Collaborative aircraft intelligent obstacle avoidance method based on adaptive artificial potential field

Abstract

The invention discloses a cooperative aircraft intelligent obstacle avoidance method based on a self-adaptive artificial potential field, and belongs to the technical field of unmanned aerial vehicle autonomous navigation. By constructing a hybrid intelligent architecture of'neural network prediction parameter-potential field method track generation ', the physical interpretability of an artificial potential field method is reserved, and the system is endowed with the ability of coping with nonlinear threats. According to the invention, an artificial neural network is added in an obstacle avoidance system of a collaborative aircraft and is trained, and dynamic adaptive prediction is carried out on core parameters of gravity and repulsive force gain coefficients of artificial potential field obstacle avoidance. According to the invention, the intelligent obstacle avoidance mode of the collaborative aircraft can be realized, and the safety and obstacle avoidance efficiency of the collaborative aircraft can be remarkably improved when the collaborative aircraft deals with a complex combat environment.

Description

Technical Field

[0001] The present invention belongs to the field of autonomous navigation technology for unmanned aerial vehicles (UAVs), and specifically relates to a dynamic adaptive parameter optimization method that combines deep learning and artificial potential fields, which is suitable for real-time obstacle avoidance control of collaborative aircraft in complex airspace environments. Background Art

[0002] Cooperative aircraft, as a key node in modern air combat systems, are the core combat element of manned / unmanned aerial vehicle (UAS) coordinated combat modes. Compared to traditional UAVs, cooperative aircraft offer numerous advantages, including greater speed, enhanced performance, and improved stealth. Their importance is particularly pronounced in low-altitude penetration scenarios. Low altitudes, due to their terrain masking effects and radar blind spots, offer a highly effective path to penetrate enemy air defenses. Equipped with advanced sensors and weapon systems, cooperative aircraft can perform forward reconnaissance, electromagnetic suppression, and precision strike missions, significantly reducing the exposure risk of manned aircraft. In complex mountainous or urban canyon environments, cooperative aircraft can leverage their low-altitude maneuverability to rapidly approach targets and provide battlefield situational awareness support to the lead aircraft via dynamic tactical data links. Current research indicates that formations equipped with cooperative aircraft can increase their penetration success rate by over 47%, and their mission effectiveness is directly impacting the battle for air superiority. Cooperative aircraft play an irreplaceable role in the modern, intelligent, and networked battlefield.

[0003] In the dynamic confrontation environment of low-altitude penetration, the obstacle avoidance system is the core subsystem that ensures the survivability and mission continuity of cooperative aircraft. Typical threats include three categories: static obstacles (such as high-voltage towers and mountain cliffs), dynamic obstacles (enemy interceptor missiles and suddenly moving vehicles), and electromagnetic soft-kill obstacles (directed energy weapon interference areas). According to statistics, 63.2% of drone accidents in the past five years were caused by obstacle avoidance failures. Especially during penetration at speeds exceeding 800 kilometers per hour, the time window for obstacle avoidance decisions is very small, and the traditional preset trajectory planning model can no longer cope with such extreme scenarios. Therefore, it is very necessary to conduct obstacle avoidance analysis of cooperative aircraft to improve the success rate of obstacle avoidance of cooperative aircraft.

[0004] Although traditional algorithms such as the Artificial Potential Field (APF) and Rapidly Exploring Random Tree (RRT) have been widely used for UAV obstacle avoidance, they have exposed significant shortcomings in real-world cooperative aircraft combat scenarios. Taking the most widely used APF as an example, its inherent problems primarily manifest in three aspects: static parameters lead to poor environmental adaptability, and fixed attraction / repulsion coefficients are unable to cope with the sudden change in threat gradients in mountainous and urban terrain; local optimal traps are frequent, easily generating oscillating trajectories in densely populated obstacle areas, resulting in a planning failure rate of at least 15%-22%; and the lack of dynamic obstacle modeling leads to response delays of up to 500 milliseconds for moving targets exceeding 300 kilometers per hour. Existing improvements, such as fuzzy logic control, can increase parameter flexibility, but their rule base construction relies on expert experience and cannot cover all scenarios in low-altitude penetration combat. While reinforcement learning-based obstacle avoidance methods offer adaptive capabilities, they face rigid constraints such as slow model convergence (requiring thousands of trial-and-error iterations) and excessive online computing resources. These problems severely restrict the practical deployment value of cooperative aircraft in high-intensity confrontation environments. How to adaptively improve the parameters of the artificial potential field according to the combat scenario has also become a key issue in improving the obstacle avoidance performance of cooperative aircraft.

