A common sense integrated communication method for collaborative work of smart agricultural equipment

Abstract

The present application belongs to the field of intelligent agriculture technology, and in particular to a sensing and communication integrated method for collaborative operation of intelligent agricultural equipment. The method comprises the following steps: step one: alternately training by re-GDQN method and IDRL method to jointly obtain optimal unmanned aerial vehicle hovering point, sensing beam width and unmanned aerial vehicle trajectory; step two: beam excitation current weight optimization based on KD-DDPG, adopting a full connection network as the network model of actor and critic network, and inputting the target direction of the surface array specification, sensing beam and communication beam equipped by the unmanned aerial vehicle at different time points into the model. The present application simultaneously considers reducing energy consumption, improving communication and sensing beam performance, and introduces coverage limitation to construct a multi-objective optimization problem, so as to realize minimization of unmanned aerial vehicle energy consumption and maximization of sensing and communication beam signal-to-noise ratio under the premise of ensuring that the sensing beam completely covers the specified farmland area.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of smart agriculture technology, in particular to a sensing and communication integrated communication method for collaborative operation of smart agricultural equipment. BACKGROUND

[0002] With the development and application of communication technology, data in farmland scenarios is gradually diversified, and real-time detection of farmland work and environmental information is currently in great demand. Unmanned aerial vehicles (UAVs) as aerial base stations benefit from their flexible deployment characteristics and have received much attention in recent years. Specifically, UAVs can monitor data such as farmland soil moisture, crop status, and agricultural machinery location by sensing sensor information in the farmland, thereby promoting the development of smart agriculture. In the application scenario, integrated sensing and communication (ISAC) technology allows sensing and communication to be performed simultaneously. Therefore, current research combines beamforming with ISAC technology to achieve collaborative operation between devices.

[0003] For the application and performance optimization of integrated sensing and communication based on UAV base stations in various scenarios, domestic and foreign research focuses on efficient use of limited resources of UAVs and optimization of sensing and communication performance, aiming to reduce interference between sensing beams and communication beams. However, the codebook design for realizing regional sensing by scanning the target region of the sensing beam is mostly based on hierarchical codebook design, and the relationship between the effective coverage range of the constructed beam and the movement of the user position and the relationship between the interference of the communication sensing beam are not considered. SUMMARY

[0004] This section aims to summarize some aspects of the embodiments of the present application and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract and title of the specification to avoid obscuring the purpose of this section, abstract and title, and such simplifications or omissions cannot be used to limit the scope of the present application.

[0005] To solve the above technical problems, according to one aspect of the present application, the present application provides the following technical solutions:

[0006] A sensing and communication integrated communication method for collaborative operation of smart agricultural equipment, comprising the following steps: step one: alternately training by re-GDQN method and IDRL method to jointly obtain optimal UAV hovering point, sensing beam width and UAV trajectory;

[0007] Step 1.1: propose an IDRL model, which combines a diffusion model with reinforcement learning and uses Q values to guide the diffusion model update to generate an optimal solution;

[0008] Step 1.2: For the selection of beam width, based on the characteristics of the hierarchical codebook, the optional beam width is expressed as a finite set of 2 n times the minimum coverage width corresponding to the coverage area radius Ψ , the re-GDQN model is proposed to obtain the optimal UAV hovering position and the corresponding beam width, and the GNN graph neural network is proposed as the Q network of reinforcement learning, wherein the message passing layer of the GNN is first reduced in dimension by the CNN convolution layer and then the attention mechanism is added to obtain the characteristics, so as to ensure the light weight of the model and improve the ability of the feature to grasp the characteristics, in addition, the center point position corresponding to the grid and the information whether it is covered are input into the re-GDQN model, and the IDRL is called to obtain the optimal access order of the generated center point, and the current optimal trajectory is obtained, the reward is obtained according to the trajectory length and the coverage efficiency, the Q network is updated, and the optimal UAV hovering position is obtained.

