Intelligent defense and attack control system for unmanned aerial vehicle

Abstract

The invention belongs to the technical field of unmanned aerial vehicle attack and defense, and discloses an intelligent defense and attack control system for an unmanned aerial vehicle. The method comprises the following steps: acquiring environment sensing data of an unmanned aerial vehicle, wherein the environment sensing data comprises state data of the unmanned aerial vehicle and image data of a surrounding environment; preprocessing the image data of the surrounding environment to obtain model input image data; performing target detection on the model input image data by using a variant Poolform algorithm to obtain comprehensive information of a target object; according to the state data of the unmanned aerial vehicle and the comprehensive information of the target object, solving an optimal attack defense strategy by using a perception integrated attack algorithm; the optimal attack defense strategy is sent to an unmanned aerial vehicle execution terminal, and the unmanned aerial vehicle is controlled to execute the optimal attack defense strategy; the optimal attack and defense strategy can be dynamically generated according to the real-time change of the environment situation, and the survivability and the attack efficiency of the unmanned aerial vehicle in the confrontation environment are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of unmanned aerial vehicle attack and defense technology, and more specifically, to an unmanned aerial vehicle intelligent defense and attack control system. Background Art

[0002] The patent with application publication number CN116820126A discloses a deep reinforcement learning system and method based on multi-UAV path planning. By building a virtual environment model of the multi-UAV path planning scenario, a deep reinforcement learning algorithm is used in the model to train the intelligent UAV, and the state information collected during the flight of the UAV is input into the neural network of deep reinforcement learning for model training; during the training process, an adversarial attack algorithm is used to interfere with the Markov decision sequence received by the intelligent agent; a state and reward dual defense framework is constructed to complete the defense against attack disturbances; the disturbed Markov sequence is input into the defense framework for data recovery; the restored data sequence is placed in the experience pool for the UAV to complete subsequent action decisions, and the optimal flight path of the UAV is obtained through multiple rounds of model training; the trained reinforcement learning model is subjected to adversarial attacks; and the model that has completed the defense test is deployed to the UAV edge computing platform.

[0003] However, the existing UAV defense and attack control systems still have many shortcomings in actual environments. In complex environments, the target recognition and tracking capabilities are weak, especially under conditions such as severe weather, complex terrain or electromagnetic interference. Targets are easily lost or misidentified, which seriously affects the effect; in terms of threat assessment, simple distance thresholds or fixed weight assessment methods are often used, which cannot accurately reflect the true threat level of the target, resulting in wrong decisions when facing multi-target threats; in terms of attack and defense strategy formulation, traditional systems mostly use preset fixed strategies, lack the ability to respond quickly to changes in the battlefield situation, and cannot adjust the attack and defense strategies in time according to the dynamic changes in the situation of both sides, which can easily lead to mission failure or UAV loss in a high-intensity confrontation environment; at the same time, the existing system is difficult to effectively balance multiple key factors in the process of executing the mission, and often makes decisions that are too radical or too conservative; in addition, when facing new threats or unknown situations, it is impossible to continuously improve combat effectiveness.

[0004] In view of this, the present invention proposes an unmanned aerial vehicle intelligent defense and attack control system to solve the above problems. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned purpose, the present invention provides the following technical solutions: an intelligent defense and attack control system for unmanned aerial vehicles, comprising: a data acquisition module, for acquiring environmental perception data of the unmanned aerial vehicle, the environmental perception data including the state data of the unmanned aerial vehicle itself and the image data of the surrounding environment;

[0006] A data processing module is used to pre-process the image data of the surrounding environment to obtain model input image data;

[0007] The target detection module is used to perform target detection on the model input image data using a variant Poolformer algorithm to obtain comprehensive information about the target object;

[0008] The strategy fitting module is used to solve the optimal attack and defense strategy based on the drone's own status data and the comprehensive information of the target object using the perception-integrated attack algorithm; the optimal attack and defense strategy is sent to the drone execution terminal to control the drone to execute the optimal attack and defense strategy; each module is connected by wired and / or wireless means.

[0009] Furthermore, the state data of the drone itself includes position data, motion state, energy state, weapon state, health management state and environment state; the environment state includes temperature, air pressure, wind speed, visibility, electromagnetic interference intensity and obstacle information;

[0010] The obstacle information includes the three-dimensional spatial position coordinates of the obstacle, the obstacle size, and the motion trajectory of the dynamic obstacle.

[0011] Furthermore, the method for obtaining the motion trajectory of the dynamic obstacle includes:

[0012] Dynamic obstacles are regarded as moving particles. Drones and other obstacles generate virtual gravitational and repulsive fields on the particles. The particles move under the combined force of the gravitational and repulsive fields.

[0013] The drone is defined as a gravitational source; the magnitude of the gravitational field is inversely proportional to the distance between the drone and the dynamic obstacle; the direction of the gravitational field points to the location of the drone;

[0014] The other obstacles are defined as repulsive sources. The magnitude of the repulsive field is inversely proportional to the distance between the repulsive source and the corresponding particle, and the direction of the repulsive field points away from the repulsive source.

[0015] For each time step, the resultant force acting on the particle is calculated, which is equal to the sum of the gravitational field plus the sum of the repulsive field; the resultant force is substituted into Newton's equation of motion, the acceleration a is obtained, and numerical integration is performed to obtain the velocity and displacement of the particle. A time range is preset, and the changes in the velocity and displacement of the particle within the time range are constructed to obtain the three-dimensional position sequence of the drone, which is the motion trajectory of the dynamic obstacle.

[0016] Furthermore, the method of preprocessing the image data of the surrounding environment includes:

[0017] Perform attention fusion equalization processing on the original image data to obtain preliminary input image data; perform image cropping on the preliminary input image data to obtain a key area image; scale the pixel values ​​in the key area image to a range of 0-1 or -1 to 1 to obtain model input image data;

[0018] The method of performing attention fusion and equalization processing includes:

[0019] Define an attention structure based on a deep convolutional neural network; the input of the attention structure is the original image data, and the output is an attention weight map of the same size as the original image data; each pixel value in the attention weight map represents the importance of the corresponding position;

[0020] Collect images in various scenes and conditions to form an image dataset, where the images contain target objects; manually annotate the images in the image dataset, and mark the location and category of the target objects with bounding boxes on the images;

[0021] The attention structure is integrated into the classification network or the target detection network to obtain a comprehensive network model; in the feature extraction stage of the classification network or the target detection network, the attention structure multiplies the output attention weight map as a mask with the original image data to obtain a feature map;

[0022] Perform evolutionary histogram equalization on the original image data to obtain a globally enhanced image; multiply the feature map and the globally enhanced image to obtain preliminary input image data.

