location, otherwise it is considered that there is no fire at this location. p is the probability threshold of fire. When a fire is determined to exist, G xy The fire detection benefit at the location is 0, which means that it is not encouraged to detect this location again. xy When there is a fire at G, it is called fire confirmation. The fire confirmation rate is defined as the percentage of the number of grids confirmed to have a fire to the number of grids that actually have a fire. xy When a fire is detected but not actually present, it is called a fire misjudgment. The fire misjudgment rate is defined as the percentage of the number of grids misjudged to have a fire to the total number of grids judged to have a fire.

[0167] Environmental exploration benefits e It represents the environmental uncertainty of the area detected by the drone, and is used to guide the drone to detect areas with higher uncertainty. Its expression is:

[0168]

[0169] Potential Exploration BenefitJ p It represents the uncertainty of the potential detection area after the drone executes the current planned route. This benefit focuses on the overall inspection task and guides the drone to approach areas with greater potential detection value, in order to obtain greater long-term returns in subsequent planning. p The expression is:

[0170]

[0171] Among them, A i (k) is the potential detection area of ​​the i-th UAV at time k, is the number of grids in the area. Combined with the horizontal action space of the drone, the entire mission area can be evenly divided into 9 sub-areas. After the drone executes the planned route at time k, its distance relative to the grid at the center of each sub-area will change. When the relative distance decreases, it means that the drone is approaching the corresponding sub-area. In the present invention, the three sub-areas with the largest relative distance reduction (excluding the sub-area where the drone is currently located) are selected as potential detection areas A. i (k). Similar to step 3, the grid coordinates need to be converted here.

[0172] During the collaborative inspection process, the detection areas of the drones may overlap, which will not only reduce the efficiency of task execution, but also increase the risk of collision between drones. To reduce this phenomenon, the collaborative benefit J can be introduced. c Assumptions is the detection range of the drone group at time k. By merging the repeated elements in this set, we can get its subset R s (k), which represents the non-overlapping detection range of the drone swarm during the inspection process. Let A(k) be the potential detection range of the drone swarm, then J c It can be expressed as:

[0173]

[0174] In low-altitude mountainous environments, there may be both static obstacles such as terrain and dynamic obstacles such as intruders, among which dynamic obstacles can usually be approximated as spheres. In the present invention, all obstacles are enveloping with standard convex polyhedrons such as spheres, hemispheres, cylinders, and cuboids. After being enveloping with convex polyhedrons, the obstacles can be uniformly expressed in the inertial system as:

[0175]

[0176] Among them, a, b, c and d, e, f are all constants, which determine the size and shape of the obstacle respectively; p o =[x o ,y o ,z o ] T Represents the coordinates of the obstacle center in the inertial system; Γ(x, y, z) = 1 represents the obstacle surface.

[0177] In order to avoid collisions between formation members and chain reactions caused by collision avoidance maneuvers, each drone is given a different priority. Low-priority drones regard high-priority drones as spherical motion obstacles, whose radius is the minimum safe distance of the drone. The drones are called U 1 , U 2 , …,

[0178] Step 5: For the current sampling time, determine whether the total time of the forest fire inspection task has reached T f , if yes, the algorithm ends; otherwise, go to step 6;

[0179] Step 6: Estimate the posterior probability distribution of potential forest fire source terms;

[0180] In the forest fire inspection problem, due to the continuous spread of forest fires, there is a correlation between the targets to be searched, that is, the grids where fire conditions exist. By making full use of the correlation between fire conditions, it is possible to greatly improve the efficiency of forest fire inspections. The fire conditions are estimated to provide a reasonable basis for updating the fire condition probability map. To this end, the present invention first estimates the actual fire conditions based on the Bayesian reasoning framework combined with particle filtering. Then, based on the estimation results, the fire condition probability map and the environmental uncertainty map are dynamically updated, so as to describe the current situation more timely and accurately, and provide a strong basis for efficient inspection planning.

[0181] According to the forest fire spread model in step 2, it can be known that the forest fire contour (x f ,y f ) parameters include the location of the fire point (x f0 ,y f0 ), forest fire spread time t f , the spreading speed of the forest fire along the x-axis and y-axis in the inertial system is a f and b f , the speed of the ellipse center along the wind direction c W , and wind direction angle φ W Among these parameters, (x f0 ,y f0 ) and t f Unknown, parameter a related to the speed of forest fire spread f , b f , c W and φ W It depends on the fuel bed, topography, weather, etc. of the search area, which can be obtained based on local geographical and meteorological data. f =[x f0 ,y f0 ,t f ] T By making an estimate, we can get an estimate of the forest fire contour, thereby guiding the drone to focus on patrolling high-risk areas in a targeted manner and improve the efficiency of mission execution.