[0009] Step 2: KD-DDPG-based beam excitation current weight optimization, a full connection network is used as the network model of the actor and critic network, the target direction of the planar array specification, the perception beam and the communication beam of the UAV at different time points are input into the model, and the excitation current weight of the perception beam and the communication beam is optimized, and a self-distillation module is added to the critic network to assist training.

[0010] As a preferred scheme of the integrated communication method for the collaborative operation of the intelligent agricultural equipment, in the step one, the method of alternating training is that, first, the IDRL is trained, so that it can obtain the optimal UAV trajectory for a group of hovering positions and corresponding communication and perception beam widths and target directions, so that the UAV can move in the shortest distance while avoiding the overlap of the perception beam and the communication beam; after p rounds, the IDRL model is fixed, and the re-GDQN method is trained, so that it can obtain reasonable UAV hovering points and corresponding perception beam widths for irregular farmland, so that the UAV moving energy consumption is minimized while the perception beam can scan the entire target area.

[0011] As a preferred scheme of the integrated communication method for the collaborative operation of the intelligent agricultural equipment, in the step 1.1, the proposed IDRL uses the reverse generation process of the diffusion model as the policy network of reinforcement learning, in each training process, the diffusion model network composed of U-Net generates the probability of selecting each path, and then the probability matrix is converted into a Hamiltonian loop by using the H operator, and then the corresponding Q value is calculated, the Q value is transmitted, and the diffusion model is updated, so as to obtain the optimal UAV access order, so that the UAV moving energy consumption is minimized and the target direction of the perception beam and the communication beam is avoided to overlap.

[0012] As a preferred scheme of the omniscience integration communication method for the collaborative operation of intelligent agricultural equipment, in the step 1.1, the environment where the intelligent agent is located contains the hovering position of the UAV X The perception beam coverage range corresponding to the hovering position of the UAV φt The target position aligned with the communication beam in each time slot And the corresponding communication beam coverage range And the selected Hamiltonian loop represented by the adjacency matrix A tsp_t , The time allocation of the moving track of the agricultural machine and the overall environment T And the time step parameter required by the diffusion model t Therefore, the environment state space is represented as: Wherein, A t The corresponding state in different models s t The selected action

[0013] The state space of the intelligent agent is set as a continuous space of edge probability distribution composed of floating-point numbers from 0 to 1, and the reward function is set as: Wherein R collection The overlap penalty

[0014] As a preferred scheme of the omniscience integration communication method for the collaborative operation of intelligent agricultural equipment, in the step 1.2, for the description of the environment of the intelligent agent, first, a quadtree is introduced to divide the target farmland area into uniform grids, and the length of the edge of the quadtree leaf node is equal to the length of the edge of the inscribed square of the beam coverage range corresponding to the smallest beam that can be constructed by the UAV. Based on this, the environment is constructed as a graph structure, which specifically contains two types of nodes: the center points of the grids divided by the quadtree and the center points corresponding to the areas covered by the perception beams, and three types of edges: grid points-grid points, coverage areas-grid points, and coverage areas-coverage areas. The graph structure indicates the topological relationship.

[0015] As a preferred scheme of the omniscience integration communication method for the collaborative operation of intelligent agricultural equipment, in the step 1.2, the state is represented as Wherein G is the grid topological structure, V is the node feature matrix containing the coordinate position of the node and the feature whether it is covered, E is the edge relationship represented by the adjacency matrix, and A tis an action mask matrix, which identifies the legal action space, wherein the action space is discrete, and is a set of grid centers with arbitrary coverage width that can be connected to the covered sensing area under the current state, and each action is represented as , node i is the selected grid center coordinate belonging to the optional action space, is the selected beam width corresponding to the position, and the re-GDQN returns a reward to the agent after completely covering the target area, that is, a sparse reward is set, wherein μ is the coverage efficiency of the current selected scheme, that is, the ratio of the covered area to the total area covered by the currently set sensing beam, and E(A t ) is the total energy consumption of the UAV calculated by the optimal access order obtained by the IDRL under the current scheme.