[0005] Neural networks demonstrate unique advantages in solving problems. Through a multi-layer perceptron architecture and nonlinear activation functions, the algorithm can achieve high-dimensional mapping and feature extraction of complex battlefield situations. Compared with traditional machine learning methods, its core value is reflected in three aspects: first, dynamic modeling capabilities. The network weights can be updated online, and can adaptively learn the mapping relationship of the most advantageous field coefficients under different terrain, weather, and electromagnetic environments; second, excellent inference efficiency. The lightweight neural network model after pruning and quantization can achieve millisecond-level real-time prediction on the embedded AI chip, meeting the extreme timeliness requirements of penetration scenarios; finally, strong generalization performance. By integrating prior knowledge (such as aerodynamic constraints) with incremental learning mechanisms, the network can still maintain a success rate of more than 90% for obstacle avoidance types that have never been seen. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, the present invention provides a method for intelligent obstacle avoidance for collaborative aircraft based on an adaptive artificial potential field. By constructing a hybrid intelligent architecture of "neural network prediction parameters - potential field method trajectory generation", the method not only retains the physical interpretability of the artificial potential field method, but also gives the system the ability to cope with nonlinear threats. The present invention adds an artificial neural network to the obstacle avoidance system of the collaborative aircraft and trains it, and dynamically and adaptively predicts the gravitational and repulsive gain coefficients, the core parameters of the artificial potential field obstacle avoidance. Through the present invention, an intelligent obstacle avoidance method for collaborative aircraft can be realized, which can significantly improve the safety and obstacle avoidance efficiency of collaborative aircraft when dealing with complex combat environments.

[0007] The technical solution of the present invention:

[0008] A collaborative aircraft intelligent obstacle avoidance method based on an adaptive artificial potential field comprises the following steps:

[0009] Step 1: Establish an artificial potential field model for coordinated aircraft motion;

[0010] Step 2: Adaptive artificial potential field model based on neural network prediction;

[0011] Step 3: Adaptive artificial potential field parameter prediction and model optimization.

[0012] Furthermore, the step 1 is specifically as follows:

[0013] Assume that the cooperative aircraft is a particle in a spatial region, and there is a potential field in the spatial region; in the potential field environment, all particles and obstacles will have forces, and the potential field will also generate forces on all particles; considering the low-altitude penetration and obstacle avoidance combat scenario of the cooperative aircraft, the altitude change of the aircraft will not be considered, and the problem will be simplified to a two-dimensional problem; suppose the position X of the cooperative aircraft in the specified airspace is wingman for:

[0014] X wingman =(x wingman ,y wingman )

[0015] where x wingman is the X position of the collaborative aircraft, y wingman is the Y position of the cooperative aircraft; the X position of the target particle goal The gravitational potential function U generated on the cooperative aircraft att (X wingman )for:

[0016]

[0017] Where k att is the attraction coefficient, that is, the proportional gain coefficient; d goal For the coordinated aircraft and target particle X goal =(x goal ,y goal ), so we have:

[0018]

[0019] Among them, x goal is the X position of the target particle, y goal is the Y position of the cooperative aircraft; the X position of the target particle goal Collaborative Aircraft X wingman The resulting gravitational force is expressed as:

[0020]

[0021] in, is the gradient operator, which is the rate of change of a function at a point in a vector field;

[0022] Theoretically, if there is no obstacle, the cooperative aircraft will continue to fly towards the target particle due to the attraction. If an obstacle appears in the potential field, then the repulsive force U generated by the obstacle on the cooperative aircraft is rep (X wingman )for:

[0023]

[0024] Where k rep is the obstacle repulsion coefficient, d obs is the distance between the obstacle and the cooperative aircraft, d r is the distance at which the obstacle affects the cooperative aircraft, which is set as a constant; then the repulsive force of the obstacle on the cooperative aircraft is expressed as:

[0025]

