Unmanned aerial vehicle data collection method in forest fire scene

Abstract

The invention relates to the technical field of unmanned aerial vehicles and Internet of Things, in particular to an unmanned aerial vehicle data collection method in a forest fire scene, which comprises the following steps: acquiring unmanned aerial vehicle data and node data, and establishing a region model and a forest fire model; establishing constraint conditions according to the unmanned aerial vehicle data, the node data, the region model and the forest fire model; further establishing an optimization problem according to the constraint condition; solving the optimization problem based on an unmanned aerial vehicle path planning algorithm of the dual-depth Q network, and taking the solved optimal parameter as the initial parameter of the dual-depth Q network to obtain a trained model; the model is deployed on the unmanned aerial vehicle, the model gives an action instruction of the unmanned aerial vehicle at the next moment according to real-time data captured by the unmanned aerial vehicle, and the unmanned aerial vehicle executes a data collection task according to the action instruction. The decision-making rationality is enhanced, the unmanned aerial vehicle is prevented from collecting invalid nodes, and the resource utilization rate is improved; by adopting the new communication scheduling method, the data transmission efficiency is improved, and the communication stability is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicles (UAVs) and the Internet of Things (IoT), and in particular to a method for collecting UAV data in a forest fire scenario. Background Art

[0002] Fires pose significant risks to forest resource protection. Firefighters face serious life and safety threats in every forest fire fighting operation. Timely global situational awareness at the fire scene can minimize this threat.

[0003] The rapid development of wireless communication technology and sensor performance has made the deployment and use of wireless sensors simpler and more reliable, which has promoted the progress of the Internet of Things and wireless sensor networks. By properly deploying wireless sensors, specific data can be quickly collected for specific areas in a convenient and efficient manner. It can collect data on demand and adapt to dynamically changing complex environments, while saving energy consumption for data transmission. This feature makes the drone-assisted Internet of Things and wireless sensor networks no longer restricted by network connections. The deployment of device nodes only needs to consider the needs of the application and the geographical structure of the target area.

[0004] Existing technologies such as CN2024119045509A disclose a multi-UAV efficient inspection method for low-altitude environments based on terrain elevation maps. This method takes into account the coordination of multiple UAVs, directly collects forest fire site information during each UAV inspection, and uniformly plans the initial paths of multiple UAVs by maintaining a global inspection information map. In the planning process of a single UAV, the environmental obstacle information is combined to obtain a safe path for a single UAV. This technology has the following problems:

[0005] 1. The efficiency of data and resource utilization is low. The solution in this paper is mainly based on the current drone detection data and limited prior information for analysis and decision-making. It does not fully explore and utilize a large amount of historical fire data and related data in different forest environments. This makes the system's adaptability and generalization capabilities relatively weak when facing new fire scenarios or complex situations, making it difficult to make optimal decisions quickly and accurately. For example, forest vegetation, topography, and climate conditions vary greatly in different regions, and this solution may require re-adjusting parameters and models for each new scenario, which is inefficient.

[0006] 2. The impact of wind speed on drones is not considered: When drones are performing data collection tasks, wind speed will limit their flight.

[0007] Therefore, there is an urgent need for a method to collect UAV data in forest fire scenarios. Summary of the invention

[0008] In view of this, the present invention discloses a method for collecting data from a drone in a forest fire scenario to solve the above problems; the method comprises:

[0009] S1, obtain drone data and node data;

[0010] S2. Establish a regional model based on the node data, use cellular automata to describe the forest fire spread process and establish a forest fire model;

[0011] S3. Establish UAV flight constraints based on UAV data; establish obstacle avoidance constraints based on regional models; establish communication constraints based on failed nodes, UAV data and node data;

[0012] S4, integrating UAV flight constraints, obstacle avoidance constraints, and communication constraints to establish an optimization problem;

[0013] S5. Solve the optimization problem based on the UAV path planning algorithm of the dual-depth Q network, obtain the optimal parameters of the dual-depth Q network, use the optimal parameters as the initial parameters of the path planning algorithm, and obtain a trained UAV path planning model;

[0014] S6. Deploy the trained model to the drone. The drone captures real-time data and inputs it into the trained model. The model gives the drone action instructions for the next moment, and the drone performs data collection tasks according to the action instructions.