[0016] As a preferred scheme of the omnidirectional sensing and communication method for collaborative operation of intelligent agricultural equipment, in step two, the state is represented as wherein represents the depression angle and elevation angle of the target arrival angle corresponding to the sensing or communication beam, represents the depression angle and elevation angle of the target arrival angle of the corresponding communication / sensing beam under the same time slot, is the selected action, and the action space is represented as a 1×N vector composed of 0~1 floating point numbers, wherein N = N x × N y , , the reward function is constructed , the optimal excitation current weight distribution of the sensing beam and the communication beam under different time slots is obtained in sequence, and thus the UAV hovering position and the corresponding sensing beam width, the UAV access order, that is, the UAV trajectory and the excitation current weight of the beam are optimized.

[0017] Compared with the prior art, the present application has the beneficial effect that the present application simultaneously considers reducing energy consumption, improving communication and sensing beam performance, introduces coverage restriction to construct a multi-objective optimization problem, and decomposes it into three sub-problems, and respectively proposes corresponding algorithms for collaborative optimization. Specifically, the proposed re-GDQN model and IDRL model are trained alternately to jointly optimize the hovering position of the unmanned aerial vehicle, the sensing beam width corresponding to each position, and the access order of the unmanned aerial vehicle (i.e. the trajectory of the unmanned aerial vehicle); on this basis, further use the KD-DDPG algorithm to optimize the excitation current weight configuration of the sensing and communication beams in each time slot to improve the beam signal-to-noise ratio (SNR). In the method design, the IDRL algorithm uses a diffusion model as a policy network to generate actions, while the re-GDQN algorithm models the environment as a graph structure and introduces a convolution layer to construct a DQN network, both of which are lightweight models that are trained alternately to determine the optimal unmanned aerial vehicle trajectory and related beam parameters, converge faster and have shorter training time; subsequently, the KD-DDPG algorithm finely adjusts the excitation current weight, so that the gain of the communication beam and the sensing beam in each other's direction is close to zero, thereby effectively suppressing mutual interference. Finally, under the premise of ensuring that the sensing beam completely covers the specified farmland area, the minimization of the energy consumption of the unmanned aerial vehicle and the maximization of the signal-to-noise ratio of the communication and sensing beams are realized. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the present application will be described in detail below with reference to the drawings and detailed embodiments. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor. Among them:

[0019] Figure 1 A time allocation diagram for the perception and communication integration of the present application for a sensing and communication integrated communication method for intelligent agricultural equipment cooperative operation;

[0020] Figure 2 A scheme overall flowchart for the present application for a sensing and communication integrated communication method for intelligent agricultural equipment cooperative operation;

[0021] Figure 3 A step one model alternation training flowchart for the present application for a sensing and communication integrated communication method for intelligent agricultural equipment cooperative operation;

[0022] Figure 4 An IDRL training flowchart for the present application for a sensing and communication integrated communication method for intelligent agricultural equipment cooperative operation;

[0023] Figure 5A re-GDQN training flowchart of a common sense integration communication method for collaborative work of intelligent agricultural equipment according to the present application;

[0024] Figure 6 A KD-DDPG training flowchart of a common sense integration communication method for collaborative work of intelligent agricultural equipment according to the present application. DETAILED DESCRIPTION

[0025] In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0026] Secondly, the present application is described in detail in combination with the schematic diagram. In the detailed description of the embodiments of the present application, the cross-sectional view of the device structure will be partially enlarged without the general proportion for the convenience of description, and the schematic diagram is only an example, which should not limit the scope of protection of the present application here. In addition, the three-dimensional spatial dimensions of length, width and depth should be included in the actual manufacture.

[0027] In order to make the objectives, technical solutions and advantages of the present application more clear, the embodiments of the present application will be further described in detail below with reference to the accompanying drawings.