[0026] According to the principle of potential field superposition, the sum of the forces acting on the cooperative aircraft in the potential field is:

[0027] F total =F att (X wingman )+F rep (X wingman )

[0028] When there are many obstacles in the potential field, the net force acting on the cooperative aircraft is:

[0029]

[0030] Where n is the number of obstacles;

[0031] The cooperative aircraft can start to move under the influence of the resultant force in this potential field; the trajectory of the cooperative aircraft is affected by its own position, obstacles, and the target particle. First, use the geometric method to determine the angle φ between the cooperative aircraft and the target particle:

[0032]

[0033] Then the gravitational components of the gravitational vector of the target particle on the cooperative aircraft on the X-axis and Y-axis are:

[0034] F att (x wingman )=F att (X wingman )·cos(φ)

[0035] F att (y wingman )=F att (X wingman )·sin(φ)

[0036] Similarly, when there are multiple obstacles, the positions of these obstacles can be positioned as X obs1 =(x obs1 ,y obs1 ), X obs2 =(x obs2 ,y obs2 ),…,X obsn =(x obsn ,y obsn ), where n is the number of obstacles; then the cooperative aircraft and the i-th obstacle X obsi =(x obsi ,y obsi ) It can be obtained through geometric relations:

[0037]

[0038] Then the repulsive force components of the i-th obstacle's repulsive force on the cooperative aircraft on the X-axis and Y-axis are:

[0039]

[0040] By calculating the attraction and repulsion, the net force F total Decompose to X axis F total (x) and Y axis F total (y) Get the angle between the movement direction of the cooperative aircraft and the target direction:

[0041]

[0042] And where to move next:

[0043]

[0044] Furthermore, the step 2 is specifically as follows:

[0045] Analyze the input features of the neural network, including the distance d between the collaborative aircraft and the target particle goal ; The distance d between the cooperative aircraft and the obstacle obs ; Surrounding obstacle density, that is, the number of obstacles within the sensor range n; The current speed of the cooperative aircraft v wingman ;The angle θ between the moving direction and the target direction;

[0046] Construct an artificial neural network with 5 nodes in the input layer. According to expert experience, the hidden layer nodes are set to 16 and 32 respectively, and the ReLU activation function is used. The output layer has 2 nodes and uses the Sigmoid function. The output layer corresponds to the attraction coefficient k att and repulsion coefficient k rep ;

[0047] The output of a single neuron is expressed as:

[0048]

[0049] x i For the input of the i-th node, w ij represents the weight between the i-th input neuron node and the j-th hidden layer neuron node, b j Represents the bias term of the jth neuron, f is the activation function; the input data is forward propagated through the network to obtain the l-layer output a (l) :

[0050] a (l) =f (l-1) (W (l-1) a (l-1) +b (l-1) )

[0051] Among them, f (l) 、W (l) 、b (l) are the activation function, weight vector and bias vector of the lth layer of the neural network respectively;

[0052] Output layer metric function for:

[0053]

[0054] In the formula, α and β are weight coefficients, balancing the importance of parameters; They are the predicted value and true value of the attraction coefficient, and the predicted value and true value of the obstacle repulsion coefficient; after backpropagation of the error, the weights and neuron biases of the neural network are updated to complete the training of the entire neural network:

[0055]

[0056] Among them, η is the neural network learning rate.

[0057] Furthermore, the step 3 is specifically as follows:

[0058] The current environmental characteristics are collected every Δt seconds and input into the adaptive artificial potential field model based on neural network prediction to obtain the attraction coefficient k att and repulsion coefficient k repThe predicted value of the artificial potential field is updated in real time:

[0059]

[0060] Among them, q goal ,q,q obs,i are the coordinates of the target particle, the cooperative aircraft, and the i-th obstacle in the three-dimensional space respectively; Represents the coordinates of the intelligent body's configuration space;

[0061] Use neural network weight pruning technology to optimize the neural network structure:

[0062]

[0063] τ is the pruning threshold.