[0015] The beneficial effects of the present invention include:

[0016] By obtaining the ambient wind speed collected by the ground node and adding the wind speed factor to the UAV flight constraints, the stability and adaptability of the UAV in actual mission execution are increased; by considering the uncertainty of the life cycle of the ground node caused by the fire, the rationality of the decision is enhanced, and the collection of invalid nodes by the UAV is avoided, thereby improving resource utilization; by adopting a new communication scheduling method, the data transmission efficiency is improved and the communication stability is guaranteed. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of the system model of the present invention;

[0018] Figure 2 It is a comparison diagram of the convergence process after smoothing under different algorithms of the present invention;

[0019] Figure 3 This is a comparison chart of the data collection rate of the UAV in each mission under different algorithms of the present invention;

[0020] Figure 4 This is a comparison chart of the success rates of safe path planning for drones under different algorithms of the present invention; DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solution, characteristics and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0022] like Figure 1 The figure shows a schematic diagram of a scene corresponding to this embodiment. The method for collecting drone data in a forest fire scene described in this embodiment includes the following steps:

[0023] S1. Obtain drone data and node data.

[0024] S2. Establish a regional model and a forest fire model based on the node data. The forest fire model is used to calculate the failed nodes.

[0025] Specifically, at the scene of a forest fire, wind is an important influencing factor. Wind force affects the spread of forest fires and also affects the air density around the rotor blades of drones, resulting in changes in motor power. w [n] is represented by in represents the data of the UAV wind speed sensor, n represents the nth time slot obtained by evenly dividing the maximum mission time, Δv w [n] represents the uncertainty of wind speed, which is determined by the difference between meteorological data and on-site wind speed. The larger the difference, the greater the uncertainty, the faster the wind changes, and the maximum uncertainty of wind speed. The larger the value, the more it is acquired by the ground node.

[0026] Furthermore, the forest fire model uses cellular automata to describe the forest fire spread process.

[0027] Specifically, the cellular automaton divides the task area into closely adjacent square grids, each of which is a cell. The properties of the cell are determined by factors such as the spatial terrain, vegetation, meteorological conditions, ignition time and combustion state inside the cell. The state of the cell can be divided into five stages: unburned, burned, completely burned, unextinguished and completely extinguished, represented by F0, F1, F2, F3 and F4 respectively.

[0028] The initial rate of fire spread R F0 The formula is:

[0029] R F0 =a*T F +b*V F +c*(100-H F )+d

[0030] Among them, T F is the temperature, V F is the wind force level, H Fis the relative humidity, a, b, c, and d are all empirical parameters, and their preferred values ​​are 0.03, 0.05, 0.01, and -0.3, respectively. Further considering the influence of fuel type, terrain slope, and wind vector on the initial spread rate, the corrected forest fire spread rate is R = R F0 *K f *K s *K w , where R is the corrected forest fire spread rate; K f is the correction factor for combustible materials; K s is the terrain slope correction coefficient, K s Describes the effect of different slopes on the spread rate of forest fire; K w is the wind vector correction coefficient, K w The effects of different wind force levels and wind direction conditions on the spread rate are described.

[0031] Furthermore, the coordinate index of each cell is defined as u and v, and the forest fire spread rate of the cell (u, v) is defined as The distance between a cell and its adjacent cells is the side length L of the cell. The state transition rule of the cell is: when the unburned cell is ignited, F0 changes to F1; after the internal combustion time After that, it is converted to F2, and F0 parallel to the fire line is ignited within Δn time, where R IN is the rate of forest fire spread inside the cell, L is the side length of the cell; F2 ignites all surrounding F0, and when there is no F0 around the cell in F2 state, it transforms to F3, which takes Δn time; F3 enters F4 state after Δn time. represents the area where the forest fire has spread, which is composed of cells that enter the F1 state and later. If the coordinate q of node k is k exist , then node k is a failed node.