[0028] The present application proposes a common sense integration communication method for collaborative work of intelligent agricultural equipment. Specifically, first, any polygon is divided into uniform grid by using quadtree algorithm, then the improved re-GDQN and IDRL are alternately optimized, the hovering position of each flight cycle of the unmanned aerial vehicle, the corresponding perception beam width and the unmanned aerial vehicle access order, that is, the unmanned aerial vehicle trajectory are optimized, so that the flight of the unmanned aerial vehicle in a cycle can completely perceive the whole area of the target farmland through the scanning of the perception beam, and the mobile energy consumption of the unmanned aerial vehicle is minimized under the condition of avoiding the overlap of the target coverage range of the perception beam and the communication beam; then for each hovering point of the unmanned aerial vehicle, based on the zero-forcing beamforming, the beam weight of the perception beam and the communication beam is optimized by using the proposed KD-DDPG, further reducing the interference between the perception beam and the communication beam.

[0029] The method of the application comprises: the proposed re-GDQN model is alternately trained with the proposed IDRL model, optimal unmanned aerial vehicle hovering positions, unmanned aerial vehicle perception beam widths at corresponding positions and unmanned aerial vehicle access sequences, i.e., unmanned aerial vehicle trajectories, are obtained; subsequently, for each time slot, the KD-DDPG model is used to optimize the excitation current weight configuration of the perception and communication beams, and the beam signal-to-noise ratio is improved. The application models and decomposes the multi-objective NP difficult problem of reducing energy consumption and improving communication and perception beam performance into three small problems, and proposes the alternately trained IDRL algorithm and re-GDQN algorithm and the KD-DDPG algorithm to solve the problems. The IDRL algorithm and the re-GDQN algorithm jointly optimize the unmanned aerial vehicle trajectory and the width of the perception beam corresponding to the unmanned aerial vehicle hovering point, the IDRL algorithm generates actions by taking the diffusion model as a policy network, the re-GDQN algorithm models the environment as a graph structure and combines a convolution layer to form a DQN algorithm, the two lightweight models are alternately optimized to obtain the optimal unmanned aerial vehicle hovering position, the perception beam width of the unmanned aerial vehicle at the corresponding position and the unmanned aerial vehicle access sequence, i.e., the unmanned aerial vehicle trajectory, and then the KD-DDPG algorithm is used to optimize the excitation current weight of the beam, so that the gain of the communication beam and the perception beam in the direction of each other approaches 0, thereby reducing the interference between each other, and finally the perception beam can completely cover the designated farmland while minimizing the energy consumption of the unmanned aerial vehicle and maximizing the communication and perception beam SNR.

[0030] Specifically, in the scenario of the application, the unmanned aerial vehicle performs perception scanning by actively constructing a perception beam to perceive farmland environment information and the positions of sensors that need to establish a link. The application regards the target farmland as a polygon, and the unmanned aerial vehicle constructs a perception beam every time it hovers directly above the perception codebook, while the unmanned aerial vehicle constructs a communication beam to communicate with the moving agricultural machinery, i.e., the construction of the perception beam is synchronized with the construction of the communication beam every time, and the unmanned aerial vehicle hovers directly above the target direction every time the beam is perceived, and then moves to the next target perception direction in the next time slot, and the specific time allocation is as shown in the table. Figure 1 Based on this, the application proposes a scheme of a perception codebook, trajectory optimization and beam weight dynamic adjustment, and the flow is as shown in the figure. Figure 2

[0031] ​The 3db width of the perception beam, the trajectory of the unmanned aerial vehicle and the beam excitation current weight are jointly optimized, the maximum communication rate is realized while ensuring the optimal perception performance, and the energy consumption of the unmanned aerial vehicle is minimized. In the scene, the unmanned aerial vehicle sends an ISAC signal through the equipped Nx×Ny planar array and realizes the perception of the farmland by receiving the reflected ISAC signal from the perception target, and the unmanned aerial vehicle communicates with the agricultural machine through the equipped Nx×Ny planar array. The scene of the application is a farmland scene, so the channel is modeled as a LoS ideal channel, and the channel models of the perception and communication beams are independent of each other. Specifically, for any selected communication beam with a transmit beamforming weight of wcom, the signal received at the agricultural machine can be represented as:

[0032] (1)

