A heuristic cluster predictive control method for multi-unmanned aerial vehicle system

Abstract

The application discloses a heuristic cluster prediction control method for a multi-unmanned aerial vehicle system, which utilizes an initial solution to screen a control space, and solves the calculation amount problem of the prediction control. The initial solution is determined by a correction gradient of the system at present, and indicates a rough direction for the next motion of the unmanned system. The method can greatly reduce the calculation amount and minimally affect the optimal control input of the system. A new obstacle avoidance mechanism based on a relative speed direction is proposed, which can ensure that the unmanned system safely and smoothly avoids the obstacles. The mechanism is applicable to a simple static obstacle scene, and can effectively deal with the obstacle avoidance problem of dense dynamic obstacles.

Description

Technical Field

[0001] This invention relates to the field of multi-UAV technology, and in particular to a heuristic swarm predictive control method for multi-UAV systems. Background Technology

[0002] The collective behavior of organisms in nature offers valuable insights for controlling aerial swarm systems. The self-organization, rapid response, and orderly structure exhibited by social animals are all targets for distributed swarm control. Many scholars have modeled biological swarm behavior and proposed numerous classic swarm control algorithms. The earliest was Reynolds' three swarm rules: cohesion, separation, and alignment. These rules describe the basic behavior of biological groups and offer some inspiration for designing swarm control methods for multiple unmanned systems. However, Reynolds' rules are too simplistic, and other classic models, such as the Vicsek model and the Cucker-Smale model, struggle to effectively handle swarm control problems in complex scenarios. In later research on swarm control methods, the concept of artificial potential fields has been widely used. Artificial potential fields correspond to gravitational fields in physics, similar to using gravitational potential to describe a gravitational field. Researchers describe the relationship between the unmanned system and its external environment by establishing potential functions, measuring certain characteristics of this relationship using the magnitude of the potential. The direction of gradient descent of the potential function can indicate the direction of motion of the unmanned system, because according to the principle of minimum potential energy, the system will spontaneously move in the direction of decreasing energy. Currently, gradient descent is used in almost all potential function scenarios, but it has certain limitations. The superposition and combination of multiple potential fields often produces local minima, which can trap the unmanned system, preventing it from reaching the optimal solution. Furthermore, artificial potential field methods cannot avoid unnatural oscillations, problems that swarm control algorithms must address. Model predictive control (MMC) can solve multi-constraint optimization problems of state variables and has great potential for handling swarm obstacle avoidance problems. This method incorporates the constraints of the unmanned system and each obstacle into the cost function optimization problem, achieving fast and stable control. However, MMC requires iteratively solving the optimization problem at each time step, which is computationally very resource-intensive. For low-cost unmanned systems forming a swarm, the computational load is enormous, making it difficult to meet the computational requirements. Additionally, due to the inherent inaccuracy of the model, long-term predictions often do not match reality.

[0003] Existing technology provides a multi-UAV formation cooperative control method based on model predictive control, comprising the following steps: Step 1: Initialize task requirements and relevant control parameters, etc.; Step 2: Perform preliminary trajectory planning for the lead UAV; Step 3: Detect the environmental conditions of the flight area in real time through sensors, determine and select a suitable flight formation, and calculate (update) the virtual formation guidance point; Step 4: Based on the calculated virtual formation guidance point, use the guidance point as a cost calculation reference value, calculate the cost function, and use a particle swarm optimization strategy based on distributed model predictive control for flight control; Step 5: Repeat steps 2, 3, and 4 to control the multi-UAV formation to fly cooperatively until the target position is reached. Summary of the Invention

[0004] To overcome the shortcomings of large computational load and inaccurate obstacle avoidance prediction in UAV swarms, this invention provides a heuristic swarm prediction control method for multi-UAV systems.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] This invention provides a heuristic swarm predictive control method for multi-UAV systems, the method comprising:

[0007] S1: Construct a drone swarm model, including several drones and several obstacles; establish a state equation for each drone, and set the motion mode and obstacle perception set of the virtual drone;

[0008] S2: Construct the original control space of the UAV, input it into the state equation, and obtain the original predicted state set of the UAV;

[0009] S3: Construct the potential function and cost function based on the obstacle perception set, the motion pattern of the virtual drone, and the original predicted state set;

[0010] S4: Determine the initial solution based on the spatial gradient of the potential function and the expected direction of the cost function, and filter the initial solution to obtain the preliminary control space;

[0011] S5: Transform the UAV swarm model into a Gibbs free field model, establish a joint probability density distribution problem, and set an update criterion based on the joint probability density distribution problem; input the initial control space into the state equation to obtain the initial predicted state set;

[0012] S6: Use the update criterion to initially predict the state set and perform iterative optimization to calculate the corresponding joint probability density value; when the joint probability density value is the maximum, use the corresponding initially screened control space as the optimal control space, and use the optimal control space to control the UAV swarm.