[0064] A collaborative aircraft intelligent obstacle avoidance method based on adaptive artificial potential field cleverly solves the obstacle avoidance problem of collaborative aircraft by constructing a hybrid intelligent architecture of "neural network prediction parameters-potential field method to generate trajectory". In theory, the distance between the collaborative aircraft and the target, the distance between the collaborative aircraft and the nearest obstacle, the density of obstacles (the number of surrounding obstacles), the current speed of the collaborative aircraft, the angle between the direction of movement of the collaborative aircraft and the target direction, etc. are learned through the neural network. These features can reflect the complexity and dynamic changes of the current battlefield environment, help the network determine when to increase or decrease the gravity and repulsion, thereby solving the problem of poor environmental adaptability caused by static parameters; each time the path planning is iterated, the current environmental state is input into the neural network to obtain a new attraction coefficient k att and repulsion coefficient k rep , updating the potential field calculation and generating the next direction of motion. This forms a closed-loop system that must ensure system stability and convergence, addressing the frequent occurrence of local optimal traps and the proneness to oscillating trajectories in densely populated obstacle areas. Finally, path planning for collaborative aircraft typically requires rapid response, so the neural network design should be as lightweight as possible to reduce computational latency. Using a smaller network structure and employing model compression techniques (such as quantization and pruning) address the high latency caused by the lack of dynamic obstacle modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a specific implementation flow chart of the present invention.

[0066] Figure 2 Schematic diagram of the artificial potential field for coordinated aircraft motion according to the present invention.

[0067] Figure 3 It is the artificial neural network prediction model diagram of the present invention.

[0068] Figure 4This is the simulation result diagram of the traditional artificial potential field method for cooperative aircraft under fixed parameters.

[0069] Figure 5 This is a diagram of the simulation results of the traditional artificial potential field method for a collaborative aircraft that is trapped in a local optimum.

[0070] Figure 6 This is a diagram of the simulation results of the traditional artificial potential field method for cooperative aircraft with increased speed.

[0071] Figure 7 This is the simulation result of cooperative aircraft intelligent obstacle avoidance based on adaptive artificial potential field. DETAILED DESCRIPTION

[0072] 1. Introduction to the principle of artificial potential field method

[0073] The artificial potential field method is a path planning algorithm based on a virtual force field, first proposed by Oussama Khatib. It is an effective method to guide and control the direction of movement of objects (such as robots, drones, etc.) by simulating the interaction between gravity and repulsion in the physical field. In the construction of the physical model of the artificial potential field method, the system realizes motion control and obstacle avoidance functions through the charge interaction mechanism. Specifically, the moving subject can be abstracted as a positive charge carrier, the target point can be modeled as a negative potential attraction source, and potential collision bodies such as obstacles can be configured as entities with the same polarity charge. Based on the Coulomb action law in electromagnetic field theory, the attraction between opposite polarity charges drives the intelligent body to move gradually towards the target posture, while the repulsive effect between like charges forms a dynamic safety barrier within its radius of action. This dual potential field coupling mechanism can be expressed mathematically as:

[0074]

[0075] in represents the coordinates of the intelligent body’s configuration space, k att and k rep are the potential field gain coefficients. This model ensures two key characteristics through the principle of vector superposition: 1) forming a gradient descent path pointing to the target in free space; 2) generating a repulsive potential field in the vicinity of the obstacle that increases inversely with the distance. Theoretical analysis and simulation experiments show that when ||q goal -q||<ε and ||qq obs,i ||>d safe When , the system can simultaneously achieve the dual goals of gradual convergence and collision avoidance, where ε is the convergence threshold, d safe The preset safety distance.

[0076] The core idea of ​​the artificial potential field method is to first generate a gravitational field through the target point to attract the intelligent agent to move in the target direction, while at the same time modeling obstacles to generate a repulsive field to repel the intelligent agent away from the obstacle area. Finally, through combined force navigation, the intelligent agent moves according to the direction of the synthetic vector of gravity and repulsion to achieve dynamic obstacle avoidance.

[0077] 2. Introduction to the principles of neural networks

[0078] A neural network is a computational model inspired by the information processing mechanisms of biological nervous systems. Its core design concept stems from the abstract simulation of biological neuron structures and information transmission mechanisms. In biological nervous systems, neurons form networks through synaptic connections, integrating and transmitting information through electrochemical signals. Artificial neural networks transform this process into a computable form through mathematical modeling. Each artificial neuron receives multiple input signals, and after weighted summation, it generates an output through a nonlinear activation function (such as ReLU, Sigmoid, etc.), simulating the threshold response characteristics of biological neurons. By interconnecting a large number of such neurons in a hierarchical structure (input layer, hidden layer, output layer), neural networks can construct deep nonlinear mapping relationships. Their multi-layer stacking characteristics give them powerful feature abstraction capabilities. Shallow networks capture local detailed features (such as edges and textures), while deep networks gradually fuse layers to form global semantic representations (such as object contours and semantic categories).