[0033] wherein P represents the transmit power, , represents a channel coefficient vector , N = Nx x Ny , x is a normalized power bearing information signal, i.e. E[ x [ n ] 2 |]=1, w c is a communication beamforming vector. Wherein n c represents an additive white Gaussian noise satisfying , wherein is the target information transmitted by the unmanned aerial vehicle to the agricultural machine, is the information transmitted by the unmanned aerial vehicle to other non-target agricultural machines, which is mainly the reference information in the direction of the perception beam in the scene. Therefore, the signal-to-noise ratio received at the target agricultural machine is:

[0034] (2)

[0035] For the ISAC system, consider beam scanning for target sensing, wherein the unmanned aerial vehicle serves as an aerial base station to construct a directional, ground-perpendicular beam, and the unmanned aerial vehicle moves to the location of interest over time to realize the perception of the target area. The scanning scheme allows the unmanned aerial vehicle to sense in a wide area, and each time the beam is precisely directed to help obtain greater beamforming gain in the direction of interest while reducing the impact of clutter in other directions. The sensing process lasts for a specified time Δt . Therefore, for each duration , for t , the perception beam weight w t of the perception beam is selected, and the signal received at the target direction is:

[0036] (3)

[0037] where is the channel coefficient vector corresponding to the sensing beam, where θ and ψ are the elevation and depression angles of the angle of arrival (AoA) and angle of departure (AoD), assuming the same AoA and AoD since the target distance is usually much larger than the array aperture. x s is the reference signal of the normalized power, w s is the sensing beamforming vector, n s denotes the additive white Gaussian noise, satisfying . Therefore, the beam gain of the sensing target can be expressed as:

[0038] (4)

[0039] The corresponding echo signal-to-noise ratio (SNR) of the sensing target is:

[0040] (5)

[0041] The energy consumption of the UAV during hovering and moving can be modeled as:

[0042] (6)

[0043] where PH is the hovering energy consumption of the UAV, and PF is a moving energy consumption D i is the distance moved in the i th time slot, calculated as:

[0044] (7)

[0045] To reduce the energy consumption of the UAV while ensuring the sensing performance, the optimization objectives are as follows:

[0046] Objective 1: Improving communication quality requires maximizing the SNR between the base station and the user, which serves as the main optimization objective. Therefore, our first objective can be expressed as:

[0047] (8)

[0048] Objective 2: The echo rate is a commonly used indicator to evaluate the sensing performance in ISAC systems. Therefore, the second objective aims to maximize the echo rate of the target, which is expressed as:

[0049] (9)

[0050] Objective three: UAV is an energy-limited aerial platform, reducing its energy consumption can prolong the operation time of the considered system. Therefore, the third objective focuses on minimizing the energy consumption of the UAV, which is expressed as follows:

[0051] (10)

[0052] According to the above optimization variables and optimization objectives, the considered multi-objective optimization problem can be expressed as follows:

[0053] (11)

[0054] In the scheme, the target angles of the perception beam and the communication beam constructed by the UAV need to be located in different directions. If the other two beams are directed to the same area, the dual-beam will degenerate into a single-beam, which will cause the intelligent matching of perception and communication to one of them, resulting in the performance degradation of one of the functions. In addition, all areas of the target farmland need to be scanned by the perception beam, and the moving distance of the UAV in one time slot needs to meet the speed limit of the UAV. This is an NP-hard problem, and the three objectives are related to each other. To simplify the problem, the present application proposes a phased lightweight optimization framework based on reinforcement learning.

[0055] A method for integrated sensing and communication of intelligent agricultural equipment collaborative operation, comprising the following steps:

[0056] Step 1: joint optimization of perception beam width and UAV trajectory:

[0057] In the first phase, the proposed re-GDQN method and the IDRL method are alternately trained to jointly obtain the optimal UAV hovering point, perception beam width and UAV trajectory. Specifically, in the scheme, first train IDRL to obtain the optimal UAV trajectory for a set of hovering positions and corresponding communication, perception beam width and target direction, so that it can minimize the distance while avoiding overlap between the perception beam and the communication beam. In the p After the wheel, the IDRL model is fixed, and the re-GDQN method is trained to obtain reasonable UAV hovering points and corresponding perception beam widths for irregular farmland, so that the UAV moving energy consumption is minimized while the perception beam can scan the entire target area. The algorithm flow is shown in Figure 3 .