[0182] In step 4, the detection range of multiple drones at time k is defined as The detection results of multiple drones at time k are recorded as d k (R(k)), the historical detection results of multiple drones before time k are d1:k-1 ={d 1 (R(1)),d 2 (R(2)),...,d k-1 (R(k-1))}. Assume that parameter a f , b f , c W and φ W It has been obtained based on external information. The forest fire source term estimation at time k The posterior probability density function (PDF) Description, that is, when the historical detection result is d 1:k Under the condition of It should be noted that the probability here is the same as

[0183] Different, the latter only means that in the grid G xy The probability of fire existing at the location. Under the Bayesian reasoning framework, It can be expressed as the following recursive form:

[0184]

[0185] in, is the normalization coefficient; is the prior probability density function, that is, the source term estimation result at time k-1; is the likelihood, that is, the forest fire source term is The detection result is d k Combining the forest fire spread model, we have:

[0186]

[0187] Among them, f(x,y,k)∈{0,1} means that when the forest fire source term is Is there a fire at (x,y)? Substituting into the forest fire spread model, the corresponding forest fire outline, i.e., the fire line, can be obtained. When (x, y) is outside the fire line, f(x, y, k) = 0; otherwise, f(x, y, k) = 1.

[0188] The expression is:

[0189]

[0190] According to the above recursive formula, by traversing the rasterized task area, we can get To reduce the computational burden, the present invention combines particle filtering (PF) to calculate Particle filtering is also known as sequential importance resampling (SIR), which is suitable for state estimation of nonlinear and non-Gaussian systems. In particle filtering, the posterior probability distribution It can be represented by a set of weighted sampling points, i.e. particles. Assuming the number of particles is N, the jth particle at time k is denoted as Represents possible forest fire source items, and its weight express is the probability of the real forest fire source term, that is Ideally, particles are sampled from the unknown actual forest fire source term distribution. According to the importance sampling method, the target distribution can be approximated by a known distribution, such as the prior estimated forest fire source term distribution, which is called an important distribution or a suggested distribution. In summary, the posterior probability density function of the forest fire source term can be approximated by the particle filter as follows:

[0191]

[0192] Where δ(·) is the unit pulse function.

[0193] At time k, the particle and its weight The update rule of is as follows. Since the location of the fire point remains unchanged, and the forest fire spread time is monotonically increasing, we have Combined likelihood The unnormalized weight of the corresponding particle It can be expressed as:

[0194]

[0195] After normalization, the weight is:

[0196]

[0197] In particle filtering, as the number of iterations increases, particle degradation will occur, that is, the weights of most particles will become very small, and the relative weights of a few particles will be very large, resulting in a decrease in particle diversity. To address this phenomenon, the effective number of particles is taken as When N eff When it is less than a certain threshold, the particles are resampled using the polynomial method.

[0198] When N eff When it is greater than the threshold, the forest fire source item The estimate can be obtained by the weighted sum of particles:

[0199]

[0200] Step 7: Calculate potential forest fire source items and their credibility;

[0201] The above particles contain the posterior probability distribution characteristics of the forest fire source items. In order to quantify and extract them and facilitate the dynamic update of the inspection information map, the present invention uses a Gaussian mixture model (GMM) to cluster the particles and then obtain the potential forest fire source items in the mission area.

[0202] GMM can be expressed as a weighted average of L Gaussian distributions, where the mean, covariance and mixing weight corresponding to the lth Gaussian distribution at time k are μ k,l ∈Μ k ={μ k,1 ,μ k,2 ,...,μ k,L},σ k,l ∈∑={σ k,1 ,σ k,2 ,...,σ k,L} and π k,l ∈Π={π k,1 ,π k,2 ,...,π k,L},and

[0203] In the present invention, the particles Treat as a dataset The data points in , i.e., the d-dimensional data vector to be clustered (d=3 in the present invention), then their probability can be expressed by GMM as:

[0204]

[0205] in, for The mean and covariance are μ k,l and σ k,l The probability of a Gaussian distribution is:

[0206]

[0207] When using the GMM method for particle clustering decomposition, the K-means clustering algorithm is first used to preliminarily divide the particles to obtain the initial values ​​of the mean, covariance, and mixing weight; then, it is iterated based on the maximum expectation method (EM) until the deviation between the two GMM decomposition results is less than the given tolerance. It should be noted that when the Gaussian model weight π in the GMM k,l Too small is π k,l <π t When , the model can be eliminated. In the present invention, the weight threshold is π t =0.05. Therefore, the number of Gaussian models N obtained in the end G ≤L.

[0208] Based on the Gaussian model parameters obtained by the above iterative decomposition, we can extract the mean μk,l As the center, π k,l is the mixed weight, corresponding to N within three standard deviations (probability and 99.7%) G The shape of the particle cluster is the same as the outline of the forest fire, which is an ellipse. k,l As a potential source of forest fires.

[0209] The credibility of potential forest fire source items is further evaluated. k for The covariance matrix of Then Σ k It reflects the convergence degree of particle filtering. 0 =trace(Σ 0 ) and T k =trace(Σ k ) can represent the uncertainty of the forest fire source item estimation at the initial time of the mission and at time k, respectively, and thus reflect its credibility. The smaller their values ​​are, the smaller the uncertainty is and the greater the credibility is. Therefore, the lth potential forest fire source item μ at time k is k,l The credibility of can be defined as:

[0210]

[0211] As the inspection progressed, T k / T 0 The overall trend is decreasing, which means that the uncertainty of the forest fire source estimation is reduced and the credibility is increased. In the early stage of the patrol mission, due to the randomness of particle distribution, T k / T 0 ≥1 situation.