[0032] S3. Establish UAV flight constraints based on UAV data; establish obstacle avoidance constraints based on regional models; establish communication constraints based on failed nodes, UAV data and node data.

[0033] Specifically, regarding the UAV flight constraints: the UAV spatial position is:

[0034] q[n]=(x[n],y[n],h u )

[0035] Among them, h u is the flight altitude of the UAV, and the projection coordinates of the UAV on the flight plane are:

[0036] q u [n]=(x[n],y[n])

[0037] Drone's heading Furthermore, there are K wireless sensor nodes in the system, where the position of the k∈{1,2,…,K}th node is q k =(x k ,y k ,z k ). Limited by the performance of the drone itself, there are constraints and Limited by the influence of wind speed in the regional model, there are constraints:

[0038] Specifically, regarding the obstacle avoidance constraints: the mountain terrain is expressed as:

[0039]

[0040] Among them, (x, y) represents the geographic coordinates, z is the height of the location, h i is the terrain correction parameter, controlling the overall height of the i-th mountain, x si With y si are the attenuation of mountain i along the x-axis and y-axis respectively, which control the slope of the mountain. During the flight of the drone, it is necessary to avoid known terrain obstacles and unknown environmental obstacles. The area of ​​obstacles in the drone's flight altitude plane is represented by a set D = {D1, D2, ... D j}, where D j represents the jth obstacle, D j That is, z(x,y)>h u is the set of points whose altitude is higher than the flight altitude of the UAV; the obstacle avoidance constraint of the UAV is

[0041] Specifically, regarding communication constraints: Communication constraints include scheduling constraints and data transmission constraints. In forest fire scenarios, the channel between the UAV and the node is mainly affected by vegetation. The present invention adopts the path loss model PL k The formula for [n] is:

[0042] PL k [n] = 20log 10 (f c )+10n1log 10 (θ k [n])+10n2log 10 (d k [n])+X σ

[0043]

[0044] Among them, f crepresents the carrier frequency, θ k [n] represents the elevation angle of the drone relative to the ground node; X σ Indicates shadow fading, which conforms to Gaussian distribution; d k [n] represents the distance between the drone and node k; the spatial position of the drone is (x[n], y[n], h u ); the position of the kth node is (x k ,y k ,z k ); n1 and n2 are the elevation angle influence factor and path loss index respectively.

[0045] Furthermore, the node transmission power is defined as p0 and the noise power of the drone receiving end is defined as p n , the signal-to-noise ratio γ at the drone receiver k [n] is:

[0046] γ k [n] = p0-PL k [n]-p n

[0047] Furthermore, the communication rate R between the drone and the node K [n] is:

[0048]

[0049] Wherein, B represents the channel bandwidth.

[0050] Furthermore, in order to support stable data stream transmission, the receiving signal-to-noise ratio γ at the drone should be k [n]≥γ0, the data collection task begins, where γ0 is a manually set signal-to-noise ratio threshold, preferably in the range of -5dB to -10dB, and preferably -5dB in this embodiment. The drone performs data collection tasks with at most one node at a time. represents the connection between the UAV and the kth node, and the scheduling constraint is expressed as:

[0051] Specifically, the UAV will prioritize data transmission tasks with nodes with good and stable communication channels. The formula for the scheduling strategy is:

[0052]

[0053] Among them, active nodes are defined as nodes with a signal-to-noise ratio γk[n] ≥ γ0, and otherwise refers to nodes with a signal-to-noise ratio lower than the threshold γ0 and failed nodes.