[0013] Preferably, in S1, a state equation is established for each of the UAVs, and the state equation is:

[0014]

[0015] Where, p i v represents the location of the drone. i For the speed of the drone, x i For the state of the drone, u i Here, m represents the control input for the drone, and m represents the mass of the drone. The state of the drone after inputting control input.

[0016] Preferably, in S2, the initial control space of the UAV is constructed, wherein the initial control space of the UAV is:

[0017]

[0018] Among them, u t Let e(k1Δθ, k2Δφ) be an arithmetic sequence, where k1, k2, k are the directions of the control input. θ k φ Δθ is an integer; Δφ is the angular interval of the horizontal control input direction; Δφ is the angular interval of the vertical control input direction. It is a set of integers.

[0019] Preferably, in S3, a potential function is constructed based on the obstacle perception set, the UAV's motion pattern, and the original predicted state set. The potential function includes the potential function between UAVs and the potential function between the UAV and the obstacle.

[0020] The formula for the potential function between the drones is:

[0021]

[0022] Where, k a k is the first parameter. t It is the second parameter; r f It is the expected distance between adjacent drones; ||p ij || Represents drone q i With drones q j The distance between them; ψ ar Let be the potential function between the drones;

[0023] The potential function between the UAV and the obstacle includes the repulsive potential function between the UAV and the obstacle and the deviation potential function between the UAV and the obstacle, and the formula is:

[0024] ψ o =ψ ob +ψ or

[0025] Where, ψ o Let ψ be the potential function between the drone and the obstacle; ob Let ψ be the repulsive potential function between the drone and the obstacle; or Let be the deviation potential function between the drone and the obstacle.

[0026] Preferably, the potential function between the UAV and the obstacle includes the repulsive potential function between the UAV and the obstacle and the deviation potential function between the UAV and the obstacle. The formula for the repulsive potential function between the UAV and the obstacle is:

[0027]

[0028] Where, k or The third parameter; ||p iβ || represents the distance between the drone and the obstacle;

[0029] The formula for the deviation potential function between the drone and the obstacle is:

[0030]

[0031] Where, k ob r is the fourth parameter. ρ The minimum distance between the obstacle and the drone's current velocity direction; u go The desired direction to bypass the obstacle; v i The current velocity is represented by ρ(λ, δ, z); ρ(λ, δ, z) is the transition function; k δ This is the fifth parameter.

[0032] Preferably, in S3, the cost function includes a cost function for the virtual drone's position and a cost function for the virtual drone's speed;

[0033] The formula for the cost function of the virtual drone's location is:

[0034] ψ rp =k rp (exp(||p i1 ||)-1)

[0035] Where, ψ rp The cost function for the virtual drone's location; k rp The sixth parameter; ||p i1 ||For drones q i Distance to the virtual drone;

[0036] The formula for the cost function of the virtual drone's speed is:

[0037] ψ rv =krv (exp(||v i1 ||)-1)

[0038] Where, ψ rv The cost function for the speed of the virtual drone; k rv The seventh parameter; ||v i1 ||For drones q i The relative speed with the virtual drone.

[0039] Preferably, the specific method of S6 is as follows:

[0040] S61: Initialization settings update criteria;

[0041] S62: Set up a preliminary operation set, store the random variables in the preliminary prediction state set into the preliminary operation set, and calculate the variables using the set update criteria to obtain the predicted marginal probability distribution value.

[0042] S63: Select a variable from the preliminary operation set and remove it from the preliminary operation set. Calculate the variable using the set update criteria to obtain the preliminary marginal probability distribution value. If the difference between the predicted marginal probability distribution value and the preliminary marginal probability distribution value is greater than a preset value, then move the variable into the prediction operation set. Otherwise, maintain the state of removing the variable from the prediction operation set until all variables in the prediction operation set are used up; obtain the optimized operation set.

[0043] S64: Calculate the joint probability density based on the optimized operation set; when the joint probability density value is the maximum, take the corresponding initially screened control space as the optimal control space, and use the optimal control space to control the UAV swarm.

[0044] Preferably, in S61, the update criterion is:

[0045]

[0046] Among them, Z i For the allocation function; ψ ar Let ψ be the potential function between the drones; o Let ψ be the potential function between the drone and the obstacle; rp The cost function for the virtual drone's location; ψ rv x is the cost function for the speed of the virtual drone; i and x j The parameters are for predicting the state set.