[0058] Step 1.1: UAV access sequence optimization based on IDRL:

[0059] In this part, the present application proposes an IDRL model to minimize the UAV trajectory while avoiding the overlap of the sensing beam and the communication beam. Wherein the IDRL combines the diffusion model with reinforcement learning, and the diffusion model is updated by the Q value to generate the optimal solution. The proposed IDRL uses the reverse generation process of the diffusion model as the policy network of reinforcement learning. In each training process, the diffusion model network composed of U-Net generates the probability of selecting each path, and then the probability matrix is converted into a Hamiltonian loop by using the H operator. Then the corresponding Q value under the "action" is calculated, the Q value is passed to guide the diffusion model to update, so as to obtain the optimal UAV access sequence to minimize the energy consumption of UAV movement and to avoid the overlap of the target direction of the sensing beam and the communication beam. The specific process is as shown in Figure 4

[0060] In the algorithm, the environment in which the agent is located contains the hovering position of the UAV X , the sensing beam coverage corresponding to the hovering position of the UAV φt , the target position of the communication beam in each time slot , the corresponding communication beam coverage , and the selected Hamiltonian loop represented by the adjacency matrix A tsp_t , The moving trajectory of the agricultural machine and the time allocation of the overall environment T , and the time step parameter required by the diffusion model t . Therefore, the environment state space can be represented as:

[0061] (12)

[0062] The state space of the agent is set to a continuous space of edge probability distribution composed of floating point numbers from 0 to 1.

[0063] According to the combination of formula 6 and formula 10, in this part, the present application sets the reward function as:

[0064] (13)

[0065] Wherein R collection is the overlap penalty, that is, the present application adds the overlap limit in formula 10 to the reward function in the form of a penalty function to guide the agent to make decisions.

[0066] Step 2.2: Optimization of the hovering position of the UAV and the width of the sensing beam based on re-GDQN:

[0067] ​In this part, for the selection of beam width, based on the characteristics of hierarchical codebook, the optional beam width is expressed as a finite set of 2 n times of the minimum coverage width corresponding to the coverage area radius Ψ . In order to simplify the calculation, in this part, the scanning problem of the perception beam is modeled as a square coverage problem, that is, the coverage range of the beam is simplified as the inscribed square of the corresponding coverage area.

[0068] Specifically, the present application proposes a re-GDQN model for obtaining the optimal unmanned aerial vehicle hovering position and the corresponding beam width. In the method, for the description of the agent environment, the present application first introduces a quadtree to divide the target farmland area into a uniform grid in this stage, and the length of the edge of the grid divided as the leaf node of the quadtree is equal to the length of the edge of the inscribed square of the beam coverage range corresponding to the minimum beam that can be constructed by the unmanned aerial vehicle. Based on this, the environment is constructed as a graph structure, which specifically includes two types of nodes: the center points of the grid divided by the quadtree, and the center points corresponding to the coverage area covered by the perception beam. The edges include three types: grid point-grid point, coverage area-grid point, and coverage area-coverage area. The graph structure indicates the topological relationship. Based on this, the present application proposes a GNN graph neural network as the Q network of reinforcement learning, wherein the message passing layer of the GNN is first reduced in dimension by a CNN convolution layer and then joins an attention mechanism to obtain features, so as to ensure the lightweight of the model while improving the ability of the agent to grasp features. In addition, the position of the center point corresponding to the grid and the information whether it is covered are input into the re-GDQN model, and the IDRL is called to obtain the optimal access order of the generated center point according to the optimal trajectory, and the reward is obtained according to the trajectory length and the coverage efficiency, so as to guide the Q network to update and obtain the optimal unmanned aerial vehicle hovering position. The specific process is shown in Figure 5 .