[0079] This hierarchical feature learning mechanism enables neural networks to break through the limitations of traditional linear models and theoretically satisfy the Universal Approximation Theorem, which states that a single hidden layer network can approximate any continuous function with arbitrary precision. Compared to traditional machine learning that relies on manual feature engineering, neural networks automatically extract the inherent laws of data through end-to-end training, demonstrating excellent complex pattern modeling capabilities in areas such as image recognition, natural language processing, and reinforcement learning. Its training process uses a backpropagation algorithm and gradient descent optimization to minimize prediction errors by dynamically adjusting network weights. It also incorporates techniques such as Dropout pruning and batch normalization to enhance the model's generalization performance, ultimately achieving an intelligent computing paradigm that autonomously learns multi-level distributed representations from high-dimensional data.

[0080] 3. Introduction to a collaborative aircraft intelligent obstacle avoidance method based on adaptive artificial potential field

[0081] In the artificial potential field method, the target point generates an attractive force, and the obstacle generates a repulsive force, and the combined force guides the robot to move. att and repulsion coefficient k repThese parameters are typically fixed, but dynamic changes in the real environment may require dynamic adjustment. For example, in complex or dynamic obstacle environments, fixed parameters may lead to local minima or oscillation. Therefore, dynamically adjusting these coefficients may improve the adaptability and robustness of path planning.

[0082] Neural networks can learn the complex nonlinear relationship between input features and output through training. If the environment state (such as the robot's current position, target position, obstacle distribution, speed, etc.) can be used as input, the output attraction coefficient k att and repulsion coefficient k rep The optimal value of , then the neural network can adjust these parameters in real time to adapt to the needs of different scenarios.

[0083] Cooperative aircraft encounter three major issues when using artificial potential field methods for obstacle avoidance: 1. Static parameters lead to poor environmental adaptability. Fixed attraction / repulsion coefficients are difficult to handle in mountainous and urban terrain, such as the sudden threat gradient. 2. Local optimal traps frequently occur, easily generating oscillating trajectories in densely populated areas, resulting in planning failures. The lack of dynamic obstacle modeling leads to high latency, with response delays of up to 500 milliseconds for moving targets exceeding 300 kilometers per hour. This problem is equivalent to the obstacle avoidance problem of a high-speed cooperative aircraft dealing with stationary obstacles. The speed advantage of cooperative aircraft compared to traditional drones becomes a disadvantage in the obstacle avoidance field, resulting in delayed avoidance of obstacles.

[0084] A collaborative aircraft intelligent obstacle avoidance method based on adaptive artificial potential field, such as Figure 1 As shown, the following steps are included:

[0085] Step 1: Establish an artificial potential field model for coordinated aircraft motion.

[0086] Assume that the cooperative aircraft is a point mass in a spatial region, and there is a potential field in the spatial region. In this potential field environment, all points of mass and obstacles will have forces acting on them, and the potential field will also generate forces on all points of mass. Considering the low-altitude penetration and obstacle avoidance combat scenario of the cooperative aircraft, the altitude change of the aircraft will not be considered, and the problem will be simplified to a two-dimensional problem. Assume that the position X of the cooperative aircraft in the specified airspace is wingman for:

[0087] X wingman =(x wingman ,y wingman )

[0088] where x wingman is the X position of the collaborative aircraft, y wingman is the Y position of the cooperative aircraft; the X position of the target particle goal The gravitational potential function U generated on the cooperative aircraft att (X wingman)for:

[0089]

[0090] Where k att is the attraction coefficient, that is, the proportional gain coefficient. goal For the coordinated aircraft and target particle X goal =(x goal ,y goal ), because the cooperative aircraft is simplified into a two-dimensional point mass in this patent, we have:

[0091]

[0092] Among them, x goal is the X position of the target particle, y goal is the Y position of the cooperative aircraft; the X position of the target particle goal Collaborative Aircraft X wingman The resulting gravitational force is expressed as:

[0093]

[0094] in, is the gradient operator, which is the rate of change of a function at a point in a vector field.