[0047] Preferably, in S64, the joint probability density is calculated based on the optimized operation set, and the formula for the joint probability distribution is:

[0048]

[0049] Where p(X) is the joint probability density.

[0050] The present invention also provides a heuristic swarm prediction control system for multi-UAV systems, for implementing the above-described method, the system comprising:

[0051] The model building module constructs a drone swarm model, including several drones and several obstacles; it establishes a state equation for each drone and sets the motion mode and obstacle perception set of the virtual drone;

[0052] The space construction module constructs the original control space of the UAV, inputs it into the state equation, and obtains the original predicted state set of the UAV.

[0053] The function construction module constructs the potential function and cost function based on the obstacle perception set, the motion pattern of the virtual drone, and the original predicted state set;

[0054] The spatial filtering module determines the initial solution based on the spatial gradient of the potential function and the expected direction of the cost function, and filters the initial solution to obtain the preliminary filtered control space.

[0055] The model prediction module transforms the UAV swarm model into a Gibbs free-field model, establishes a joint probability density distribution problem, and sets update criteria based on the joint probability density distribution problem; the initial control space is input into the state equation to obtain the initial predicted state set;

[0056] The control module is optimized by using the update criteria to initially predict the state set for iterative optimization and calculate the corresponding joint probability density value. When the joint probability density value is the maximum, the corresponding initially screened control space is taken as the optimal control space, and the UAV swarm is controlled by the optimal control space.

[0057] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0058] This invention proposes a heuristic swarm predictive control method for multi-UAV systems. It utilizes an initial solution to filter the control space, thus addressing the computational complexity issue of predictive control. This initial solution is determined by the system's current correction gradient, providing a coarse direction for the UAV system's next movement. This method significantly reduces computational cost while minimizing impact on the system's optimal control input. A novel obstacle avoidance mechanism based on relative velocity direction ensures the UAV system can safely and smoothly avoid obstacles. This mechanism is applicable not only to simple static obstacle scenarios but also effectively handles obstacle avoidance problems involving dense dynamic obstacles. Attached Figure Description

[0059] Figure 1This is a flowchart of the heuristic swarm predictive control method for multi-UAV systems described in Example 1;

[0060] Figure 2 This is a schematic diagram of the structure of the drone swarm and obstacles described in Example 2;

[0061] Figure 3 This is a schematic diagram of the control space structure described in Example 2;

[0062] Figure 4 This is a schematic diagram illustrating the operational mode of the potential function and transition function described in Example 2;

[0063] Figure 5 This is a schematic diagram of the obstacle avoidance mechanism described in Example 2;

[0064] Figure 6 This is a schematic diagram of the heuristic swarm predictive control system for multi-UAV systems described in Example 3. Detailed Implementation

[0065] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0066] To better illustrate this embodiment, some parts in the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions;

[0067] It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings.

[0068] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0069] Example 1

[0070] This embodiment provides a heuristic swarm predictive control method for multi-UAV systems, such as... Figure 1 As shown, the method includes:

[0071] S1: Construct a drone swarm model, including several drones and several obstacles; establish a state equation for each drone, and set the motion mode and obstacle perception set of the virtual drone;

[0072] S2: Construct the original control space of the UAV, input it into the state equation, and obtain the original predicted state set of the UAV;

[0073] S3: Construct the potential function and cost function based on the obstacle perception set, the motion pattern of the virtual drone, and the original predicted state set;

[0074] S4: Determine the initial solution based on the spatial gradient of the potential function and the expected direction of the cost function, and filter the initial solution to obtain the preliminary control space;

[0075] S5: Transform the UAV swarm model into a Gibbs free field model, establish a joint probability density distribution problem, and set an update criterion based on the joint probability density distribution problem; input the initial control space into the state equation to obtain the initial predicted state set;

[0076] S6: Use the update criterion to initially predict the state set and perform iterative optimization to calculate the corresponding joint probability density value; when the joint probability density value is the maximum, use the corresponding initially screened control space as the optimal control space, and use the optimal control space to control the UAV swarm.

[0077] Example 2

[0078] This embodiment provides a heuristic swarm predictive control method for multi-UAV systems, the method comprising:

[0079] S1: Construct a drone swarm model, including several drones and several obstacles; establish a state equation for each drone, and set the motion mode and obstacle perception set of the virtual drone;

[0080] S = {q1, q2, ..., q} N}, where N > 1 is the set of N individual UAV systems in the cluster system. Each UAV system q i For a node i in the topology graph of the corresponding cluster, all nodes are in the node set v = {1, 2, ..., N}.