[0023] Furthermore, the method of performing evolutionary histogram equalization includes:

[0024] The original image data is divided into a number of overlapping small blocks, the size of each small block is preset; a certain overlapping area is set between adjacent small blocks; each small block is subjected to histogram equalization processing to obtain a reconstructed small block;

[0025] For each reconstructed small block, calculate the normalized coordinates of its center point in the original image data; record them as the center point coordinates; calculate the Euclidean distance d from the center of the reconstructed small block to the center of the original image data based on the center point coordinates; for each reconstructed small block, count the number of pixels overlapping with other reconstructed small blocks; for each reconstructed small block, calculate its grayscale co-occurrence matrix; based on the comprehensive features in the grayscale co-occurrence matrix; the comprehensive features include contrast DA, energy DH, homogeneity DF and entropy DS;

[0026] Contrast ratio DA = ∑ I ∑ J (IJ) γ ×P(I,J) δ; Where I and J are the row and column indices of the grayscale co-occurrence matrix; P(I,J) is an element in the grayscale co-occurrence matrix, which indicates the frequency of occurrence of pixel pairs between grayscale values ​​I and J; γ is an exponential factor that controls the degree of influence of grayscale difference on contrast, and δ is an exponential factor that controls the degree of influence of the frequency of occurrence of pixel pairs on contrast;

[0027] Homogeneity Where a and b are parameters that control the shape of the function; entropy DS = ∑ I ∑ J P(I,J)×log(P(I,J));

[0028] The comprehensive features are normalized and weightedly fused to obtain the comprehensive texture complexity index of the corresponding reconstructed small block area; for each reconstructed small block area, the edge detection operator is applied to extract the number of edge pixels and the cumulative gradient amplitude, and weighted fusion is performed to obtain the edge strength index of the reconstructed small block area; the comprehensive texture complexity index and the edge strength index are fused into a comprehensive feature index C;

[0029] Define the position weight of the reconstructed small area h Among them, σ is the parameter that controls the weight decay speed; ω1 is the influence parameter; ω2 is the exponential factor that controls the distance, which is used to control the influence of distance on the weight;

[0030] Define the overlap weight of the reconstructed small area h Where a1 is the scaling factor; No is the number of pixels that overlap between the reconstructed small block h and other reconstructed small blocks, Nb is the total number of pixels in the reconstructed small block h; a2 is the overlap index factor; γ2 is a control parameter; the position weight and overlap weight of the reconstructed small block h are multiplied to obtain the comprehensive weight coefficient of the corresponding reconstructed small block;

[0031] For the overlapping areas in the reconstructed small blocks, the pixel values ​​therein are weighted averaged according to the comprehensive weight coefficients of the corresponding reconstructed small blocks to obtain the fused pixel values; for the non-overlapping areas in the reconstructed small blocks, the corresponding pixel values ​​are directly used to obtain a complete globally enhanced image.

[0032] Furthermore, the method of cropping the preliminary input image data includes:

[0033] Based on the motion trajectory of the dynamic obstacle, the field of view volume of each time step within the time range is determined. For each time step t on the motion trajectory, the camera coordinate system is established with the position of the drone as the origin. According to the attitude angle of the drone, the directions of the three coordinate axes of the camera coordinate system are determined; according to the field of view angle, the coordinates of the eight vertices of the camera cone in the camera coordinate system are calculated; using the position and attitude angle of the drone, the vertex coordinates in the camera coordinate system are converted to the world coordinate system, and the coordinates of the eight vertices in the world coordinate system are obtained. Using the coordinates of the eight vertices in the world coordinate system, a three-dimensional cone or pyramid is constructed; the three-dimensional cone or pyramid is the field of view of the drone at time step t; and then a series of time-varying three-dimensional field of view volumes V(t) are obtained;

[0034] For each time step t, the corresponding three-dimensional field of view volume V(t) is projected onto the two-dimensional image plane to obtain a series of two-dimensional polygonal regions P(t);

[0035] Initialize a mask image M of the same size as the global enhanced image, and all pixel values ​​in the mask image M are initially 0; for each time step t, set the pixel values ​​in the corresponding two-dimensional polygonal area P(t) on the mask image M to 1; multiply the global enhanced image and the mask image M pixel by pixel to obtain the key area image.

[0036] Furthermore, the method of performing target detection on the model input image data includes:

[0037] Construct a variant Poolformer model. The basic structure of the variant Poolformer model includes an encoder and a decoder. Define the encoder to include a trunk feature branch and a parallel Fourier transform branch. The Fourier transform branch performs a two-dimensional discrete Fourier transform on the model input image data to convert it from the spatial domain to the frequency domain. In the frequency domain, the model input image data is low-pass filtered and high-pass filtered.

[0038] Insert a Conv1x1 convolution layer between the Fourier transform branch and the trunk feature branch; the weight W1 of the Conv1x1 layer corresponding to the trunk feature branch is initialized to 1, and the weight W2 of the Conv1x1 layer corresponding to the Fourier transform branch is initialized to 0;

[0039] Initially, the weight of the Conv1x1 layer corresponding to the Fourier transform branch is 0. During the training process, the gradients of W1 and W2 are calculated using the back propagation algorithm, and W1 and W2 are updated using the optimizer; the weight of the Conv1x1 layer corresponding to the Fourier transform branch gradually changes, so that the Fourier transform branch affects the output;

[0040] After the Conv1x1 layer, the features of the backbone feature branch and the Fourier transform branch are fused to obtain a fused feature tensor, which is fed into the decoder. The decoder applies a series of convolutional layers and prediction heads to the fused feature tensor. The prediction head outputs the target detection result, which includes the target bounding box and category probability of the target object.

[0041] Extract comprehensive information of each target object from the target detection results, including position information, speed information and threat level; the position information is the coordinate position of the target object and the size of the target bounding box; by tracking the displacement of the target object in the continuous frame model input image data, the movement speed is calculated, which is the speed information;

[0042] According to the category probability and speed information of the target object, its potential threat to the drone is evaluated and a corresponding threat level value is assigned, which is the threat level.