[0212] Step 8: Dynamically update the inspection information map based on potential forest fire sources;

[0213] The potential forest fire source item μ l Substituting into the forest fire spread model, the potential forest fire outline can be obtained, and then the grid G ​​can be judged. xy Is it within the corresponding potential fire line envelope? Represents G xy There is fire within the line of fire. Represents G xy There is no fire outside the fire line. xy The probability of fire existing at time k is Combined with the credibility of forest fire source term κ k,l , assuming that other potential forest fire sources are not considered, μ l In G xy The probability increment generated at That is, in G xyWhen there is no fire, the potential forest fire μ l The probability of fire spreading is Considering all potential forest fire sources, in G xy The joint probability increment at is:

[0214]

[0215] Therefore, the dynamic update mechanism of the fire probability map can be expressed as:

[0216]

[0217] in, is the threshold for dynamic probability update. When , the probability of fire in this grid is not dynamically updated, and the probability is only updated when the drone detects this grid. When the grid is large and there is no fire in reality, if the detection result shows a false alarm, the drone is likely to misjudge the fire. Therefore, it is necessary to set a dynamic probability update threshold so that the drone can further increase the probability of fire existence only through detection and finally accurately determine the fire.

[0218] The dynamic update mechanism of the environmental uncertainty graph can be expressed as:

[0219]

[0220] Step 9: Calculate the multi-UAV horizontal route corresponding to the highest collaborative inspection benefit;

[0221] In the mountain forest fire patrol problem, there are special terrains such as peaks and hills. The UAV needs to change the flight altitude to ensure that the field of view of the downward sensor can cover the special terrain. To this end, the present invention proposes a three-dimensional collaborative patrol planning method based on information graphs and terrain elevation to achieve online optimization and terrain following of the patrol route, and eliminate collision conflicts between drones.

[0222] First, the influence of terrain elevation is not considered, and only the horizontal movement of the drone is considered. The solution space of the original optimization problem is too large, and due to the dynamic changes of forest fire conditions and patrol information maps, it is difficult to solve the problem directly. In order to plan a patrol route with high quality online, based on the rolling horizon optimization (RHC) framework, the original problem is transformed into a series of small-scale optimization problems in a short time domain, namely:

[0223]

[0224] st(6),(7)and(8)

[0225] Among them, N r is the rolling time domain step size.

[0226] The steps of rolling horizon optimization are as follows: Assuming the current planning time is k, the greedy algorithm can be used to solve the future N of each drone r The optimal patrol route within the step length, and each aircraft only executes the first step, that is, As the planned horizontal speed at the current moment; when reaching the next planned time k+1, repeat the above steps until the inspection process is completed.

[0227] In the multi-machine collaborative planning problem, the size of the solution space grows exponentially with the number of drones. In order to speed up the solution, the pruning method is used to reduce the amount of calculation. The steps are as follows:

[0228] (1) For each UAV, calculate the patrol benefits of all feasible horizontal routes in the rolling time domain. This benefit only includes the fire detection benefit J in formula (30): f 、Environmental exploration benefitsJ e and potential detection benefit J p , excluding synergy benefits J c ; Then, n horizontal routes with the greatest benefit for each UAV are selected. In the present invention, n=5.

[0229] (2) Arrange and combine the selected UAV routes to obtain (N u ) n The routes of multiple UAVs are grouped and the complete collaborative inspection benefit of each route group is calculated according to formula (30); then, the route combination with the largest benefit is selected as the optimal horizontal planning route for multiple UAVs.

[0230] Step 10: Plan and obtain feasible three-dimensional waypoints that meet the requirements of terrain following and aircraft collision avoidance at the next moment;

[0231] The planned route obtained through the above steps does not take into account the altitude change and flight safety of the drone. For this reason, the IFDS algorithm is introduced in the present invention to achieve terrain following and collision avoidance between multiple drones during forest fire inspection. In order to make the drone return to the preset flight altitude after passing through the undulating terrain area, the following altitude keeping law is designed:

[0232]

[0233] Among them, η z >0 and η z >0 is a constant, and z i (k) are the expected height and current height of the ith UAV, respectively. To reduce the risk of collision between drones, the expected height of each UAV can be taken as different values.

[0234] The drone considers obstacles such as mountains and hills and other drones with high priority as obstacles, and the planned horizontal movement speed The speed maintained with altitude is used as the confluence velocity u in the IFDS algorithm i (k), namely:

[0235]

[0236] Then, the disturbed flow velocity is generated based on the IFDS algorithm To achieve mountain terrain following and collision avoidance, there are:

[0237]

[0238] in, and are the disturbance matrix, obstacle reference speed and the speed of the jth obstacle in the IFDS algorithm. Figure 4 As shown in the figure, for terrain obstacles, since the UAV still needs to detect them, the UAV should climb when the current altitude is lower than the terrain elevation, that is, the directional coefficient in the IFDS algorithm is taken as θ i =π / 2.

[0239] Finally, Perform kinematic constraints to obtain the actual planned speed v of the drone i (k).