[0081] For each of the aforementioned UAVs, a state equation is established, which is:

[0082]

[0083] Where, p i v represents the location of the drone. i For the speed of the drone, x i For the state of the drone, u i Here, m represents the control input for the drone, and m represents the mass of the drone. The state of the drone after inputting control input.

[0084] like Figure 2 As shown, the motion mode of the virtual drone is:

[0085]

[0086] Where m is the unmanned aerial vehicle system q iQuality; g represents the cluster There is a unique desired target point at any given time; For clusters The expected rate of motion.

[0087] Cluster There is a unique clustering algorithm at any given time. in, It is a cluster The set of all unmanned aerial vehicle (UAV) system states. It is a collection of control commands for drone systems other than virtual drones. It is the set of perceived positions of the edge points of all obstacles in the environment.

[0088]

[0089] Where, p ki It is an unmanned aerial vehicle system q i The edge point of the k-th obstacle is perceived, O i It is an unmanned aerial vehicle system q i The set of edge points of each perceived obstacle, the total number of obstacles is N. O .

[0090] S2: Construct the original control space of the UAV, input it into the state equation, and obtain the original predicted state set of the UAV;

[0091] Constructing the original control space for drones, such as Figure 3 As shown, the initial control space of the UAV is:

[0092]

[0093] Among them, u t Let e(k1Δθ, k2Δφ) be an arithmetic sequence, where k1, k2, k are the directions of the control input. θ k φ Δθ is an integer; Δθ is the angular interval of the horizontal control input direction; Δφ is the angular interval of the vertical control input direction; Z is a set of integers; the direction vector e(θ, φ) = [cosφcosθ, cosφsinθ, sinφ] T The parameter k represents the direction of the control input. θ k φ Satisfying k θ Δθ=k φ Δφ=2πn unThe selected value cannot be too large, otherwise it will cause excessive computational burden; it cannot be too small, otherwise the desired control input cannot be applied to the UAV system, and the desired cluster state cannot be achieved. In this algorithm, the discrete control space of all UAV systems remains consistent, and the selectable control input of each UAV system is not affected by the current system state.

[0094] To predict the system's state, the continuous state transition function is rewritten in discrete form based on the backward Euler method. x i It is a drone q i state, u i It is a drone q i The input is G and K, which are constant matrices, where k represents the sampling time and H represents the planning period.

[0095]

[0096] The planning period and sampling interval are not necessarily the same; the planning period may be longer than the sampling interval. In predictive control methods, a planning period that is appropriately longer than the sampling interval can improve the clustering performance of an aerial trunking system. The predicted state set of the system can be written as follows:

[0097]

[0098] Where, μ k It is the control space at time k, x k+H It is the predicted state set at time k+H. It is a drone q N The predicted state at time k+H is determined by the drone q. N The control space μ at time k N,k With drones q N The state x at time k N,k The above method only satisfies the control input constraints, not necessarily the velocity constraints. Therefore, states whose predicted velocities are greater than the maximum velocity need to be removed from the predicted state set.

[0099] S3: Construct the potential function and cost function based on the obstacle perception set, the motion pattern of the virtual drone, and the original predicted state set;

[0100] Based on the obstacle perception set, the UAV's motion pattern, and the original predicted state set, a potential function is constructed, such as... Figure 4 As shown, the potential function includes the potential function between drones and the potential function between drones and obstacles;

[0101] The formula for the potential function between the drones is:

[0102]

[0103] Where, k a k is the first parameter. t It is the second parameter; r f It is the expected distance between adjacent drones; ||p ij || Represents drone q i With drones q j The distance between them; ψ ar Let k be the potential function between the drones; t ∈(1, 2], and very close to 2, have little effect on attraction and repulsion. Since the range of neighbors r... s Limited, ||p ij ||>k t r f The situation where r is taken is rare. f =0.421, k a =0.2, k t =1.95.

[0104] The potential function between the UAV and the obstacle includes the repulsive potential function between the UAV and the obstacle and the deviation potential function between the UAV and the obstacle, and the formula is:

[0105] ψ o =ψ ob +ψ or

[0106] Where, ψ o Let ψ be the potential function between the drone and the obstacle; ob Let ψ be the repulsive potential function between the drone and the obstacle; or Let be the deviation potential function between the drone and the obstacle.