[0095] Theoretically, if there is no obstacle, the cooperative aircraft will continue to fly towards the target particle due to the attraction. If an obstacle appears in the potential field, then the repulsive force U generated by the obstacle on the cooperative aircraft is rep (X wingman )for:

[0096]

[0097] Where k rep is the obstacle repulsion coefficient, d obs is the distance between the obstacle and the cooperative aircraft, d r The distance at which the obstacle affects the cooperative aircraft is set to a constant because it is determined by the obstacle itself. Then the repulsive force of the obstacle on the cooperative aircraft is expressed as:

[0098]

[0099] According to the principle of potential field superposition, the sum of the forces acting on the cooperative aircraft in the potential field is:

[0100] F total =F att (X wingman )+F rep (X wingman )

[0101] When there are many obstacles in the potential field, the net force acting on the cooperative aircraft is:

[0102]

[0103] Where n is the number of obstacles;

[0104] like Figure 2 As shown, the cooperative aircraft can start to move under the influence of the resultant force in this potential field. The trajectory of the cooperative aircraft is affected by its own position, obstacles, and the target particle. First, use the geometric method to determine the angle φ between the cooperative aircraft and the target particle:

[0105]

[0106] Then the gravitational components of the gravitational vector of the target particle on the cooperative aircraft on the X-axis and Y-axis are:

[0107] F att (x wingman )=F att (X wingman )·cos(φ)

[0108] F att (y wingman )=F att (X wingman )·sin(φ)

[0109] Similarly, when there are multiple obstacles, the positions of these obstacles can be positioned as X obs1 =(x obs1 ,y obs1 ), X obs2 =(x obs2 ,y obs2 ),…,X obsn =(x obsn ,y obsn ), where n is the number of obstacles. Then the cooperative aircraft and the i-th obstacle X obsi =(x obsi ,y obsi ) It can be obtained through geometric relations:

[0110]

[0111] Then the repulsive force components of the i-th obstacle's repulsive force on the cooperative aircraft on the X-axis and Y-axis are:

[0112]

[0113] By calculating the attraction and repulsion, the net force F totalDecompose to X axis F total (x) and Y axis F total (y) Get the angle between the movement direction of the cooperative aircraft and the target direction:

[0114]

[0115] And where to move next:

[0116]

[0117] Step 2: Adaptive artificial potential field model based on neural network prediction

[0118] Analyze the input features of the neural network, including the distance d between the collaborative aircraft and the target particle goal ; The distance d between the cooperative aircraft and the obstacle obs ; Surrounding obstacle density, that is, the number of obstacles within the sensor range n; The current speed of the cooperative aircraft v wingman ; The angle θ between the moving direction and the target direction.

[0119] like Figure 3 As shown in the figure, an artificial neural network is constructed with 5 nodes in the input layer. Based on expert experience, the hidden layer nodes are set to 16 and 32 respectively, and the ReLU activation function is used. The output layer has 2 nodes and uses the Sigmoid function. The output layer corresponds to the attraction coefficient k att and repulsion coefficient k rep .

[0120] The output of a single neuron is expressed as:

[0121]

[0122] x i For the input of the i-th node, w ij represents the weight between the i-th input neuron node and the j-th hidden layer neuron node, b j Represents the bias term of the jth neuron, and f is the activation function. The input data is forward propagated through the network to obtain the l-layer output a (l) :

[0123] a (l) =f (l-1) (W (l-1) a (l-1) +b (l-1) )

[0124] Among them, f (l) 、W (l) 、b (l) are the activation function, weight vector and bias vector of the lth layer of the neural network respectively;

[0125] Output layer metric function for:

[0126]

[0127] In the formula, α and β are weight coefficients, balancing the importance of parameters; They are the predicted value and true value of the attraction coefficient, and the predicted value and true value of the obstacle repulsion coefficient. After backpropagation of the error, the weights and neuron biases of the neural network are updated to complete the training of the entire neural network:

[0128]

[0129] Where η is the neural network learning rate;

[0130] Step 3: Adaptive artificial potential field parameter prediction and model optimization