[0240] The expression of the next desired waypoint of the i-th UAV is:

[0241] p i (k+1)=v i (k)ΔT (34)

[0242] Step 11: Update the inspection information map according to the detection results of the drone during flight;

[0243] During the forest fire patrol process, the drone continuously updates the probability of fire in the detected area based on the sensor detection results, combined with the sensor detection probability and false alarm probability According to the Bayesian criterion, this updating process can be expressed as:

[0244]

[0245] Among them, d(x,y,k)∈{0,1} is the sensor’s observation of the grid G ​​at time k. xy The detection result of G xy If the value is less than 1, it means that a fire is detected, otherwise it means that no fire is detected. d(x,y,k) is a random event and obeys the following probability distribution:

[0246]

[0247] in, Represents the probability of a random event.

[0248] The update formula of the environmental uncertainty graph ξ(x,y,k) is:

[0249]

[0250] Step 12: Return to step 4 to plan the multi-UAV forest fire inspection route for the next sampling time.

[0251] Example

[0252] In order to verify the effectiveness of the proposed method, the present invention conducts comparative simulation verification for the scene of multi-UAV collaborative patrol of forest fires. The comparison methods are selected as the standard probabilistic heuristic patrol method and the raster patrol method, wherein the former is the dynamic information map update mechanism based on fire estimation removed from the proposed method. To ensure fairness, the terrain following and inter-machine collision avoidance mechanisms are introduced into the comparison method as in the proposed method. The indicators for evaluating the performance of the method are selected as the fire discovery time, the fire confirmation rate, the fire misjudgment rate and the average single-step calculation time. Among them, the fire discovery time is defined as the time required for the first confirmation of the fire, and the fire confirmation rate and the fire misjudgment rate are defined as described in step four.

[0253] The planning results obtained by the raster inspection method are as follows: Figure 5 As shown. This method divides the task area evenly into N u The sub-areas are assigned to each drone, and each drone conducts a full coverage inspection of the fire situation in the corresponding sub-area. The fire probability map and environmental uncertainty map at the end of the inspection are shown in the figure below: Figure 5 (a) and 5(b). Combined with the patrol route, it can be seen that the probability of fire and environmental uncertainty at the grids visited by the drone will change significantly compared to the initial state, but this method does not use prior information and patrol information maps, and does not focus on searching high-value areas within the limited patrol time, so the overall change of the patrol information map is small. Figure 5 The squares on the UAV routes in (c) and 5(d) represent the positions of the UAVs at the current moment. It can be seen that after combining the proposed terrain following method, each UAV can adjust its flight altitude following the terrain undulations, thus achieving avoidance and detection of terrain obstacles with little impact on the horizontal route. Figure 5(e) and 5(f) are the enlarged image of the forest fire area and the fire judgment result, respectively. The green circle and red circle in the latter represent the grids with and without fire, respectively. When a fire is judged, a red cross will be marked on the corresponding grid. It can be seen that only UAV 4 visited the forest fire area, and the fire was not judged in the end. The fire confirmation rate is 0%, and there is no misjudgment rate of fire. The average single-step calculation time is 0.0183s, which is much less than the planned step length of 20s, and can meet the real-time requirements of the collaborative inspection task.

[0254] The planning results obtained by the standard probabilistic heuristic inspection method are as follows: Figure 6 As shown. Figure 6 (a) and Figure 6 (b) It can be seen that compared with the raster inspection method, the inspection information map changes more significantly after using this method. Figure 6 (c) and Figure 6 (d) It can be seen that this is because during the inspection process, the multiple drones did not conduct a full coverage inspection of the vast mission area, but quickly arrived near the high-value areas determined by prior information based on the inspection information map and carried out key inspections there; in this process, the drones were able to achieve terrain following. Figure 6 (e) and Figure 6 (f) It can be seen that multiple drones can only patrol a small part of the forest fire area. This is because there is a large deviation between the prior information of the fire situation and the actual initial situation, and the fire situation is constantly developing dynamically. In the standard probabilistic heuristic patrol method, the fire situation is not estimated online and the patrol information map is dynamically updated on this basis. The patrol efficiency needs to be further improved. When this method is used, the fire detection time is 2820s, the fire confirmation rate and misjudgment rate are 16.99% and 0% respectively, which is better than the raster patrol method; the average single-step calculation time is 0.5932s, which is greater than the raster patrol method, but still much smaller than the planning step size, which can meet the real-time requirements.

[0255] The forest fire inspection planning results using the method proposed by the present invention are as follows: Figure 7 As shown, to explain in detail the update of fire situation estimation and patrol information map, Figure 7 (a) Figure 7 (b) and Figure 7 (c) shows the planning results when the fire is first confirmed (t = 2720s), when the fire confirmation rate reaches 50.63% (t = 3740s), and when the task is completed (t = 7000s). When the fire is first confirmed, due to the dynamic information graph update mechanism, the information graph changes significantly compared to the initial state, such as Figure 7 (a1) Fire existence probability map and Figure 7 (a2) Environmental uncertainty diagram. Figure 7 (a3) Three-dimensional routes and Figure 7(a4) The route is visible when projected on the horizontal plane. The planned route at this time is similar to the result of using the standard heuristic method, but the proposed method uses the cumulative detection results of multiple drones in the patrol area, so that the estimated fire line continues to approach the real fire line. Figure 7 (a5) is an enlarged view of the forest fire area and Figure 7 (a6) Fire situation determination result. During the inspection, there was one misjudgment of the fire situation, and UAV 1 confirmed the fire situation for the first time when it arrived over the real fire line.