[0107] like Figure 5 As shown, the interaction between the UAV system and obstacles differs from its interaction with its neighbors; obstacles only have a repulsive effect on the UAV system. To safely and effectively avoid dense static and dynamic obstacle clusters, we divide the area around the UAV system into four zones, representing different collision risk levels. Zones I, II, and III indicate a risk of collision, while zone IV indicates a near-impossible collision, with zero risk. The axis of symmetry for zone I is the UAV system's q. i The line connecting the obstacle β has an apex angle θ. I =arcsin(2r coll / ||p iβ ||) Determined by the safety radius r coll And the distance between the unmanned system and obstacles ||p iβ|| Confirmed, indicating that the drone system will collide with the target area if it moves at its current speed. Areas II and III expand outwards sequentially, indicating that the drone system is at risk of collision if it moves at its current speed.

[0108] Region III's outer edge is determined by parameter k. δ Safety radius r coll and the distance between the drone system and the obstacle ||p iβ || Confirmed, θ III =arcsin((2+k) δ )r coll / ||p iβ ||). Distance between the drone system and obstacles ||p iβ || Angle θ between the unmanned aerial vehicle system and the line connecting the obstacle and the obstacle iβ The risk level of a collision between the two was determined. When θ iβ ≥θ III At that time, the unmanned aerial vehicle system q i It will not collide with obstacles. When θ iβ <θ III ,||p iβ ||Greater than the expected distance rf UAV sensing radius r s At that time, the unmanned aerial vehicle system q i It can detect obstacles and there is a varying degree of collision risk. Therefore, the unmanned aerial vehicle (UAV) system q i The action of avoiding the obstacle needs to be generated, using velocity v. ob It means that the regulations v ob Along the outer edge of Region III. Within the obstacle's influence range, the unmanned aerial vehicle system q i Along v ob Directional movement. This is beneficial for unmanned aerial vehicle systems. i Smoothly avoid obstacles without significant oscillations. When θ iβ <θ III ,||p iβ When the distance is less than the expected distance rf, there exists a repulsive potential function ψ. or >0 enables the unmanned aerial vehicle system q i Move away from the obstacle. At this point, the desired direction is to avoid it, v. ob It still works, but v ob Insufficient constraints could lead to collisions, therefore explicit constraints on the distances between unmanned aerial vehicle (UAV) systems are needed. or Let be the repulsive potential function between the drone and the obstacle.

[0109] The potential function between the UAV and the obstacle includes the repulsive potential function between the UAV and the obstacle and the deviation potential function between the UAV and the obstacle. The formula for the repulsive potential function between the UAV and the obstacle is:

[0110]

[0111] Where, k or The third parameter; ||p iβ || represents the distance between the drone and the obstacles; where r is taken as... f =0.421, k or =0.2. Potential function ψ or A gradient component is provided, which is related to v ob The superposition determined the q of the unmanned aerial vehicle system. i The desired direction u to bypass the obstacle go , ψ ob To represent the actual speed of the unmanned aerial vehicle system relative to u go Deviation between.

[0112] The formula for the deviation potential function between the drone and the obstacle is:

[0113]

[0114] Where, k ob r is the fourth parameter. ρ The minimum distance between the obstacle and the drone's current velocity direction; u go The desired direction to bypass the obstacle; v i The current velocity is represented by ρ(λ, δ, z); ρ(λ, δ, z) is the transition function; k δ The fifth parameter; r ρ =||p iβ ||sinθ iβ It is an obstacle-free unmanned aerial vehicle system q i The minimum distance in the current velocity direction, and θ iβ =arccos(v i ·u go / ||v i ||·||u go ||). λ=2r coll / ||p iβ ||.

[0115] ρ(λ, δ, z) is the transition function k used to classify risk levels. ρ Parameters where both δ and φ are positive, such as Figure 4 As shown, the transition function is:

[0116]

[0117] The cost function includes a cost function for the virtual drone's location and a cost function for the virtual drone's speed;

[0118] The formula for the cost function of the virtual drone's location is:

[0119] ψ rp =k rp (exp(||p i1 ||)-1)

[0120] Where, ψ rp The cost function for the virtual drone's location; k rp The sixth parameter; ||p i1 ||For drones q i Distance to the virtual drone;

[0121] The formula for the cost function of the virtual drone's speed is:

[0122] ψ rv =k rv (exp(||v i1 ||)-1)

[0123] Where, ψ rv The cost function for the speed of the virtual drone; k rv The seventh parameter; ||v i1 ||For drones q i The relative speed with the virtual drone.