[0043] Furthermore, the method of using the perception-integrated attack algorithm to solve the optimal attack and defense strategy includes:

[0044] Define a series of attack and defense behaviors that the drone can take to form a decision space; each attack and defense behavior is parameterized by one or several continuous variables;

[0045] According to the type of target object, the kinematic model of the target object is defined as a constant velocity model, a constant acceleration model or a maneuvering target model; the parameters of the kinematic model of the target object include position and velocity, and the initial values ​​of the parameters are obtained from the target detection results; the parameters of the kinematic model are updated online using the Kalman filter algorithm; the kinematic model of the UAV is defined as a 6-DOF rigid body motion model;

[0046] Combine the state data of the drone itself and the comprehensive information of the target object into a state vector; define the control input of the drone to include thrust and control surface deflection; combine the control input of the drone into a control vector;

[0047] The state transfer equation is defined as: x(T+1)=f(x(T),u(T)), where x(T+1) is the state vector at time T+1; x(T) is the state vector at time T, u(T) represents the control vector of the UAV at time T; f() is a nonlinear function obtained from the kinematic model of the target object and the UAV;

[0048] The reward function is defined using the state transition equation as a constraint; the formula of the reward function is:

[0049] R(s,a)=v1×Rki(s,a)+v2×Rsu(s,a)+v3×Rdi(s,a)+v4×Rti(s,a);

[0050] Among them, v1, v2, v3 and v4 are the weight balance coefficients of the corresponding items; R(s,a) is the benefit of the control vector a taken by the drone when the state vector is s; Rki(s,a) is the threat elimination function; Rsu(s,a) is the survival function; Rdi(s,a) is the distance penalty function; Rti(s,a) is the time penalty function;

[0051] All state vectors constitute the state space, and the state space, decision space, state transfer equation and reward function are constructed as a reinforcement learning problem model; the optimal attack and defense strategy is obtained based on the reinforcement learning problem model.

[0052] Furthermore, the threat elimination function Rki(s,a)=wki×th(s,a); where wki is the weight coefficient for eliminating the threat; th(s,a) represents the threat level posed by the target object to the drone when the state vector is the control vector a taken by the drone under the condition of s;

[0053] Self-survival function Among them, wda is the weight coefficient of survival penalty; dam(s,a) is the damage value when the drone takes control vector a when the state vector is s; hea(s) is the health status of the drone when the state vector is s; τ1 and τ2 are adjustable parameters;

[0054] Distance Penalty Function Among them, dis(s,a) is the distance between the drone and the target object when the state vector is s and the control vector is a; wea is the effective range of the drone's weapon; σ1 is the rate adjustment parameter, and δ1 is the degree adjustment parameter;

[0055] Time penalty function Among them, tmax is the maximum time step that the UAV can execute, that is, the time limit of the task; t1 is the time step that the UAV has executed; is the time penalty exponent parameter.

[0056] Furthermore, the method of solving the optimal attack and defense strategy based on the reinforcement learning problem model includes:

[0057] A decision strategy network is constructed based on the reinforcement learning problem model. The decision strategy network consists of an actor network and a critic network, both of which use a deep neural network structure. The input of the actor network is a state vector, and the output is a control vector. The input of the critic network is a state vector and a control vector, and the output is a benefit. The network structure of the actor network and the critic network is a fully connected network or a convolutional network. A deep reinforcement learning algorithm is used to optimize the decision strategy network to obtain a trained decision strategy network.

[0058] Obtain the current state data of the drone. For each target object, the current state data of the drone and the comprehensive information of the target object form a real-time state vector. The real-time state vector is used as input. Through the trained decision strategy network, a series of continuous behavior vectors are output. The attack and defense behaviors and their parameters are decoded as the preliminary attack and defense strategy.

[0059] Continuous behavior vectors constitute discrete behavior sequences; the discrete behavior sequences are convolved with Gaussian kernels to obtain smooth behavior sequences; the smooth behavior sequences are decoded to obtain the optimal attack and defense strategy.

[0060] The technical effects and advantages of the unmanned aerial vehicle intelligent defense and attack control system of the present invention are as follows:

[0061] The present invention can obtain environmental situation information in an all-round and multi-dimensional manner, which not only improves the ability to understand complex environments, but also can effectively identify hidden targets, greatly reducing the probability of missed detection and false detection; in terms of target assessment, by comprehensively analyzing the target's motion characteristics, threat level and attack intention, a more scientific and accurate threat assessment system is established, providing a reliable basis for subsequent decision-making; in terms of attack and defense decision-making, it breaks through the limitations of traditional fixed strategies, and can dynamically generate the optimal attack and defense strategy according to the real-time changes in the environmental situation, significantly improving the survivability and strike effectiveness of the UAV in a confrontational environment; it can independently weigh multiple key factors such as attack timing, threat avoidance and energy consumption to ensure efficient completion of tasks; in addition, it also has the ability of autonomous learning and experience accumulation, can continuously optimize and improve decision-making models, and continuously improve combat effectiveness; in practical applications, it significantly reduces the burden on operators, improves the coordinated combat capability of UAVs, and lays an important foundation for building an intelligent and networked UAV system. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 A schematic diagram of an intelligent defense and attack control system for unmanned aerial vehicles according to the present invention;

[0063] Figure 2 The diagram is a schematic diagram of a security authentication system based on a secure computer according to the present invention. DETAILED DESCRIPTION

[0064] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] Example 1

[0066] See also Figure 1 As shown, the drone intelligent defense and attack control system described in this embodiment includes:

[0067] A data acquisition module is used to acquire the environmental perception data of the UAV, which includes the state data of the UAV itself and the image data of the surrounding environment;

[0068] A data processing module is used to pre-process the image data of the surrounding environment to obtain model input image data;

[0069] The target detection module is used to perform target detection on the model input image data using a variant Poolformer algorithm to obtain comprehensive information about the target object;

[0070] The strategy fitting module is used to solve the optimal attack and defense strategy based on the UAV's own status data and the comprehensive information of the target object using the perception-integrated attack algorithm; the optimal attack and defense strategy is sent to the UAV execution terminal to control the UAV to execute the optimal attack and defense strategy; each module is connected by wired and / or wireless means to realize data transmission between modules.

[0071] The drone's own status data includes position data, motion status, energy status, weapon status, health management status and environmental status; the position data is the three-dimensional spatial position coordinates of the drone (longitude, latitude, altitude), the motion status includes the drone's velocity vector (horizontal speed, vertical speed), the drone's attitude angle (yaw angle, pitch angle, roll angle) and the drone's acceleration vector; the energy status includes the remaining fuel / battery, engine / motor power output and energy consumption rate.

[0072] The weapon status is the remaining quantity and status of various weapons; for each weapon, a status indicator is defined to represent its working status; for example, discrete values ​​or continuous values ​​are used for quantification, 0 represents complete failure, 1 represents slight failure, and 2 represents normal operation; or a real number between 0 and 1, 0 represents complete failure, 1 represents intact, and intermediate values ​​represent different degrees of failure or aging.

[0073] The health management status refers to the health status of various subsystems (aircraft, avionics, control, etc.) (also quantified as discrete or continuous values); the environmental status includes temperature, air pressure, wind speed, visibility, electromagnetic interference intensity and obstacle information; obstacle information includes the three-dimensional spatial position coordinates of the obstacle, obstacle size and the motion trajectory of dynamic obstacles; these status data are acquired and integrated through the drone's own sensors.