[0131] The current environmental characteristics are collected every Δt seconds and input into the adaptive artificial potential field model based on neural network prediction to obtain the attraction coefficient k att and repulsion coefficient k rep The predicted value of the artificial potential field is updated in real time:

[0132]

[0133] Among them, q goal ,q,q obs,i are the coordinates of the target particle, the cooperative aircraft, and the i-th obstacle in the three-dimensional space respectively; Represents the coordinates of the intelligent body's configuration space;

[0134] Use neural network weight pruning technology to optimize the neural network structure:

[0135]

[0136] τ is the pruning threshold, such as the 10% quantile of the absolute value distribution of the weight.

[0137] Implementation examples:

[0138] The traditional artificial potential field method will have the following problems when simulating cooperative aircraft obstacle avoidance scenarios: 1. Static parameters lead to poor environmental adaptability, and the simulation effect is as follows: Figure 4 As shown; 2. Local optimal traps occur frequently, and oscillating trajectories are easily generated in dense obstacle areas, resulting in planning failure rate. The simulation results are shown as follows Figure 5 As shown; 3. The high delay caused by the lack of dynamic obstacle modeling will lead to the problem of not being able to avoid obstacles in time. The simulation effect is as follows Figure 6According to the above specific implementation steps, a collaborative aircraft intelligent obstacle avoidance method based on adaptive artificial potential field can be realized, and the simulation effect is as follows: Figure 7 The following table shows the performance comparison between the traditional artificial potential field method and the method of the present invention:

[0139] Table 1 Summary and comparison of results

[0140]

[0141] The beneficial effects of the present invention are as follows:

[0142] 1. The present invention uses a neural network to perceive the battlefield situation in real time (including characteristics such as the cooperative aircraft-target distance, the cooperative aircraft-obstacle distance, obstacle density, speed and heading angle), and constructs a "dynamic perception-parameter prediction-potential field update" closed-loop system, which enhances the dynamic environment adaptability of the cooperative aircraft and optimizes the obstacle avoidance success rate.

[0143] 2. The present invention proposes a hybrid architecture based on neural network and artificial potential field method, which suppresses local minima through a closed-loop feedback mechanism. Through the hybrid architecture of potential field and neural network, the closed-loop feedback mechanism suppresses local minima, effectively reduces the oscillating trajectory of the cooperative aircraft under high-speed motion, and accelerates the global convergence speed.

[0144] 3. This invention uses a lightweight neural network (two hidden layers, less than 10k parameters) and model compression technology (weight pruning) to achieve millisecond-level inference latency. This improves computational efficiency and can more effectively meet the requirements of high-speed combat scenarios involving coordinated aircraft.

Claims

1. A collaborative aircraft intelligent obstacle avoidance method based on adaptive artificial potential field, characterized in that: The following steps are involved: Step 1: Establish an artificial potential field model for coordinated aircraft motion; Step 2: Adaptive artificial potential field model based on neural network prediction; Step 3: Adaptive artificial potential field parameter prediction and model optimization.