[0124] S4: Determine the initial solution based on the spatial gradient of the potential function and the expected direction of the cost function, and filter the initial solution to obtain the preliminary control space;

[0125] A correct initial solution can effectively reduce computational overhead while minimizing the impact on the optimal control input of the UAV system. According to the principle of minimum potential energy, the aerial swarm system will move in the direction of decreasing energy. The energy function established by this algorithm includes both a potential function and a performance cost. For the potential function, the potential energy can be reduced along the negative gradient direction. For the cost function, the spatial gradient cannot be solved. Therefore, the direction of energy reduction in the system does not represent the direction of gradient descent of the potential function. However, appropriate correction to the negative gradient direction of the potential function can minimize the system energy. This correction includes the desired direction of the cost, such as v. ob In summary, the selection of the initial solution involves two parts: the spatial gradient of the potential function and the expected direction of the cost function. The superposition of these two parts roughly conforms to the principle of minimum potential energy. Using optimization methods, the optimal control input under this initial solution condition can be obtained. The spatial negative gradient can be obtained from the potential function.

[0126]

[0127]

[0128]

[0129] To reduce r f When the position of the unmanned aerial vehicle system changes, a damping force u is added. gav =-k av v ij u gav Pointing to the unmanned aerial vehicle system q i Compared to unmanned aerial vehicle systems q j The opposite direction of the movement.

[0130] To achieve better obstacle avoidance performance, the desired obstacle avoidance velocity u is added to the initial solution. gob =k3n ob Where k3 is a constant, n ob The velocity v of the obstacle ob The direction can be determined by the following formula.

[0131]

[0132] In summary, the total initial solution can be obtained by superimposing the two initial gradient solutions.

[0133] u g =u gar +u gav +u gor +u gr +u gob

[0134] The direction of the initial solution is used to filter the control input, but the magnitude of the initial solution is not used to filter the control input magnitude. This is because the initial solution u... g Contains multiple parameters to be determined, u g The modulus is not easy to determine, and it is easy to encounter ||u g ||>u max The situation is as follows. The control space after preliminary screening is:

[0135] μ * ={u|u=u t e(k1Δθ,k2Δφ),k1∈[0,k θ ), k2∈[0, k φ ),

[0136]

[0137] u t ={u max / n un ,2u max / n un , ..., u max},∠(u,u g ) = arccos(u·u g ||u||·||ug ||);

[0138] k u The semi-apex angle of the envelope cone after reducing the motion space.

[0139] S5: Transform the UAV swarm model into a Gibbs free field model, establish a joint probability density distribution problem, and set an update criterion based on the joint probability density distribution problem; input the initial control space into the state equation to obtain the initial predicted state set;

[0140] S6: Use the update criterion to initially predict the state set and perform iterative optimization to calculate the corresponding joint probability density value; when the joint probability density value is the maximum, use the corresponding initially screened control space as the optimal control space, and use the optimal control space to control the UAV swarm.

[0141] S61: Initialization settings update criteria;

[0142] Determining the probability distribution remains challenging due to the mutual influence between random variables. To address this issue, variational inference can be employed, aiming to find an approximate distribution q(X) that closely approximates p(X). The degree of similarity between these two distributions is typically expressed using the KL divergence. The goal of measuring the difference between these two distributions is to minimize the KL divergence. Assuming the random variables are independent, the q(X) distribution satisfies the form q(X) = ∏ i q(X i Then, we can use the idea of ​​coordinate ascent to minimize D(q|p) until convergence. We can obtain the minimum point that minimizes the KL divergence and take this minimum point as the update criterion for iteration.

[0143] The update criteria are as follows:

[0144]

[0145] Among them, Z i For the allocation function; ψ ar Let ψ be the potential function between the drones; o Let ψ be the potential function between the drone and the obstacle; rp The cost function for the virtual drone's location; ψ rv x is the cost function for the speed of the virtual drone; i and x j The parameters are for predicting the state set.

[0146] S62: Set up a preliminary operation set, store the random variables in the preliminary prediction state set into the preliminary operation set, and calculate the variables using the set update criteria to obtain the predicted marginal probability distribution value.

[0147] S63: Select a variable from the preliminary operation set and remove it from the preliminary operation set. Calculate the variable using the set update criteria to obtain the preliminary marginal probability distribution value. If the difference between the predicted marginal probability distribution value and the preliminary marginal probability distribution value is greater than a preset value, then move the variable into the prediction operation set. Otherwise, maintain the state of removing the variable from the prediction operation set until all variables in the prediction operation set are used up; obtain the optimized operation set.

[0148] S64: Calculate the joint probability density based on the optimized operation set; when the joint probability density value is the maximum, take the corresponding initially screened control space as the optimal control space, and use the optimal control space to control the UAV swarm.