[0054] Further, represents the total amount of data collected from node k at time slot n. The superscript L is used to distinguish and Q K , the amount of data collected by the drone from node k is expressed as the data transmission constraint: The total amount of data that can be collected by the drone during the data collection process of all nodes is The maximum mission time is evenly discretized into N time slots, so that the position of the drone in each time slot can be approximately regarded as unchanged. The time slot length of the nth time slot is δ[n] = T max / N, in this embodiment, the number of time slots N is preferably 300.

[0055] S4. Integrate the UAV flight constraints, obstacle avoidance constraints, and communication constraints to establish an optimization problem.

[0056] Specifically, to maximize the total amount of data quickly collected from all nodes, the optimization problem is established as follows:

[0057]

[0058] subject to

[0059]

[0060] C7:T≤T max

[0061] Among them, C1 is the UAV flight speed performance constraint, C2 is the UAV steering performance constraint, C3 is the uncertainty constraint of environmental wind force, C4 is the obstacle avoidance constraint, C5 is the scheduling constraint, C6 is the data transmission constraint, and C7 represents the task time constraint.

[0062] S5. The UAV path planning algorithm based on the dual-depth Q network solves the optimization problem and obtains the optimal parameters of the dual-depth Q network. The optimal parameters are used as the initial parameters of the path planning algorithm to obtain a trained UAV path planning model.

[0063] Specifically, the sequential decision problem is expressed as a Markov decision process, that is, a formal expression using a tuple The specific meaning of the decision-making process is as follows:

[0064] Represents the state space, the environmental information vector S [n] =(S u [n],S o [n],S s [n],S t ), where S u [n] is the state information vector of the UAV, v[n] represents the flight speed of the drone, and the formula is v[n] = v u[n]+λv w [n],v u [n] is the ground speed of the UAV in a windless environment, λ is the discount factor for the influence of wind speed, preferably in the range of 0.005 to 0.01, and preferably 0.01 in this embodiment; S o [n] is the surrounding environment information vector detected by the UAV; S o [n]=(d1[n],d2[n],…,d8[n]), where d i [n], i∈{1,2,…,8} represents the distance between the UAV and the obstacle detected by the sensor in a certain direction. The sensor has a maximum detection distance limit; S s [n] is the global node information vector, S t The time consumed for task execution; further, s k [n] represents the information vector of a single node k.

[0065] The action of the agent is limited by the mechanical performance of the drone and can only take executable actions sampled from the legal action set. And define safety rules. When the drone's next action will fly towards an obstacle, cancel the action and give a penalty.

[0066] Represents the transition of state in the entire system.

[0067] represents the reward function set according to the optimization problem.

[0068] The dual-depth Q network reinforcement learning algorithm is used to solve the problem. When solving the problem, the settings are as follows:

[0069] Data collection rewards are Where α1 represents the data collection amount scaling factor, which is used to reduce the data collection amount to a suitable range. In this embodiment, α1 is preferably 10 -6 , It represents the total amount of data collected from node k in time slot n+1, where k represents the number of nodes.

[0070] The remaining task time reward is α2 represents a time amount scaling factor, which is used to reduce the time amount to a suitable range. In this embodiment, α2 is preferably 0.1.

[0071] The dynamic obstacle avoidance penalty is d=min(d1[n],d2[n],…,d8[n]), α3 represents the obstacle avoidance penalty coefficient, which is used to enlarge the shortest detection distance to an appropriate range. In this embodiment, α3 is preferably 10, and d represents the minimum value of the distance between the drone and the obstacle detected by the sensor in all directions. s Indicates the longest detection distance of the drone from the sensor.

[0072] There is a static obstacle collision penalty for known obstacles. α4 represents a collision penalty value, and in this embodiment, α4 is preferably 1.

[0073] In order to reduce invalid actions, the invalid action penalty is R ia =-α5, α5 represents the invalid action penalty value, and in this embodiment, α5 is preferably 1.