[0256] like Figure 7 As shown in (b), as the fire situation is continuously confirmed during the inspection process, the accuracy of fire situation estimation continues to improve, and the key areas of multi-UAV inspections gradually tend to the real forest fire areas. This trend in turn makes it more likely for multiple UAVs to confirm the fire situation, thus forming a virtuous circle and greatly improving the inspection efficiency.

[0257] according to Figure 7 (c) It can be seen that the probability distribution of the fire at this time is consistent with the real forest fire outline, and the estimated fire line can approximate the real fire line well, indicating that the fire estimation and dynamic information map updating method proposed in the present invention has certain effectiveness. In the later stage of the patrol mission, the patrol routes of multiple drones are mainly concentrated over the dynamically expanding real forest fire area, indicating that the patrol process has a reasonable transition from the exploration stage to the utilization stage, thus achieving the effect of "highlighting the key points and taking into account the general". In the end, the fire discovery time using the proposed method was 2720s, and the fire confirmation rate was 82.42%, which was much better than the two comparison methods. The fire misjudgment rate was 0.24%, indicating that the misjudgment results accounted for a very small number of all the judgment results. The average single-step calculation time was 0.6387s, which was slightly longer than the standard heuristic search method, but it could still meet the real-time requirements of the patrol mission well.

[0258] When the method proposed by the present invention is used, the distance between each UAV and the nearest terrain obstacle surface is as follows: Figure 8 As shown in the figure, the closest distance between each drone is greater than the minimum safety distance of 10m. Therefore, the multiple drones did not collide with terrain obstacles during the entire inspection process, indicating that the proposed method has a good terrain following effect. Fig. 9 As shown, it can be seen that the minimum distance between aircraft is greater than 10m, which can meet the requirements for collision avoidance between aircraft.

[0259] The vertical motion information obtained by the method proposed by the present invention is as follows: Fig.10As shown in the figure. In the early stage of the mission, due to the large number of terrain obstacles on the patrol route, each drone needs to follow the terrain, so the flight altitude, vertical speed and acceleration vary greatly. When multiple drones arrive near the forest fire area, due to the relatively flat terrain, the vertical movement of the drones is mainly caused by the requirements of collision avoidance between drones, which is relatively gentle; the flight altitude fluctuates around the expected altitude, and the vertical speed and acceleration fluctuate slightly around zero. During the entire patrol process, the vertical speed and acceleration meet the kinematic constraints, indicating that the planned route is feasible.

[0260] Due to the randomness of particle filtering and sensor detection results, in order to further verify the performance of the proposed method, 50 Monte Carlo simulations were performed for each method. The statistical results of various indicators are shown in Table 1. In terms of fire detection time, the proposed method is similar to the standard probabilistic heuristic patrol method, and both are far superior to the raster patrol method. The latter cannot determine the fire within the specified task time and has a very low patrol efficiency. Due to the introduction of a dynamic information graph update mechanism based on fire estimation, the proposed method is far superior to the two comparison methods in terms of fire confirmation rate, and the misjudgment rate is very low. In terms of single-step calculation time, the proposed method is similar to the standard probabilistic heuristic patrol method, both are greater than the raster patrol method, but all three are far less than the planning step size, and have good real-time performance.

[0261] Table 1 Comparison of forest fire inspection performance under 50 Monte Carlo simulations

[0262]

[0263] The comparison of the average fire confirmation rate and the average single-step calculation time when different methods are adopted are as follows: Fig.11 and Fig.12 As shown. It can be seen that the method proposed in the present invention is far superior to the comparative method in terms of forest fire inspection efficiency, and the additional computational burden is relatively small. It should be noted that when using the method proposed in the present invention, in the early stage of forest fire inspection, due to the large particle update amplitude and high resampling frequency, the average single-step calculation time is relatively long; with the improvement of the accuracy of forest fire source estimation, the particle update amplitude becomes smaller and the resampling frequency is lower, so the average single-step calculation time is relatively short in the early stage.

[0264] In summary, for the collaborative forest fire patrol task in a three-dimensional environment, the method proposed in the present invention can reasonably balance utilization and exploration while meeting the requirements of terrain following and collision avoidance between machines, thereby significantly improving the efficiency of forest fire patrol and providing a strong guarantee for emergency management.