[0074] The methods for obtaining the motion trajectory of dynamic obstacles include:

[0075] Dynamic obstacles are regarded as moving particles. UAVs and other obstacles generate virtual gravitational fields and repulsive fields on the particles. The particles move under the combined force of the gravitational field and the repulsive field. The UAV is defined as a gravitational source. The size of the gravitational field is inversely proportional to the distance between the UAV and the dynamic obstacle. The direction of the gravitational field points to the location of the UAV. Other obstacles (fixed or dynamic) are defined as repulsive sources. The size of the repulsive field is inversely proportional to the distance between the repulsive source and the corresponding particle, and the direction of the repulsive field points away from the repulsive source. For each time step, the resultant force acting on the particle is calculated, and the resultant force is equal to the sum of the gravitational field plus the sum of the repulsive field. The resultant force is substituted into Newton's equation of motion, and the acceleration a is obtained by solving it, and the velocity and displacement of the particle are obtained by numerical integration. A time range is preset, and the changes in the velocity and displacement of the particle within the time range are constructed to obtain a three-dimensional position sequence of the UAV, which is used to represent the motion trajectory of the dynamic obstacle.

[0076] Methods for preprocessing image data of the surrounding environment include:

[0077] The original image data is processed by attention fusion and equalization to obtain preliminary input image data; the preliminary input image data is cropped to obtain the key area image; the pixel values ​​in the key area image are scaled to the range of 0-1 or -1 to 1 (using methods such as minimum-maximum normalization or mean-variance normalization) to obtain the model input image data.

[0078] The methods for attention fusion and balance processing include:

[0079] Define an attention structure based on a deep convolutional neural network to learn the importance weights of different areas in the image; the input of the attention structure is the original image data, and the output is an attention weight map of the same size as the original image data; each pixel value in the attention weight map represents the importance of the corresponding position, and the larger the value, the more important it is.

[0080] An image dataset is formed by collecting images under various scenes and conditions, wherein the images contain target objects of interest (such as people, vehicles, airplanes, etc.); the images in the image dataset are manually annotated, and the positions and categories of the target objects are marked with bounding boxes on the images; the attention structure is integrated into the classification network or the target detection network to obtain a comprehensive network model; in the feature extraction stage of the classification network or the target detection network, the attention structure outputs the attention weight map as a mask and multiplies it with the original image data to obtain a feature map; the original image data is subjected to evolutionary histogram equalization to obtain a globally enhanced image; the feature map is multiplied with the globally enhanced image to obtain preliminary input image data.

[0081] The training process of the comprehensive network model includes:

[0082] Define the loss function, use the image data set to train the comprehensive network structure, calculate the gradient of the attention structure according to the value of the loss function of the comprehensive network model, and update its parameters. After training, the attention structure automatically focuses on the area where the target object is located.

[0083] Ways to perform evolutionary histogram equalization include:

[0084] The original image data is divided into several overlapping small blocks, usually using a sliding window method; the size of each small block is pre-set, such as 64x64 or 128x128 pixels; a certain overlapping area is set between adjacent small blocks to ensure smooth boundary processing.

[0085] Each small block is subjected to histogram equalization processing to obtain a reconstructed small block. Specifically, each small block is converted into a corresponding grayscale image. For each grayscale image, the number of pixels at each grayscale level (0-255) is counted and a grayscale histogram is constructed. The grayscale histogram reflects the distribution of pixels of different grayscale levels in the image.

[0086] Based on the grayscale histogram, calculate the cumulative distribution function value CDF(i) of each grayscale level i;

[0087] Among them, n(0) represents the number of pixels with gray level 0, n(1) represents the number of pixels with gray level 1, and so on, n(i) is the number of pixels with gray level i; N is the total number of pixels in the grayscale image; for each gray level i, calculate its new mapping gray level j = (L-1) × CDF(i) according to its cumulative distribution function value CDF(i); where L is the total number of gray levels, usually 256; traverse each pixel in the small block area, map its gray level i to the new mapping gray level j; obtain the remapped small block area; record it as the reconstructed small block area.

[0088] For each reconstructed small block, calculate the normalized coordinates of its center point in the original image data; record them as the center point coordinates; based on the center point coordinates, calculate the Euclidean distance d from the reconstructed small block to the center of the original image data; for each reconstructed small block, count the number of pixels overlapping with other reconstructed small blocks; for each reconstructed small block, calculate its grayscale co-occurrence matrix (the grayscale co-occurrence matrix reflects the spatial distribution characteristics between pixels of different grayscale values ​​in the image); based on the comprehensive features in the grayscale co-occurrence matrix; the comprehensive features include contrast DA, energy DH (which is the mean square value, reflecting the uniformity of grayscale distribution in the image), homogeneity DF and entropy DS.

[0089] Contrast ratio DA = ∑ I ∑ J(IJ) γ ×P(I,J) δ ; Where I and J are the row and column indices of the grayscale co-occurrence matrix, corresponding to different grayscale values; P(I,J) is an element in the grayscale co-occurrence matrix, indicating the frequency of occurrence of pixel pairs between grayscale values ​​I and J; γ is an exponential factor that controls the degree of influence of grayscale difference on contrast, and δ is an exponential factor that controls the degree of influence of pixel pair frequency on contrast.

[0090] Homogeneity Where a and b are parameters that control the shape of the function; entropy DS = ∑ I ∑ J P(I,J)×log(P(I,J));

[0091] The comprehensive features are normalized and weightedly fused to obtain the comprehensive texture complexity index of the corresponding reconstructed small block area; for each reconstructed small block area, edge detection operators (such as Sobel, Canny, etc.) are applied to extract the number of edge pixels and the cumulative gradient amplitude, and weighted fusion is performed to obtain the edge strength index of the reconstructed small block area; the comprehensive texture complexity index and the edge strength index are fused into a comprehensive feature index C; specifically, a simple weighted summation or nonlinear fusion function is used, such as a multi-layer perceptron (MLP).

[0092] Define the position weight of the reconstructed small area h Among them, σ is a parameter that controls the weight decay speed; ω1 is an influencing parameter that adjusts the influence of the comprehensive feature index C on the weight; ω2 is an exponential factor that controls the distance, which is used to control the influence of the distance on the weight.

[0093] Define the overlap weight of the reconstructed small area h Among them, a1 is the scaling factor, which is used to adjust the magnitude of the overlapping weight; No is the number of pixels that overlap between the reconstructed small block area h and other reconstructed small blocks, and Nb is the total number of pixels in the reconstructed small block area h; a2 is the overlapping index factor, which is used to control the influence of the number of overlapping pixels on the weight; γ2 is a control parameter used to adjust the influence of the comprehensive feature index on the weight.

[0094] The position weight and overlap weight of the reconstructed small block area h are multiplied to obtain the comprehensive weight coefficient of the corresponding reconstructed small block area; for the overlapping area in the reconstructed small block area, the pixel values ​​therein are weighted averaged according to the comprehensive weight coefficient of the corresponding reconstructed small block area to obtain the fused pixel value; for the non-overlapping area in the reconstructed small block area, the corresponding pixel value is directly used to obtain a complete global enhanced image.