[0149] Viewing the cluster as a Gibbs random field, the unmanned aerial vehicle (UAV) systems within the cluster correspond sequentially to the set of random variables X = {X1, X2, ..., X...}. N Elements in}. Random variable X i The domain of change is determined by the unmanned aerial vehicle system q i The predicted state set is determined. When X i The cluster exhibits different energy distributions when different values ​​are selected.

[0150] The joint probability density is calculated based on the optimized operation set, and the formula for the joint probability distribution is:

[0151]

[0152] Where p(X) is the joint probability density, and the assignment function is... To make the calculated control commands smoother, an inertial element is introduced at the end of the algorithm. Where α is a positive number close to 1, u * These are control commands calculated during the current observation period. It is the control command calculated from the previous observation period. It is the final control command.

[0153] Example 3

[0154] This embodiment also provides a heuristic swarm prediction control system for multi-UAV systems, used to implement the methods of Embodiments 1 and 2, such as... Figure 6 As shown, the system includes:

[0155] The model building module constructs a drone swarm model, including several drones and several obstacles; it establishes a state equation for each drone and sets the motion mode and obstacle perception set of the virtual drone;

[0156] The space construction module constructs the original control space of the UAV, inputs it into the state equation, and obtains the original predicted state set of the UAV.

[0157] The function construction module constructs the potential function and cost function based on the obstacle perception set, the motion pattern of the virtual drone, and the original predicted state set;

[0158] The spatial filtering module determines the initial solution based on the spatial gradient of the potential function and the expected direction of the cost function, and filters the initial solution to obtain the preliminary filtered control space.

[0159] The model prediction module transforms the UAV swarm model into a Gibbs free-field model, establishes a joint probability density distribution problem, and sets update criteria based on the joint probability density distribution problem; the initial control space is input into the state equation to obtain the initial predicted state set;

[0160] The control module is optimized by using the update criteria to initially predict the state set for iterative optimization and calculate the corresponding joint probability density value. When the joint probability density value is the maximum, the corresponding initially screened control space is taken as the optimal control space, and the UAV swarm is controlled by the optimal control space.

[0161] The same or similar labels correspond to the same or similar parts;

[0162] The terms used to describe positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent.

[0163] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A heuristic swarm predictive control method for multi-UAV systems, characterized in that, The method includes: S1: Construct a drone swarm model, including several drones and several obstacles; establish a state equation for each drone, and set the motion mode and obstacle perception set of the virtual drone; S2: Construct the original control space of the UAV, input it into the state equation, and obtain the original predicted state set of the UAV; S3: Construct the potential function and cost function based on the obstacle perception set, the motion pattern of the virtual drone, and the original predicted state set; S4: Determine the initial solution based on the spatial gradient of the potential function and the expected direction of the cost function, and filter the initial solution to obtain the preliminary control space; S5: Transform the UAV swarm model into a Gibbs free field model, establish a joint probability density distribution problem, and set an update criterion based on the joint probability density distribution problem; input the initial control space into the state equation to obtain the initial predicted state set; S6: Use the update criterion to initially predict the state set and perform iterative optimization to calculate the corresponding joint probability density value; when the joint probability density value is the maximum, use the corresponding initially screened control space as the optimal control space, and use the optimal control space to control the UAV swarm.

2. The heuristic swarm predictive control method for multi-UAV systems according to claim 1, characterized in that, In S1, a state equation is established for each of the UAVs, and the state equation is as follows: Where, p i v represents the location of the drone. i For the speed of the drone, x i For the state of the drone, u i Here, m represents the control input for the drone, and m represents the mass of the drone. The state of the drone after inputting control input.

3. The heuristic swarm predictive control method for multi-UAV systems according to claim 1, characterized in that, In S2, the initial control space of the UAV is constructed, and the initial control space of the UAV is as follows: Among them, u t The sequence is an arithmetic progression, e(k1Δθ,k2Δφ) is the direction of the control input, and k1, k2, k θ ,k φ Δθ is an integer; Δφ is the angular interval of the horizontal control input direction; Δφ is the angular interval of the vertical control input direction. It is a set of integers.

4. The heuristic swarm predictive control method for multi-UAV systems according to claim 1, characterized in that, In S3, a potential function is constructed based on the obstacle perception set, the drone's motion pattern, and the original predicted state set. The potential function includes the potential function between drones and the potential function between drones and obstacles. The formula for the potential function between the drones is: Where, k a k is the first parameter. t It is the second parameter; r f It is the expected distance between adjacent drones; ||p ij || Represents drone q i With drones q l The distance between them; ψ ar Let the potential function be the relationship between the drones; The potential function between the UAV and the obstacle includes the repulsive potential function between the UAV and the obstacle and the deviation potential function between the UAV and the obstacle, and the formula is: ψ o =ψ ob +ψ or Where, ψ o Let ψ be the potential function between the drone and the obstacle; ob Let ψ be the repulsive potential function between the drone and the obstacle; or Let be the deviation potential function between the drone and the obstacle.