[0069] In the scheme, the state is represented as:

[0070] (14)

[0071] where G is the grid topology structure, V is the node feature matrix, which includes the coordinate position of the node and the feature whether it is covered, E is the edge relationship, which is represented by an adjacency matrix, and A is the action mask matrix, which identifies the legal action space. In the scheme, the action space is discrete, which is the set of grid centers that can be connected with the covered perception area at any coverage width under the current state, and each action is represented as:

[0072] (15)

[0073] node i For the selected center coordinates of the grid belonging to the optional action space, is the selected beam width of the corresponding position. In the scenario, the perception beam needs to cover the target area comprehensively, therefore, the re-GDQN returns the reward to the agent after each full coverage of the target area, i.e., sparse reward is set, which is specifically expressed in combination with formula 10 as follows:

[0074] (16)

[0075] wherein μ is the coverage efficiency of the current selected scheme, i.e., the ratio of the area covered to the total area covered by the perception beam currently set, and D is the total energy consumption of the UAV calculated by the optimal access order obtained by the IDRL under the current scheme. Based on this model, the re-GDQN is trained to obtain the optimal UAV hovering position and the corresponding perception beam coverage range, so that the perception beam constructed by the UAV minimizes the energy consumption of the UAV while comprehensively scanning the target area.

[0076] Step two: beam excitation current weight optimization based on KD-DDPG:

[0077] According to the hovering position of the UAV at different times obtained by the model trained in the first stage, combined with the motion trajectory of the agricultural machine, the arrival angle of the UAV relative to the agricultural machine at different time points can be calculated, so that the actual target direction of the communication beam can be obtained. In order to further improve the sensing performance, in this part, I optimize the excitation current weight in the beam weight, so that the gain of the perception beam and the communication beam in each other forms a null.

[0078] In the method, KD-DDPG uses a fully connected network as the network model of the actor and critic network, inputs the specifications of the planar array equipped by the UAV, the target direction of the perception beam and the communication beam at different time points into the model, optimizes the excitation current weight of the perception beam and the communication beam, and in the training process, in order to ensure the lightweight of the model, a self-distillation module is added to the critic network for auxiliary training. The process is as follows Figure 6 .

[0079] In this part, the state is represented as:

[0080] (17)

[0081] wherein represents the depression angle and the elevation angle of the target arrival angle corresponding to the perception or communication beam, represents the depression angle and the elevation angle of the target arrival angle of the communication / perception beam at the corresponding same time slot, is the selected action, and in the algorithm, the action space can be represented as a 1xN vector composed of 0~1 floating point numbers, wherein N = Nx × N y

[0082] (18)

[0083] Construct the reward function using Formula 10:

[0084] (19)

[0085] The scheme sequentially obtains the optimal excitation current weight allocation of the sensing beam and the communication beam under different time slots. At this point, the hovering position of the UAV and the corresponding sensing beam width, the UAV access sequence, i.e. the UAV trajectory and the excitation current weight of the beam are optimized, and the optimization problem corresponding to formula (10) is solved.

[0086] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A sensor-integrated communication method for collaborative operation of smart agricultural equipment, characterized in that, Includes the following steps: Step 1: The re-GDQN and IDRL methods are trained alternately to jointly obtain the optimal UAV hovering point, sensing beamwidth, and UAV trajectory. The alternating training method is as follows: First, the IDRL is trained to obtain the optimal UAV trajectory for a set of hovering positions and corresponding communication, sensing beamwidths, and target directions, so as to minimize the distance while avoiding overlap between the sensing beam and the communication beam. After p rounds, the IDRL model is fixed, and the re-GDQN method is trained to obtain a reasonable UAV hovering point and corresponding sensing beamwidth for irregular farmland, so that the UAV's movement energy consumption is minimized while the sensing beam can scan the entire target area. Step 1.1: Propose the IDRL model. IDRL combines the diffusion model with reinforcement learning. The Q-value guides the update of the diffusion model to generate the optimal solution. The proposed IDRL uses the reverse generation process of the diffusion model as the policy network for reinforcement learning. In each training process, the diffusion model network composed of U-Net generates the probability of selecting each path. Then, the H operator is used to convert the probability matrix into a Hamiltonian loop. Subsequently, the corresponding Q-value is calculated and passed to guide the update of the diffusion model, thereby obtaining the optimal UAV access order so that the UAV movement energy consumption is minimized and the target direction of the sensing beam and the communication beam is avoided from overlapping. Step 1.2: A re-GDQN model is proposed to obtain the optimal hovering position of the UAV and the corresponding beamwidth. A Q-network with a GNN graph neural network as reinforcement learning is proposed. The information of the center point position and whether it is covered by the grid is input into the re-GDQN model, and IDRL is called to obtain the current optimal trajectory according to the optimal access order of the generated center points. The reward is obtained according to the trajectory length and coverage efficiency to guide the Q-network update and obtain the optimal hovering position of the UAV. Step 2: Optimize beam excitation current weights based on KD-DDPG. A fully connected network is used as the network model for both the actor and critic networks. The array specifications of the UAV, the sensing beam and the communication beam, and the target directions at different time points are input into the model to optimize the excitation current weights of the sensing beam and the communication beam. A self-distillation module is added to the critic network to assist in training.