[0095] The methods of cropping the preliminary input image data include:

[0096] Based on the motion trajectory of the dynamic obstacle, the field of view volume of each time step within the time range is determined. For each time step t on the motion trajectory, the corresponding field of view is calculated according to the position and attitude angle of the UAV at the time step t, and a series of time-varying three-dimensional field of view volumes V(t) are obtained. Specifically, a camera coordinate system is established with the position of the UAV as the origin, and the directions of the three coordinate axes of the camera coordinate system are determined according to the attitude angle of the UAV. According to the field of view angle (calculated by the intrinsic parameters of the camera), the coordinates of the eight vertices of the camera cone are calculated in the camera coordinate system. Using the position and attitude angle of the UAV, the vertex coordinates in the camera coordinate system are converted to the world coordinate system to obtain the coordinates of the eight vertices in the world coordinate system, and a three-dimensional cone or pyramid is constructed using the coordinates of the eight vertices in the world coordinate system. The three-dimensional cone or pyramid is the field of view of the UAV at time step t. Then a series of time-varying three-dimensional field of view volumes V(t) are obtained.

[0097] For each time step t, the corresponding three-dimensional field of view volume V(t) is projected onto the two-dimensional image plane to obtain a series of two-dimensional polygonal areas P(t), which represent the image areas that the image acquisition equipment of the UAV can observe at each time step; a mask image M of the same size as the globally enhanced image is initialized, and all pixel values ​​in the mask image M are initially 0; for each time step t, the pixel values ​​in the corresponding two-dimensional polygonal area P(t) are set to 1 on the mask image M, and the final mask image M contains the image area that the UAV can observe within the entire time range; the globally enhanced image is multiplied pixel by pixel by pixel by the mask image M to obtain the key area image.

[0098] Methods for performing target detection on model input image data include:

[0099] Construct a variant Poolformer model. The basic structure of the variant Poolformer model includes an encoder and a decoder. The encoder is defined to include a trunk feature branch and a parallel Fourier transform branch. The trunk feature branch uses a multi-head self-attention mechanism to extract features from the model input image data, and the decoder uses a pooled attention mechanism to aggregate global context information.

[0100] The Fourier transform branch performs a two-dimensional discrete Fourier transform (2D-DFT) on the model input image data, converting it from the spatial domain to the frequency domain, and performing low-pass filtering and high-pass filtering on the model input image data in the frequency domain; a Conv1x1 convolutional layer is inserted between the Fourier transform branch and the trunk feature branch; the weight W1 of the Conv1x1 layer corresponding to the trunk feature branch is initialized to 1, and the weight W2 of the Conv1x1 layer corresponding to the Fourier transform branch is initialized to 0.

[0101] Initially, since the weight of the Conv1x1 layer corresponding to the Fourier transform branch is 0, the Fourier transform branch has no contribution to the output; during the training process, the back propagation algorithm is used to calculate the gradients of W1 and W2, and the optimizer (such as SGD, Adam, etc.) is used to update W1 and W2; the weight of the Conv1x1 layer corresponding to the Fourier transform branch gradually changes, so that the Fourier transform branch affects the output.

[0102] After the Conv1x1 layer, the features of the backbone feature branch and the Fourier transform branch are fused to obtain a fused feature tensor, which is sent to the decoder. The decoder applies a series of convolutional layers and prediction heads to the fused feature tensor; the prediction head outputs the target detection result, which includes the target bounding box and category probability of the target object.

[0103] The training process is as follows:

[0104] Collect and prepare training datasets, including images in various scenes and conditions, and annotate target objects (people, vehicles, airplanes, etc.) in the images to obtain bounding boxes and category labels of target objects; you can use public datasets such as COCO, VOC, etc., or use custom datasets; divide the training dataset into training set and validation set; initialize other trainable parameters of the variant Poolformer model (such as convolution kernel weights, batch normalization parameters, etc.), and use initialization methods such as Xavier or Kaiming; define loss functions, including classification loss (such as cross entropy loss) and regression loss (such as smooth L1 loss), for supervised target detection tasks; use small batch stochastic gradient The training is performed in a descent manner. In each training iteration, a batch of image data and corresponding labels are sampled from the training dataset, and the image data is input into the variant Poolformer model to obtain the predicted bounding box and category probability, the value of the calculated loss function, and the gradient of the trainable parameters (including the Conv1x1 layer weight) is calculated using the back-propagation algorithm. The trainable parameters are updated according to the gradient using an optimizer (such as SGD, Adam, etc.), and the model performance is periodically evaluated on the validation set, such as mAP (average precision) and other indicators; according to the model performance on the validation set, the hyperparameters, such as the learning rate and regularization strength, are adjusted to obtain better performance; and the variant Poolformer model with the best model performance is saved.

[0105] The comprehensive information of each target object is extracted from the target detection results, and the comprehensive information includes position information, speed information and threat level; the position information is the coordinate position of the target object (the center coordinate of the target bounding box) and the size of the target bounding box; the movement speed is calculated by tracking the displacement of the target object in the continuous frame model input image data, which is the speed information; according to the category probability and speed information of the target object, its potential threat level to the drone is evaluated, and the corresponding threat level value is assigned, which is the threat level; specifically, for different categories of target objects, their potential threat level values ​​to the drone are pre-defined; for example, the threat level of pedestrians is low, while the threat level of equipment is high; stationary target objects pose a smaller threat, while high-speed moving target objects pose a greater threat, and the potential threat level value is adjusted according to the speed information of the target object to obtain the threat level.

[0106] Ways to use the perception-integrated attack algorithm to solve the optimal attack and defense strategy include:

[0107] Define a series of attack and defense behaviors that the UAV can take to form a decision space; for example, acceleration, deceleration, change of course, firing weapons, evasion, etc.; each attack and defense behavior is parameterized by one or several continuous variables, such as acceleration, heading angle, etc.; according to the type of target object, define the kinematic model of the target object as a constant velocity model, a constant acceleration model or a maneuvering target model; the parameters of the kinematic model of the target object include position and velocity, and the initial values ​​of the parameters are obtained from the target detection results; use the Kalman filter algorithm to update the parameters of the kinematic model online; define the kinematic model of the UAV as a 6-degree-of-freedom rigid body motion model.