5. The heuristic swarm predictive control method for multi-UAV systems according to claim 4, characterized in that, The potential function between the UAV and the obstacle includes the repulsive potential function between the UAV and the obstacle and the deviation potential function between the UAV and the obstacle. The formula for the repulsive potential function between the UAV and the obstacle is: Where, k or The third parameter; ||p iβ || represents the distance between the drone and the obstacle; The formula for the deviation potential function between the drone and the obstacle is: Where, k ob r is the fourth parameter. ρ v represents the minimum distance between the obstacle and the drone's current velocity direction; go The desired direction to bypass the obstacle; v i Represents the current velocity; ρ(λ,δ,z) is the transition function; k δ This is the fifth parameter.

6. The heuristic swarm predictive control method for multi-UAV systems according to claim 1, characterized in that, In S3, the cost function includes a cost function for the virtual drone's position and a cost function for the virtual drone's speed; The formula for the cost function of the virtual drone's location is: ψ rp =k rp (exp(||p i1 ||)-1) Where, ψ rp The cost function for the virtual drone's location; k rp The sixth parameter; ||p i1 ||For drones q i Distance to the virtual drone; The formula for the cost function of the virtual drone's speed is: ψ rv =k rv (exp(||v i1 ||)-1) Where, ψ rv The cost function for the speed of the virtual drone; k rv The seventh parameter; ||v i1 ||For drones q i The relative speed with the virtual drone.

7. The heuristic swarm predictive control method for multi-UAV systems according to claim 1, characterized in that, The specific method of S6 is as follows: S61: Update criteria for initialization settings; S62: Set up a preliminary operation set, store the random variables in the preliminary prediction state set into the preliminary operation set, and use the set update criteria to calculate the variables to obtain the predicted marginal probability distribution value. S63: Select a variable from the preliminary operation set and remove it from the preliminary operation set. Calculate the variable using the set update criteria to obtain the preliminary marginal probability distribution value. If the difference between the predicted marginal probability distribution value and the preliminary marginal probability distribution value is greater than a preset value, then move the variable into the prediction operation set. Otherwise, maintain the state of removing the variable from the prediction operation set until all variables in the prediction operation set are used up; obtain the optimized operation set. S64: Calculate the joint probability density based on the optimized operation set; when the joint probability density value is the maximum, take the corresponding initially screened control space as the optimal control space, and use the optimal control space to control the UAV swarm.

8. The heuristic swarm predictive control method for multi-UAV systems according to claim 7, characterized in that, In S61, the update criterion is as follows: Among them, Z i For the allocation function; ψ ar Let ψ be the potential function between the drones; o Let ψ be the potential function between the drone and the obstacle; rp The cost function for the virtual drone's location; ψ rv x is the cost function for the speed of the virtual drone; i and x j The parameters are for predicting the state set.

9. The heuristic swarm predictive control method for multi-UAV systems according to claim 8, characterized in that, In S64, the joint probability density is calculated based on the optimized operation set, and the formula for the joint probability distribution is: Where p(X) is the joint probability density.

10. A heuristic swarm prediction control system for multi-UAV systems, used to implement the method described in claims 1-9, characterized in that, The system includes: The model building module constructs a drone swarm model, including several drones and several obstacles; it establishes a state equation for each drone and sets the motion mode and obstacle perception set of the virtual drone; The space construction module constructs the original control space of the UAV, inputs it into the state equation, and obtains the original predicted state set of the UAV. The function construction module constructs the potential function and cost function based on the obstacle perception set, the motion pattern of the virtual drone, and the original predicted state set; The spatial filtering module determines the initial solution based on the spatial gradient of the potential function and the expected direction of the cost function, and filters the initial solution to obtain the preliminary filtered control space. The model prediction module transforms the UAV swarm model into a Gibbs free-field model, establishes a joint probability density distribution problem, and sets update criteria based on the joint probability density distribution problem; the initial control space is input into the state equation to obtain the initial predicted state set; The control module is optimized by using the update criteria to initially predict the state set for iterative optimization and calculating the corresponding joint probability density value. When the joint probability density value is the maximum, the corresponding initially screened control space is taken as the optimal control space, and the UAV swarm is controlled using the optimal control space.

Images