Claims

1. A method for efficient multi-UAV inspection in low-altitude environment based on terrain elevation map, characterized in that: The specific steps include: Step 1: Several drones perform forest fire inspection tasks, determine the parameters of the airborne sensors, and use the area to be inspected as the task area for rasterization; ① Airborne sensor parameters include detection range S, detection probability and false alarm probability ② Rasterization of the task area, specifically: Assume that the task area is a rectangle and divide it evenly into L x ×L y Discrete grid G xy , the size of each grid is Δx×Δy; any grid is located by coordinates x and y, where x=1,2,...,L x ,y=1,2,...,L y ; A forest fire is defined as an area bounded by a spreading fire line. In the mission area, the set of grids whose centers are within the fire line envelope is taken as the target area to be searched. All grids in the target area to be searched are discrete point targets that are related to each other. According to the elliptical fire contour model, it is assumed that the starting point of the forest fire is (x f0 ,y f0 ), the forest fire spread time is t f ≥0, the forest fire line is expressed as: Among them, (x f ,y f ) is the position coordinate of any point f on the line of fire, θ f ∈[0,2π] is the angle corresponding to point f, that is, the angle of the length of the opposite arc in the ellipse; a f and b f They represent the average spreading speed of forest fire along the x-axis and y-axis in the inertial system, c W Represents the average speed of the ellipse center along the wind direction; φ W is the angle between the wind direction and the positive direction of the x-axis in the inertial system; the major and minor semi-axes of the ellipse are a f t f and b f t f ; Step 2: Use prior information to initialize the inspection information map, including the fire probability map and the environmental uncertainty map; ①Use prior information to obtain the initial fire probability map, and use a two-dimensional Gaussian function to describe the initial probability distribution of forest fire; The shape of the forest fire is considered as an ellipse, and the coordinates of the geometric center point of the ellipse are [x c ,y c ] T , then the initial fire existence probability map is expressed as: Among them, [x c ,y c ] T The probability of fire existing at the peak value p c ; σ x and σ y is the variance, which determines the range of the Gaussian distribution. The semi-axes of the ellipse parallel to the x-axis and y-axis in the inertial system are expressed as and ② The environmental uncertainty map is represented based on the information entropy of the probability of fire existence; The environmental uncertainty graph ξ(x,y,k)∈[0,1] is represented by the information entropy of the corresponding fire existence probability: when or When ξ(x,y,k)=0; Step 3: Initially, multiple drones perform patrol tasks according to the initial patrol information map and obtain comprehensive information of the current patrol area; The comprehensive information specifically includes: the position and ground speed vector of each drone, forest fire inspection benefits, and the position, shape and size of each environmental obstacle, which are described as follows: ① The position and ground speed vector of the drone Decouple the three-dimensional motion of the drone into horizontal motion and vertical motion. u The state equation of a UAV at time k is expressed as: Among them, p i (k) = [x i (k),y i (k),z i (k)] T represents the position of the drone, V h,i is the horizontal velocity amplitude of the i-th UAV; i (k) is the track deviation angle; ω i (k) = Δχ i (k) is the turning angle rate; V z,i and a z,i are the speed and acceleration of the UAV in the vertical direction respectively; ΔT is the planning step length; select ω i and a z,i As control input for drones; ② Forest fire inspection income Taking maximizing forest fire patrol benefits as the optimization goal, a multi-UAV collaborative forest fire patrol planning model is established as follows: p i (k)∈F, i=1,2,...,N u (8) In formula (5), p * N u The optimal patrol route of the UAV during the entire patrol mission execution process. N u The optimal patrol route of a UAV at time k, is the optimal planning waypoint of the i-th UAV at time k, p is the solution space, is the total number of planned steps for the inspection task, where T f is the total time of the patrol task; J is the patrol benefit function, including the fire detection benefit J f , environmental exploration benefit J e , potential detection benefit J p and synergy benefits J c ; d ,λ e ,λ c is the weight of the corresponding benefit; Formula (6) represents the collision avoidance constraint between aircraft, where d min is the minimum safe distance of the UAV; Formula (7) represents the terrain elevation constraint, z i (k) and z(x i ,y i ) are respectively the flight altitude of UAV i at time k and the terrain height at the grid where it is located; Formula (8) represents the kinematic constraint and terrain constraint, where F is the set of feasible routes considering the corresponding constraints; ③Environmental obstacles Environmental obstacles are all enveloping with standard convex polyhedrons. After enveloping, the obstacles are uniformly expressed in the inertial system as follows: Among them, a, b, c and d, e, f are all constants, which determine the size and shape of the obstacle respectively; p o =[x o ,y o ,z o ] T represents the coordinates of the obstacle center in the inertial system; Γ(x,y,z)=1 represents the obstacle surface; Each drone is given a different priority. Low-priority drones regard high-priority drones as spherical motion obstacles, whose radius is the minimum safety distance of the drone. The drones are sorted from low to high in order of priority and are called Step 4: For the current sampling time, determine whether the total time of the forest fire inspection task has reached T f , if yes, the algorithm ends; otherwise, go to step five; Step 5: Estimate the posterior probability distribution of potential forest fire source items based on the detector information of multiple UAVs during the inspection process; The detection results of multiple UAVs at time k are recorded as d k (R(k)), the historical detection results of multiple drones before time k are d 1:k-1 ={d1(R(1)),d2(R(2)),...,d k-1 (R(k-1))}. Estimation of forest fire source term at time k From the posterior probability density function Description, in the Bayesian reasoning framework, It is expressed as: in, is the normalization coefficient; is the prior probability density function, that is, the source term estimation result at time k-1; is the likelihood, that is, the forest fire source term is The detection result is d k probability; Traverse the rasterized task area and combine it with particle filtering to calculate Specifically: In particle filtering, the posterior probability distribution It is represented by a set of weighted sampling points, i.e. particles; assuming the number of particles is N, the jth particle at time k is recorded as Represents possible forest fire source items, and its weight express is the probability of the real forest fire source term, that is The posterior probability density function of the forest fire source term is