[0108] Combine the state data of the drone itself and the comprehensive information of the target object into a state vector; define the control input of the drone to include thrust and control surface deflection; combine the control input of the drone into a control vector; then define the state transfer equation as: x(T+1)=f(x(T),u(T)), where x(T+1) is the state vector at time T+1; x(T) is the state vector at time T, and u(T) represents the control vector (attack and defense behavior) of the drone at time T; f() is a nonlinear function obtained from the kinematic model of the target object and the drone; specifically, based on the 6-DOF rigid body motion model, list the state data of the drone itself and its rate of change The differential equations between them, for example, the rate of change of position is equal to the corresponding velocity; the rate of change of velocity is equal to the function of the corresponding acceleration and angular velocity of rotation; the rate of change of attitude angle is equal to the nonlinear function of the corresponding angular velocity; the thrust determines the acceleration, and the control surface deflection determines the angular acceleration. The control input is substituted into the function of the corresponding acceleration and angular velocity of rotation, that is, the control input is coupled with the state vector; according to the kinematic model of the target object, the differential equations of its position and velocity are established; for example, for the constant velocity model, the rate of change of the target position is equal to the velocity of the target object; the above differential equations are discretized to obtain their differential form, and all the discretized differential equations are combined to obtain f().

[0109] The state transition equation is used as a constraint to define the reward function, which quantifies the benefits of the control vector (attack and defense behavior); the formula of the reward function is:

[0110] R(s,a)=v1×Rki(s,a)+v2×Rsu(s,a)+v3×Rdi(s,a)+v4×Rti(s,a);

[0111] Among them, v1, v2, v3 and v4 are the weight balance coefficients of the corresponding items; R(s,a) is the benefit of the control vector a taken by the drone when the state vector is s; Rki(s,a) is the threat elimination function; Rsu(s,a) is its own survival function; Rdi(s,a) is the distance penalty function; Rti(s,a) is the time penalty function.

[0112] Threat elimination function Rki(s,a)=wki×th(s,a); where wki is the weight coefficient for eliminating the threat; th(s,a) represents the threat level posed by the target object to the drone when the state vector is the control vector a taken by the drone under the state vector s; the threat level is a discrete value (such as 1-5) or a continuous value (such as between 0-1).

[0113] Self-survival function Among them, wda is the weight coefficient of the survival penalty, which is used to adjust the penalty size; dam(s,a) is the damage value when the UAV takes the control vector a when the state vector is s; hea(s) is the health status of the UAV when the state vector is s, which is the comprehensive value of the remaining fuel and structural integrity; τ1 and τ2 are adjustable parameters.

[0114] Distance Penalty Function Among them, dis(s,a) is the distance between the drone and the target object when the state vector is s and the control vector is a; wea is the effective range of the drone's weapon; σ1 is the rate adjustment parameter, which controls the rate of increase of the distance; δ1 is the degree adjustment parameter, which controls the degree of increase of the distance.

[0115] Time penalty function Among them, tmax is the maximum time step that the UAV can execute, that is, the time limit of the task; t1 is the time step that the UAV has executed; is the time penalty exponential parameter; as the execution time approaches the time limit, the time penalty increases rapidly. Controls the rate of increase.

[0116] All state vectors constitute the state space, and the state space, decision space, state transfer equation and reward function are constructed as a reinforcement learning problem model; a decision policy network is constructed based on the reinforcement learning problem model. The decision policy network consists of an Actor network and a Critic network, and usually adopts a deep neural network structure; the input of the Actor network is a state vector, and the output is a control vector; the input of the Critic network is a state vector and a control vector, and the output is a reward; the network structure of the Actor network and the Critic network is a fully connected network, a convolutional network or other types, and needs to be designed according to the representation of the state and behavior; a deep reinforcement learning algorithm (such as deep Q network (DQN), policy gradient, etc.) is used to optimize the decision policy network to obtain a trained decision policy network.

[0117] The current state data of the drone is obtained. For each target object, the current state data of the drone and the comprehensive information of the target object are used to form a real-time state vector. The real-time state vector is used as input. Through the trained decision strategy network, a series of continuous behavior vectors are output. The attack and defense behaviors and their parameters are decoded as the preliminary attack and defense strategies. Specifically, the real-time state vector is input into the trained Actor network. The Actor network will output the corresponding optimal behavior according to the real-time state vector. The optimal behavior is decoded to obtain the attack and defense behaviors and their parameters.

[0118] Continuous behavior vectors constitute discrete behavior sequences, which may undergo drastic changes. For example, between adjacent time steps, the acceleration or heading angle changes too much. Such drastic changes may cause the drone to be unable to actually execute or to execute inefficiently. The discrete behavior sequence is convolved with a Gaussian kernel to obtain a smooth behavior sequence. The smooth behavior sequence is decoded to obtain the optimal attack and defense strategy.

[0119] It should be noted that the output of the Actor network is usually a continuous behavior vector, which needs to be decoded into specific attack and defense behaviors and their parameters. For example, the different components of the behavior vector are decoded into acceleration, heading angle change rate, weapon launch parameters, etc.

[0120] This embodiment can obtain environmental situation information in an all-round and multi-dimensional manner, which not only improves the ability to understand complex environments, but also can effectively identify hidden targets, greatly reducing the probability of missed detection and false detection; in terms of target assessment, by comprehensively analyzing the target's motion characteristics, threat level and attack intention, a more scientific and accurate threat assessment system is established, providing a reliable basis for subsequent decision-making; in terms of attack and defense decision-making, it breaks through the limitations of traditional fixed strategies, and can dynamically generate the optimal attack and defense strategy according to the real-time changes in the environmental situation, significantly improving the survivability and strike effectiveness of the UAV in a confrontational environment; it can autonomously weigh multiple key factors such as attack timing, threat avoidance and energy consumption to ensure efficient completion of the task; in addition, it also has the ability of autonomous learning and experience accumulation, can continuously optimize and improve the decision-making model, and continuously improve combat effectiveness; in actual applications, it significantly reduces the burden on operators, improves the coordinated combat capability of UAVs, and lays an important foundation for building an intelligent and networked UAV system.

[0121] Example 2

[0122] See also Figure 2 As shown, the part not described in detail in this embodiment is described in Example 1, which provides a method for intelligent defense and attack control of unmanned aerial vehicles, including:

[0123] Step 1: Obtain the environmental perception data of the drone, which includes the drone's own status data and image data of the surrounding environment;

[0124] Step 2: preprocess the image data of the surrounding environment to obtain model input image data;

[0125] Step 3: Use the variant Poolformer algorithm to perform target detection on the model input image data to obtain comprehensive information of the target object;

[0126] Step 4: Based on the drone's own status data and the comprehensive information of the target object, the optimal attack and defense strategy is solved using the perception-integrated attack algorithm; the optimal attack and defense strategy is sent to the drone execution terminal to control the drone to execute the optimal attack and defense strategy.

[0127] Example 3

[0128] This embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the operation mode of the above-mentioned drone intelligent defense and attack control method is implemented.

[0129] Since the electronic device introduced in this embodiment is an electronic device used to implement a drone intelligent defense and attack control method in the embodiment of this application, based on the drone intelligent defense and attack control method introduced in the embodiment of this application, the technical personnel of this field can understand the specific implementation of the electronic device of this embodiment and its various variations, so how the electronic device implements the method in the embodiment of this application is not described in detail here. As long as the technical personnel of this field implement the electronic device used in the drone intelligent defense and attack control method in the embodiment of this application, it belongs to the scope of protection of this application.