[0074] The global reward function is R = R dc +R tk +R uo +R so +R ia .

[0075] Furthermore, the initial parameters for solving the Markov decision process are set as follows:

[0076] Map size 300*300*50 Number of ground nodes 10 Maximum flight speed of drone 12m / s Maximum steering angle π Drone flight altitude 30m Maximum task time 300s Node transmit power 0.1W Signal-to-noise ratio threshold -5dB

[0077] Furthermore, the optimization problem is solved to obtain the optimal parameters of the dual-depth Q network. The solving steps include:

[0078] Step 1: Set the signal-to-noise ratio threshold γ0 and the maximum task time T max , the maximum flight speed of the drone v max and maximum steering angle

[0079] Step 2: Initialize the replay buffer D, estimate the network ξ, and initialize the target network ξ by copying the estimated network parameters - .

[0080] Step 3: By v max , Compute the legal action space.

[0081] Step 4: Initialize the experimental environment and set all environmental parameters to initial values.

[0082] Step 5: Get the environment information vector S[n], input S[n] into the network, get action a[n], the agent interacts with the environment and gets reward And the new environment information vector S[n+1].

[0083] Step 6: Tuple Added to the replay buffer D.

[0084] Step 7: Sample a certain number of tuples from D and calculate, and use the loss function to calculate the gradient to update the network parameters. Furthermore, the target Q value function in the dual-depth Q network is as follows:

[0085] S[n+1] is a non-final state

[0086] Among them, S[n+1] is the judgment condition of the final state: all The corresponding node signal disappears, or the task execution time T is greater than the maximum task time constraint T max .

[0087] Step 8: Every N r The update copies the parameters in ξ to ξ - In, N r is a threshold value of the number of training times of the neural network set manually. In this embodiment, N r Preferably 600.

[0088] Step 9: Determine whether S[n+1] is the final state; if so, jump to step 4; if not, jump to step 5.

[0089] Step 10: If the data collection rate and the path planning success rate reach the expected values, the training cycle is stopped to obtain the optimal neural network parameters, and the optimal parameters are set as the initial values ​​of each parameter in the neural network to obtain a trained model. The expected value of the data collection rate is preferably 98%, and the expected value of the path planning success rate is preferably 96%.

[0090] Furthermore, the optimal parameters are used as the initial parameters of the path planning algorithm to obtain a trained UAV path planning model.

[0091] S6. Deploy the trained model to the drone. The drone captures real-time data and inputs it into the trained model. The model gives the drone action instructions for the next moment, and the drone performs data collection tasks according to the action instructions.

[0092] Furthermore, if Figure 2 As shown, Figure 2 The convergence process of the algorithm proposed in the present invention and the comparison algorithm after smoothing is shown. The algorithm proposed in the present invention is the drone path planning algorithm based on dual deep Q network, and the comparison algorithm is the Dueling DQN algorithm. It can be seen that the reward value of the Dueling DQN algorithm continues to fluctuate violently after rising and the reward value does not reach the maximum; the performance of the algorithm proposed in the present invention is significantly improved, and the fluctuation is small during the convergence process, showing stability in a random environment.

[0093] like Figure 3 As shown, Figure 3 The data collection rate of the drone in each mission under different algorithms is demonstrated. Affected by environmental factors and the fire spreading process, the collection rate of the Dueling DQN algorithm has been hovering around 90% after the initial rise; the drone path planning algorithm based on the dual deep Q network adopted in the present invention reaches 100% after the initial rise process, and then stabilizes at a level close to 100%, demonstrating the effectiveness of the algorithm proposed in the present invention in data collection tasks.

[0094] like Figure 4 As shown, Figure 4 The success rates of safe path planning for drones under different algorithms are demonstrated. It can be seen that the success rates of the comparison algorithm and the algorithm adopted by the present invention can both reach 80% in the initial stage, while the path planning success rate of the Dueling DQN algorithm continues to decrease and fluctuates violently; the success rate of the algorithm adopted by the present invention converges to more than 95%, demonstrating strong adaptability to the complex environment of forest fires.