approximately expressed by particle filtering as: Among them, δ(·) is the unit pulse function; At time k, the particle and its weight The update rules are as follows: first, Then, combining the likelihood The unnormalized weight of the corresponding particle It is expressed as: After normalization, the weight is: Get the effective number of particles When N eff When it is less than the threshold, the particles are resampled using the polynomial method; when N eff When it is greater than the threshold, the particles are weighted summed to obtain the forest fire source estimate The formula is: Step 6: Use Gaussian mixture model GMM to cluster particles containing the posterior probability distribution characteristics of forest fire source items, and calculate potential forest fire source items and their credibility; GMM is expressed as the weighted average of L Gaussian distributions, where the mean, covariance and mixing weight corresponding to the lth Gaussian distribution at time k are μ k,l ∈Μ k ={μ k,1 ,μ k,2 ,...,μ k,L },σ k,l ∈∑={σ k,1 ,σ k,2 ,...,σ k,L } and π k,l ∈Π={π k,1 ,π k,2 ,...,π k,L },and The particles Treat as a dataset The data points in , that is, the d-dimensional data vector to be clustered, then its probability is expressed by GMM as: in, for The mean and covariance are μ k,l and σ k,l The probability of a Gaussian distribution is: For each Gaussian model parameter, extract the mean μ k,l As the center, π k,l is the mixed weight, corresponding to N within three standard deviations G particle clusters, and the center μ of each particle cluster k,l As a potential source of forest fires; Assess the credibility of potential forest fire sources: Let Σ k for The covariance matrix of Then Σ k It reflects the convergence degree of particle filtering. T0=trace(Σ0) and T k =trace(Σ k ) represent the uncertainty of the estimated forest fire source term at the initial time of the mission and at time k, respectively. Then the lth potential forest fire source term μ at time k is k,l The credibility of is defined as: Step 7: Dynamically update the inspection information map based on potential forest fire sources; The potential forest fire source item μ l Substitute it into the forest fire spread model to obtain the potential forest fire contour, and then determine the grid G xy Is it within the corresponding potential fire line envelope? Represents G xy There is fire within the line of fire. Represents G xy There is no fire outside the fire line; the grid G xy The probability of fire existing at time k is Combined with the credibility of forest fire source term κ k,l , assuming that other potential forest fire sources are not considered, μ l In G xy The probability increment generated at That is, in G xy When there is no fire, the potential forest fire source μ l The probability of fire spreading is Considering all potential forest fire sources, in G xy The joint probability increment at is: Therefore, the dynamic update mechanism of the fire probability map is expressed as: in, is the threshold for dynamic probability update; when When the fire probability of this grid is detected, the probability is not dynamically updated, and the probability is only updated when the drone detects this grid; The dynamic update mechanism of the environmental uncertainty graph is expressed as: Step 8: Calculate the multi-UAV horizontal route with the highest collaborative inspection benefit based on the rolling horizon optimization and pruning strategy; Based on the rolling horizon optimization framework, the horizontal motion problem of the UAV is transformed into a series of small-scale optimization problems in a short time domain, namely: st(6),(7)and(8) Among them, N r is the rolling time domain step; The steps of rolling horizon optimization are as follows: Assuming the current planning time is k, the greedy algorithm is used to solve the future N of each drone r The optimal patrol route within the step length, and each aircraft only executes the first step, that is, As the planned horizontal speed at the current moment; when the next planned moment k+1 is reached, repeat the above steps until the inspection process is completed; The pruning method is used to solve the multi-machine collaborative planning problem. The steps are as follows: (1) For each UAV, calculate the patrol benefits of all feasible horizontal routes in the rolling time domain. This benefit only includes the fire detection benefit J in formula (22): f 、Environmental exploration benefitsJ e and potential detection benefit J p , excluding synergy benefits J c ; Then, select the n horizontal routes with the greatest benefit for each drone; (2) Arrange and combine the selected UAV routes to obtain (N u ) n The routes of multiple UAVs are grouped, and the complete collaborative inspection benefit of each route group is calculated according to formula (22); then, the route combination with the largest benefit is selected as the optimal horizontal planning route of multiple UAVs; Step 9: Introduce the IFDS algorithm to plan and obtain feasible three-dimensional waypoints that meet the requirements of terrain following and aircraft collision avoidance at the next moment; The process of calculating three-dimensional waypoints using the IFDS algorithm is: First, in order to make the drone return to the preset flight altitude after passing through the undulating terrain area, the following altitude holding law is designed: Among them, η z >0 and η z >0 is a constant, and z i (k) are the expected altitude and current altitude of the i-th UAV respectively; Then, the drone regards obstacles such as mountains and hills and other drones with high priority as obstacles, and changes the planned horizontal movement speed to The speed maintained with altitude is used as the confluence velocity u in the IFDS algorithm i (k), namely: Finally, the disturbed flow velocity is generated based on the IFDS algorithm To achieve terrain following and collision avoidance, there are: in, and are the disturbance matrix in the IFDS algorithm, the reference speed of obstacles, and the speed of the jth obstacle. For terrain obstacles, since the UAV needs to detect them, the UAV should climb when the current altitude is lower than the terrain elevation, that is, the directional coefficient in the IFDS algorithm is taken as θ i =π / 2; Finally, Perform kinematic constraints to obtain the actual planned speed v of the drone i (k); The expression of the next desired waypoint of the i-th UAV is: p i (k+1)=v i (k)ΔT (26) Step 10: Update the inspection information map according to the detection results of the drone during the flight, return to step 3, and plan the multi-drone forest fire inspection route at the next sampling time; During the forest fire patrol process, the drone continuously updates the probability of fire in the detected area based on the sensor detection results, combined with the sensor detection probability and false alarm probability According to the Bayesian criterion, the updating process is expressed as: Among them, d(x,y,k)∈{0,1} is the sensor’s observation of the grid G ​​at time k. xy The detection result of G xy If the value is not higher than 1, it indicates that a fire is detected; otherwise, it indicates that no fire is detected; The update formula of the environmental uncertainty graph ξ(x,y,k) is:

2. The method for efficient multi-UAV inspection in a low-altitude environment based on a terrain elevation map according to claim 1 is characterized in that: The relationship between the detection range of the airborne sensor and the grid is: S = Δx × Δy.

3. The method for efficient multi-UAV inspection in low-altitude environment based on terrain elevation map according to claim 1 is characterized in that: In the position and ground speed vector of the UAV, the horizontal and vertical motions of the UAV are: The horizontal motion space of the UAV is discretized. The UAV has 8 horizontal motion directions at time k, χ i (k)∈{1,2,3,4,5,6,7,8}, assuming that the feasible heading angle deflection range of the UAV within a planning step is ±45°, then χ i (k+1)∈{χ i (k)-1,χ i (k),χ i (k)+1}mod 8, where mod is the modulus operator. When the horizontal motion direction of the drone is parallel to the grid diagonal, the direction is called oblique; otherwise, it is called positive. Assume that the average velocity amplitudes of oblique motion and positive motion are V s and V f , The flight altitude of the drone is higher than the terrain altitude corresponding to the current grid. At the same time, the vertical motion space of the drone is continuous, satisfying the constraint V z,i ∈[V zmin ,V zmax ] and a z,i ∈[a zmin ,a zmax ].

4. The method for efficiently inspecting multiple UAVs in a low-altitude environment based on a terrain elevation map according to claim 1 is characterized in that: In the inspection revenue function, the specific definitions of each sub-revenue function are as follows: (1) Fire detection benefit J f It represents the weighted sum of the probability of fire existing in the area detected by the drone during the inspection process. The expression is: Among them, R i (k) is the detection range of the i-th UAV at time k, i.e., the UAV waypoint φ i (k) The grid directly below; ζ(x,y,k) is G xy Fire condition judgment sign at time k: Among them, ζ(x, y, k) = 1 means that it has been determined that in G xy There is a fire at this location, otherwise it is considered that there is no fire at this location. p is the probability threshold of fire existence; when it is determined that there is a fire, G xy The fire detection benefit at is 0; (2) Environmental exploration benefits J e It represents the environmental uncertainty of the area detected by the drone, and is used to guide the drone to detect areas with higher uncertainty. Its expression is: Where ξ(x,y,k) is G xy The uncertainty at time k; (3) Potential Exploration Benefits J p It represents the uncertainty of the potential detection area after the UAV executes the current planned route, and guides the UAV to approach the area with greater potential detection value. The expression is: Among them, A i (k) is the potential detection area of ​​the i-th UAV at time k, is the number of grids in the area; after the UAV executes the planned route at time k, its distance relative to the grid at the center of each sub-area will change. When the relative distance decreases, it means that the UAV is approaching the corresponding sub-area; (4) Assumptions is the detection range of the drone group at time k. By merging the repeated elements in this set, we can get its subset R s (k), which represents the non-overlapping detection range of the drone group during the inspection process; let the non-overlapping potential detection range of the drone group be A(k), then the collaborative benefit J c It is expressed as:

5. The method for efficient multi-UAV inspection in low-altitude environment based on terrain elevation map according to claim 1 is characterized in that: The likelihood It is expressed as: Among them, f(x,y,k)∈{0,1} means that when the forest fire source term is Is there a fire at (x,y)? Substituting into the forest fire spread model, we can obtain the corresponding forest fire outline, i.e., the fire line; when (x, y) is outside the fire line, f(x, y, k) = 0; otherwise, f(x, y, k) = 1.

6. The method for high-efficiency multi-UAV inspection in low-altitude environment based on terrain elevation map according to claim 1 is characterized in that: The normalization coefficient The expression is:

7. The method for efficient multi-UAV inspection in low-altitude environment based on terrain elevation map according to claim 1 is characterized in that: The GMM method is used for particle clustering decomposition: Firstly, the K-means clustering algorithm is used to preliminarily divide the particles and obtain the initial values ​​of the mean, covariance and mixing weight of each Gaussian model; Then, the maximum expectation method is used for iteration until the deviation of the two GMM decomposition results is less than the given tolerance. Finally, the relationship between the weight of each Gaussian model and the weight threshold is determined. When the Gaussian model weight π in GMM k,l Less than the weight threshold π t When , the model is eliminated; otherwise, the model is retained, and the number of Gaussian models obtained is N G ≤L.

8. The method for efficient multi-UAV inspection in low-altitude environment based on terrain elevation map according to claim 1 is characterized in that: As a random event, d(x,y,k) obeys the following probability distribution: in, Represents the probability of a random event.