[0130] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0131] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technical users in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. An intelligent defense and attack control system for unmanned aerial vehicles, characterized in that: include: A data acquisition module is used to acquire the environmental perception data of the UAV, which includes the state data of the UAV itself and the image data of the surrounding environment; A data processing module is used to pre-process the image data of the surrounding environment to obtain model input image data; The target detection module is used to perform target detection on the model input image data using a variant Poolformer algorithm to obtain comprehensive information about the target object; The strategy fitting module is used to solve the optimal attack and defense strategy based on the drone's own status data and the comprehensive information of the target object using the perception-integrated attack algorithm; the optimal attack and defense strategy is sent to the drone execution terminal to control the drone to execute the optimal attack and defense strategy; each module is connected by wired and / or wireless means.

2. The UAV intelligent defense and attack control system according to claim 1 is characterized in that: The state data of the drone itself includes location data, motion state, energy state, weapon state, health management state and environment state; the environment state includes temperature, air pressure, wind speed, visibility, electromagnetic interference intensity and obstacle information; The obstacle information includes the three-dimensional spatial position coordinates of the obstacle, the obstacle size, and the motion trajectory of the dynamic obstacle.

3. The UAV intelligent defense and attack control system according to claim 2 is characterized in that: The method for obtaining the motion trajectory of the dynamic obstacle includes: Dynamic obstacles are regarded as moving particles. Drones and other obstacles generate virtual gravitational and repulsive fields on the particles. The particles move under the combined force of the gravitational and repulsive fields. The drone is defined as a gravitational source; the magnitude of the gravitational field is inversely proportional to the distance between the drone and the dynamic obstacle; the direction of the gravitational field points to the location of the drone; The other obstacles are defined as repulsive sources. The magnitude of the repulsive field is inversely proportional to the distance between the repulsive source and the corresponding particle, and the direction of the repulsive field points away from the repulsive source. For each time step, the resultant force acting on the particle is calculated, which is equal to the sum of the gravitational field plus the sum of the repulsive field; the resultant force is substituted into Newton's equation of motion, the acceleration a is obtained, and numerical integration is performed to obtain the velocity and displacement of the particle. A time range is preset, and the changes in the velocity and displacement of the particle within the time range are constructed to obtain the three-dimensional position sequence of the drone, which is the motion trajectory of the dynamic obstacle.

4. The UAV intelligent defense and attack control system according to claim 3 is characterized in that: The method of preprocessing the image data of the surrounding environment includes: Perform attention fusion equalization processing on the original image data to obtain preliminary input image data; perform image cropping on the preliminary input image data to obtain a key area image; scale the pixel values ​​in the key area image to a range of 0-1 or -1 to 1 to obtain model input image data; The method of performing attention fusion and equalization processing includes: Define an attention structure based on a deep convolutional neural network; the input of the attention structure is the original image data, and the output is an attention weight map of the same size as the original image data; each pixel value in the attention weight map represents the importance of the corresponding position; Collect images in various scenes and conditions to form an image dataset, where the images contain target objects; manually annotate the images in the image dataset, and mark the location and category of the target objects with bounding boxes on the images; The attention structure is integrated into the classification network or the target detection network to obtain a comprehensive network model; in the feature extraction stage of the classification network or the target detection network, the attention structure multiplies the output attention weight map as a mask with the original image data to obtain a feature map; Perform evolutionary histogram equalization on the original image data to obtain a globally enhanced image; multiply the feature map and the globally enhanced image to obtain preliminary input image data.

5. The UAV intelligent defense and attack control system according to claim 4 is characterized in that: The method of performing evolutionary histogram equalization includes: The original image data is divided into a number of overlapping small blocks, the size of each small block is preset; a certain overlapping area is set between adjacent small blocks; each small block is subjected to histogram equalization processing to obtain a reconstructed small block; For each reconstructed small block, calculate the normalized coordinates of its center point in the original image data; record them as the center point coordinates; calculate the Euclidean distance d from the center of the reconstructed small block to the center of the original image data based on the center point coordinates; for each reconstructed small block, count the number of pixels overlapping with other reconstructed small blocks; for each reconstructed small block, calculate its grayscale co-occurrence matrix; based on the comprehensive features in the grayscale co-occurrence matrix; the comprehensive features include contrast DA, energy DH, homogeneity DF and entropy DS; Contrast ratio DA = ∑ I ∑ J (IJ) γ ×P(I,J) δ ; Where I and J are the row and column indices of the grayscale co-occurrence matrix; P(I,J) is an element in the grayscale co-occurrence matrix, which indicates the frequency of occurrence of pixel pairs between grayscale values ​​I and J; γ is an exponential factor that controls the degree of influence of grayscale difference on contrast, and δ is an exponential factor that controls the degree of influence of the frequency of occurrence of pixel pairs on contrast; Homogeneity Where a and b are parameters that control the shape of the function; entropy DS = ∑ I ∑ J P(I,J)×log(P(I,J)); The comprehensive features are normalized and weightedly fused to obtain the comprehensive texture complexity index of the corresponding reconstructed small block area; for each reconstructed small block area, the edge detection operator is applied to extract the number of edge pixels and the cumulative gradient amplitude, and weighted fusion is performed to obtain the edge strength index of the reconstructed small block area; the comprehensive texture complexity index and the edge strength index are fused into a comprehensive feature index C; Define the position weight of the reconstructed small area h Among them, σ is the parameter that controls the weight decay speed; ω1 is the influence parameter; ω2 is the exponential factor that controls the distance, which is used to control the influence of distance on the weight; Define the overlap weight of the reconstructed small area h Where a1 is the scaling factor; No is the number of pixels that overlap between the reconstructed small block h and other reconstructed small blocks, Nb is the total number of pixels in the reconstructed small block h; a2 is the overlap index factor; γ2 is a control parameter; the position weight and overlap weight of the reconstructed small block h are multiplied to obtain the comprehensive weight coefficient of the corresponding reconstructed small block; For the overlapping areas in the reconstructed small blocks, the pixel values ​​therein are weighted averaged according to the comprehensive weight coefficients of the corresponding reconstructed small blocks to obtain the fused pixel values; for the non-overlapping areas in the reconstructed small blocks, the corresponding pixel values ​​are directly used to obtain a complete globally enhanced image.