[0095] Finally, it should be noted that the above only describes some embodiments of the present invention. For those skilled in the art, it is conceivable that various changes, modifications, substitutions and deformations may be made to these embodiments without departing from the principles and spirit of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents, and the above-mentioned actions should be covered within the scope of protection of the present invention.

Claims

1. A method for collecting drone data in forest fire scenarios, characterized in that: include: S1, obtain drone data and node data; S2. Establish a regional model and a forest fire model based on the node data. The forest fire model is used to calculate the failure nodes; S3. Establish UAV flight constraints based on UAV data; establish obstacle avoidance constraints based on the regional model; Establish communication constraints based on failed nodes, drone data, and node data; S4, integrating UAV flight constraints, obstacle avoidance constraints, and communication constraints to establish an optimization problem; S5. Solve the optimization problem based on the UAV path planning algorithm of the dual-depth Q network, obtain the optimal parameters of the dual-depth Q network, use the optimal parameters as the initial parameters of the dual-depth Q network, and obtain a trained UAV path planning model; S6. Deploy the trained model to the drone. The drone captures real-time data and inputs it into the trained model. The model gives the drone action instructions for the next moment, and the drone performs data collection tasks according to the action instructions.

2. The method for collecting data from unmanned aerial vehicles in forest fire scenarios according to claim 1, characterized in that: The forest fire model uses cellular automata to describe the forest fire spread process.

3. The method for collecting data from unmanned aerial vehicles in forest fire scenarios according to claim 2, characterized in that: The cellular automaton includes states: unburned, burned, completely burned, unextinguished and completely extinguished, which are represented by F0, F1, F2, F3, F4 respectively. The transition rules between states are: When the unburned cell is ignited, F0 becomes F1; F1 internal combustion time After that, it is converted to F2, and F0 parallel to the fire line is ignited within Δn time, where R IN is the rate of forest fire spread inside the cell, and L represents the side length of the cell; F2 ignites all surrounding F0s. When there is no F0 around the cell in F2 state, it transforms to F3, which takes time Δ·; After Δn time, F3 enters F4 state.

4. The method for collecting data from unmanned aerial vehicles in forest fire scenarios according to claim 3, characterized in that: gather represents the area where the forest fire has spread, which is composed of cells that enter the F1 state and later. If the coordinate q of node k is k exist , then node k is a failed node.

5. The node k coordinates in the cells entering F1 and later states belong to the set Represents the area where the forest fire has spread. If the coordinates of node k Then node k is a failed node.

6. The method for collecting data from unmanned aerial vehicles in forest fire scenarios according to claim 1, characterized in that: The established optimization problems include: UAV flight speed performance constraints, UAV steering performance constraints, environmental wind uncertainty constraints, obstacle avoidance constraints, scheduling constraints, data transmission constraints, and mission time constraints.

7. The method for collecting data from unmanned aerial vehicles in forest fire scenarios according to claim 5, characterized in that: The formula of the communication scheduling strategy adopted by the scheduling constraint is: in, represents the connection between the UAV and the kth node, γ k [n] represents the signal-to-noise ratio of the kth node, and the active nodes are defined as the signal-to-noise ratio γ k [n]≥γ0, γ0 represents the signal-to-noise ratio threshold, otherwise represents nodes with a signal-to-noise ratio lower than the threshold and failed nodes.

8. The method for collecting data from unmanned aerial vehicles in forest fire scenarios according to claim 6, characterized in that: The value range of the signal-to-noise ratio threshold γ0 is: -5dB to -10dB.

9. The method for collecting data from unmanned aerial vehicles in forest fire scenarios according to claim 1, characterized in that: The discount factor λ for wind speed influence ranges from 0.005 to 0.01.

Images