2. The sensor-integrated communication method for collaborative operation of smart agricultural equipment according to claim 1, characterized in that, In step 1.1, the environment in which the agent is located includes the hovering position specified by the UAV. X The coverage area of ​​the sensing beam corresponding to the drone's hovering position φt The target position aligned with the communication beam in each time slot and the corresponding communication beam coverage area and the selected Hamiltonian cycle represented by the adjacency matrix. A tsp_t , The movement trajectory of agricultural machinery and the time allocation of the overall environment T and the time step parameters required for the diffusion model t Therefore, the environment state space is represented as: ,in, A t For the corresponding states in different models s t The action of making a selection; The agent's state space is set as a continuous space consisting of edge probability distributions of floating-point numbers from 0 to 1, and the reward function is set as follows: ,in R collection This is an overlapping penalty.

3. The sensor-integrated communication method for collaborative operation of smart agricultural equipment according to claim 1, characterized in that, In step 1.2, the description of the intelligent agent's environment first involves introducing a quadtree to divide the target farmland area into uniform grids. The side length of each leaf node in the quadtree grid is equal to the side length of the inscribed square of the beam coverage area corresponding to the smallest beam that the UAV can construct. Based on this, the environment is constructed as a graph structure, specifically containing two types of nodes: the center point of the grid divided by the quadtree and the center point corresponding to the area covered by the sensing beam. The edges include three types: grid point-grid point, coverage area-grid point, and coverage area-coverage area. The graph structure shows the topological relationships.

4. The sensor-integrated communication method for collaborative operation of smart agricultural equipment according to claim 1, characterized in that, In step 1.2, the state is represented as follows: Where G is the mesh topology, V is the node feature matrix, containing the coordinates of the nodes and whether they are covered, E is the edge relationship, represented by the adjacency matrix, and A... t This is the action mask matrix, which identifies the legal action space. The action space is discrete and represents a set of grid centers of arbitrary coverage width that can be connected to the covered perception area in the current state. Each action is represented as... , node i The center coordinates of the selected grid belonging to the optional action space. This refers to the beamwidth selected for the corresponding location. The reward is returned to the agent only after the re-GDQN completely covers the target area, i.e., sparse reward is set. Where μ is the coverage efficiency of the currently selected scheme, that is, the ratio of the covered area to the total area covered by the currently set sensing beam, E(A t Then, the total energy consumption of the UAV is calculated by the optimal access order under the current scheme obtained by IDRL.

5. The sensor-integrated communication method for collaborative operation of smart agricultural equipment according to claim 1, characterized in that, In step two, the state is represented as ,in The depression and elevation angles represent the target's angle of arrival corresponding to the sensing or communication beam. This indicates the depression and elevation angles of the target arrival angle for the communication / sensing beam in the same time slot. This represents the selected action, and the action space is represented as a 1×N vector consisting of floating-point numbers from 0 to 1. N = N x × N y , Construct the reward function The optimal excitation current weight allocation for the sensing beam and communication beam under different time slots is obtained sequentially. At this point, the drone hovering position and the corresponding sensing beam width, the drone access sequence, i.e. the drone trajectory and the excitation current weight of the beam are optimized.

Images