6. The UAV intelligent defense and attack control system according to claim 5 is characterized in that: The method of performing image cropping on the preliminary input image data includes: Based on the motion trajectory of the dynamic obstacle, the field of view volume of each time step within the time range is determined. For each time step t on the motion trajectory, the camera coordinate system is established with the position of the drone as the origin. According to the attitude angle of the drone, the directions of the three coordinate axes of the camera coordinate system are determined; according to the field of view angle, the coordinates of the eight vertices of the camera cone in the camera coordinate system are calculated; using the position and attitude angle of the drone, the vertex coordinates in the camera coordinate system are converted to the world coordinate system, and the coordinates of the eight vertices in the world coordinate system are obtained. Using the coordinates of the eight vertices in the world coordinate system, a three-dimensional cone or pyramid is constructed; the three-dimensional cone or pyramid is the field of view of the drone at time step t; and then a series of time-varying three-dimensional field of view volumes V(t) are obtained; For each time step t, the corresponding three-dimensional field of view volume V(t) is projected onto the two-dimensional image plane to obtain a series of two-dimensional polygonal regions P(t); Initialize a mask image M of the same size as the global enhanced image, and all pixel values ​​in the mask image M are initially 0; for each time step t, set the pixel values ​​in the corresponding two-dimensional polygonal area P(t) on the mask image M to 1; multiply the global enhanced image and the mask image M pixel by pixel to obtain the key area image.

7. The UAV intelligent defense and attack control system according to claim 6 is characterized in that: The method of performing target detection on the model input image data includes: Construct a variant Poolformer model. The basic structure of the variant Poolformer model includes an encoder and a decoder. Define the encoder to include a trunk feature branch and a parallel Fourier transform branch. The Fourier transform branch performs a two-dimensional discrete Fourier transform on the model input image data to convert it from the spatial domain to the frequency domain. In the frequency domain, the model input image data is low-pass filtered and high-pass filtered. Insert a Conv1 x1 convolution layer between the Fourier transform branch and the trunk feature branch; the weight W1 of the Conv1 x1 layer corresponding to the trunk feature branch is initialized to 1, and the weight W2 of the Conv1 x1 layer corresponding to the Fourier transform branch is initialized to 0; Initially, the weight of the Conv1 x1 layer corresponding to the Fourier transform branch is 0. During the training process, the back propagation algorithm is used to calculate the gradients of W1 and W2, and the optimizer is used to update W1 and W2; the weight of the Conv1 x1 layer corresponding to the Fourier transform branch gradually changes, so that the Fourier transform branch affects the output; After the Conv1 x1 layer, the features of the backbone feature branch and the Fourier transform branch are fused to obtain a fused feature tensor, which is fed into the decoder. The decoder applies a series of convolutional layers and prediction heads to the fused feature tensor. The prediction head outputs the target detection result, which includes the target bounding box and category probability of the target object. Extract comprehensive information of each target object from the target detection results, including position information, speed information and threat level; the position information is the coordinate position of the target object and the size of the target bounding box; by tracking the displacement of the target object in the continuous frame model input image data, the movement speed is calculated, which is the speed information; According to the category probability and speed information of the target object, its potential threat to the drone is evaluated and a corresponding threat level value is assigned, which is the threat level.

8. The UAV intelligent defense and attack control system according to claim 7 is characterized in that: The method of using the perception-integrated attack algorithm to solve the optimal attack and defense strategy includes: Define a series of attack and defense behaviors that the drone can take to form a decision space; each attack and defense behavior is parameterized by one or several continuous variables; According to the type of target object, the kinematic model of the target object is defined as a constant velocity model, a constant acceleration model or a maneuvering target model; the parameters of the kinematic model of the target object include position and velocity, and the initial values ​​of the parameters are obtained from the target detection results; the parameters of the kinematic model are updated online using the Kalman filter algorithm; the kinematic model of the UAV is defined as a 6-DOF rigid body motion model; Combine the state data of the drone itself and the comprehensive information of the target object into a state vector; define the control input of the drone to include thrust and control surface deflection; combine the control input of the drone into a control vector; The state transfer equation is defined as: x(T+1)=f(x(T),u(T)), where x(T+1) is the state vector at time T+1; x(T) is the state vector at time T, u(T) represents the control vector of the UAV at time T; f() is a nonlinear function obtained from the kinematic model of the target object and the UAV; The reward function is defined using the state transition equation as a constraint; the formula of the reward function is: R(s,a)=v1×Rki(s,a)+v2×Rsu(s,a)+v3×Rdi(s,a)+v4×Rti(s,a); Among them, v1, v2, v3 and v4 are the weight balance coefficients of the corresponding items; R(s,a) is the benefit of the control vector a taken by the drone when the state vector is s; Rki(s,a) is the threat elimination function; Rsu(s,a) is the survival function; Rdi(s,a) is the distance penalty function; Rti(s,a) is the time penalty function; All state vectors constitute the state space, and the state space, decision space, state transfer equation and reward function are constructed as a reinforcement learning problem model; the optimal attack and defense strategy is obtained based on the reinforcement learning problem model.

9. The UAV intelligent defense and attack control system according to claim 8, characterized in that: Threat elimination function Rki(s,a)=wki×th(s,a); where wki is the weight coefficient for eliminating the threat; th(s,a) represents the threat level posed by the target object to the drone when the state vector is the control vector a taken by the drone under the condition of s; Self-survival function Among them, wda is the weight coefficient of survival penalty; dam(s,a) is the damage value when the drone takes control vector a when the state vector is s; hea(s) is the health status of the drone when the state vector is s; τ1 and τ2 are adjustable parameters; Distance Penalty Function Among them, dis(s,a) is the distance between the drone and the target object when the state vector is s and the control vector is a; wea is the effective range of the drone's weapon; σ1 is the rate adjustment parameter, and δ1 is the degree adjustment parameter; Time penalty function Among them, tmax is the maximum time step that the UAV can execute, that is, the time limit of the task; t1 is the time step that the UAV has executed; is the time penalty exponent parameter.

10. The UAV intelligent defense and attack control system according to claim 9, characterized in that: The method of solving the optimal attack and defense strategy based on the reinforcement learning problem model includes: A decision strategy network is constructed based on the reinforcement learning problem model. The decision strategy network consists of an actor network and a critic network, both of which use a deep neural network structure. The input of the actor network is a state vector, and the output is a control vector. The input of the critic network is a state vector and a control vector, and the output is a benefit. The network structure of the actor network and the critic network is a fully connected network or a convolutional network. A deep reinforcement learning algorithm is used to optimize the decision strategy network to obtain a trained decision strategy network. Obtain the current state data of the drone. For each target object, the current state data of the drone and the comprehensive information of the target object form a real-time state vector. The real-time state vector is used as input. Through the trained decision strategy network, a series of continuous behavior vectors are output. The attack and defense behaviors and their parameters are decoded as the preliminary attack and defense strategy. Continuous behavior vectors constitute discrete behavior sequences; the discrete behavior sequences are convolved with Gaussian kernels to obtain smooth behavior sequences; the smooth behavior sequences are decoded to obtain the optimal attack and defense strategy.

Images