2. The method for cooperative aircraft intelligent obstacle avoidance based on adaptive artificial potential field according to claim 1, characterized in that: The step 1 is specifically as follows: Assume that the cooperative aircraft is a particle in a spatial region, and there is a potential field in the spatial region; in the potential field environment, all particles and obstacles will have forces, and the potential field will also generate forces on all particles; considering the low-altitude penetration and obstacle avoidance combat scenario of the cooperative aircraft, the altitude change of the aircraft will not be considered, and the problem will be simplified to a two-dimensional problem; suppose the position X of the cooperative aircraft in the specified airspace is wingman for: X wingman =(x wingman ,y wingman ) where x wingman is the X position of the collaborative aircraft, y wingman is the Y position of the cooperative aircraft; the X position of the target particle goal The gravitational potential function U generated on the cooperative aircraft att (X wingman )for: Where k att is the attraction coefficient, that is, the proportional gain coefficient; d goal For the coordinated aircraft and target particle X goal =(x goal ,y goal ), so we have: Among them, x goal is the X position of the target particle, y goal is the Y position of the cooperative aircraft; the X position of the target particle goal Collaborative Aircraft X wingman The resulting gravitational force is expressed as: in, is the gradient operator, which is the rate of change of a function at a certain point in the vector field; Theoretically, if there is no obstacle, the cooperative aircraft will continue to fly towards the target particle due to the attraction. If an obstacle appears in the potential field, then the repulsive force U generated by the obstacle on the cooperative aircraft is rep (X wingman )for: Where k rep is the obstacle repulsion coefficient, d obs is the distance between the obstacle and the cooperative aircraft, d r is the distance at which the obstacle affects the cooperative aircraft, which is set as a constant; then the repulsive force of the obstacle on the cooperative aircraft is expressed as: According to the principle of potential field superposition, the sum of the forces acting on the cooperative aircraft in the potential field is: F total =F att (X wingman )+F rep (X wingman ) When there are many obstacles in the potential field, the net force acting on the cooperative aircraft is: Where n is the number of obstacles; The cooperative aircraft can start to move under the influence of the resultant force in this potential field; the trajectory of the cooperative aircraft is affected by its own position, obstacles, and the target particle. First, use the geometric method to determine the angle φ between the cooperative aircraft and the target particle: Then the gravitational components of the gravitational vector of the target particle on the cooperative aircraft on the X-axis and Y-axis are: F att (x wingman )=F att (X wingman )·cos(φ) F att (y wingman )=F att (X wingman )·sin(φ) Similarly, when there are multiple obstacles, the positions of these obstacles can be positioned as X obs1 =(x obs1 ,y obs1 ), X obs2 =(x obs2 ,y obs2 ),…,X obsn =(x obsn ,y obsn ), where n is the number of obstacles; then the cooperative aircraft and the i-th obstacle X obsi =(x obsi ,y obsi ) It can be obtained through geometric relations: Then the repulsive force components of the i-th obstacle's repulsive force on the cooperative aircraft on the X-axis and Y-axis are: By calculating the attraction and repulsion, the net force F total Decompose to X axis F total (x) and Y axis F total (y) Get the angle between the movement direction of the cooperative aircraft and the target direction: And where to move next:

3. The method for cooperative aircraft intelligent obstacle avoidance based on adaptive artificial potential field according to claim 1, characterized in that: The step 2 is as follows: Analyze the input features of the neural network, including the distance d between the collaborative aircraft and the target particle goal ; The distance d between the cooperative aircraft and the obstacle obs ; Surrounding obstacle density, that is, the number of obstacles within the sensor range n; The current speed of the cooperative aircraft v wingman ;The angle θ between the moving direction and the target direction; Construct an artificial neural network with 5 nodes in the input layer. According to expert experience, the hidden layer nodes are set to 16 and 32 respectively, and the ReLU activation function is used. The output layer has 2 nodes and uses the Sigmoid function. The output layer corresponds to the attraction coefficient k att and repulsion coefficient k rep ; The output of a single neuron is expressed as: x i For the input of the i-th node, w ij represents the weight between the i-th input neuron node and the j-th hidden layer neuron node, b j Represents the bias term of the jth neuron, f is the activation function; the input data is forward propagated through the network to obtain the l-layer output a (l) : a (l) =f (l-1) (W (l-1) a (l-1) +b (l-1) ) Among them, f (l) 、W (l) 、b (l) are the activation function, weight vector and bias vector of the lth layer of the neural network respectively; Output layer metric function for: In the formula, α and β are weight coefficients, balancing the importance of parameters; They are the predicted value and true value of the attraction coefficient, and the predicted value and true value of the obstacle repulsion coefficient; after backpropagation of the error, the weights and neuron biases of the neural network are updated to complete the training of the entire neural network: Among them, η is the neural network learning rate.

4. The method for cooperative aircraft intelligent obstacle avoidance based on adaptive artificial potential field according to claim 1, characterized in that: The step 3 is as follows: The current environmental characteristics are collected every Δt seconds and input into the adaptive artificial potential field model based on neural network prediction to obtain the attraction coefficient k att and repulsion coefficient k rep The predicted value of the artificial potential field is updated in real time: Among them, q goal ,q,q obs,i are the coordinates of the target particle, the cooperative aircraft, and the i-th obstacle in the three-dimensional space respectively; Represents the coordinates of the intelligent body's configuration space; Use neural network weight pruning technology to optimize the neural network structure: τ is the pruning threshold.

Images