Method for actively identifying interference and decoy by inspection unmanned aerial vehicle

Abstract

The invention provides a method for actively identifying interference and decoy by an inspection unmanned aerial vehicle, which comprises the following steps: acquiring real-time feedback data of each sensor of the unmanned aerial vehicle to obtain first data; constructing an unmanned aerial vehicle virtual anti-decoy training scene, and training an intelligent anti-decoy strategy model in a layered reinforcement learning training mode; quantitatively deploying the intelligent anti-decoy strategy model to an airborne computer of the unmanned aerial vehicle; and taking the preprocessed first data as the input of the intelligent anti-decoy strategy model, performing active identification of interference decoy, and outputting and executing a flight decision in real time. Diversified data samples are provided for anti-decoy training of the unmanned aerial vehicle through a virtual simulation environment; various decoy modes are simulated, and adversarial strategy training is adopted to improve the strategy precision of the model; layered reinforcement learning is adopted, effective decision learning is realized through a hierarchical decision target, and the decision ability of a decision model is improved.

Description

Technical Field

[0001] The present application relates to the field of drone applications, and in particular to a method for proactively identifying interference and deception by inspection drones. Background Art

[0002] In recent years, drone technology has been rapidly advancing toward intelligent, multi-scenario applications. Leveraging high-precision navigation, AI visual recognition, and multi-sensor fusion technologies, it has demonstrated significant advantages in infrastructure inspection. During railway inspections, drones can rapidly scan tracks, bridges, and slopes along the lines, accurately identifying cracks and subsidence risks, achieving efficiency several times higher than manual inspections. During power grid inspections, drone swarms can autonomously detect damaged insulators and broken conductors along hundreds of kilometers of high-voltage lines, mitigating significant overhead operational risks. On highways, drones use 3D modeling to monitor pavement defects in real time, while simultaneously analyzing traffic data to aid maintenance decision-making. For river inspections, drones equipped with spectrometers and thermal imaging equipment can track water pollution sources around the clock and assess dam leakage risks.

[0003] However, when drones operate in complex environments, the GPS navigation, wireless communications, and sensor systems they rely on present significant security vulnerabilities. During railway inspections, if a drone encounters GPS signal spoofing or electromagnetic interference, it could deviate from the inspection track and collide with a high-speed train or the overhead catenary, causing derailments and economic losses. During power grid inspections, directional radio frequency interference can disable the navigation of a drone swarm, leading to collisions with high-voltage towers and regional power outages. In highway inspection scenarios, attackers could hijack an aircraft and direct it toward the road, colliding with moving vehicles and causing traffic accidents. During river monitoring, malicious interference could force a drone out of control and crash into the core area of a dam, potentially damaging gate sensors and triggering malfunctions in the flood discharge system.

[0004] Current drone anti-interference and anti-spoofing technologies have formed a multi-layered defense system. At the navigation level, multi-mode fusion positioning using GPS / Beidou / inertial navigation / visual SLAM, combined with anti-spoofing algorithms, effectively protects against conventional signal spoofing. Frequency-hopping spread spectrum technology and an AI spectrum sensing system are used in communication links, significantly enhancing anti-interception capabilities.

[0005] However, defensive frequency modulation through active jamming signals cannot effectively resist various forms of jamming deception. For example, a correct deception signal can easily bypass signal detection. Therefore, in the process of executing tasks in multiple scenarios, there must still be an anti-interference system that can actively identify deception to improve the level of drone deception resistance and risk prevention. Summary of the Invention

[0006] In view of the above problems, the present invention provides a method for patrol drones to actively identify interference and deception. Reinforcement learning is used to construct a virtual simulation environment to train the drone's anti-deception strategy to cope with complex and changeable interference and deception methods, ensure the flight safety of the drone, and improve the defense strategy.

[0007] In a first aspect, the present invention provides a method for proactively identifying jamming and deception by an inspection drone, comprising the following steps:

[0008] Acquire real-time feedback data from each sensor of the drone to obtain first data; wherein the first data includes image data and point cloud data;

[0009] Construct a virtual anti-deception training scenario for drones, and use a hierarchical reinforcement learning training model to train an intelligent anti-deception strategy model; hierarchical reinforcement learning includes a signal analysis layer and a task decision-making layer.

[0010] Among them, a reinforcement learning framework is constructed for the state space with the drone's posture state, channel state and interference pattern as elements, the reward function with positioning error, communication interruption and survival metric as reward and punishment elements, and the action space with path replanning, communication relay request and switching visual SLAM assistance as elements;

[0011] Quantitatively deploy the intelligent anti-deception strategy model on the drone's onboard computer;

[0012] The pre-processed first data is used as the input of the intelligent anti-deception strategy model to actively identify interference deception, output flight decisions in real time and execute them.

[0013] Furthermore, the training process of extracting the route from the model training of the first data is as follows:

[0014] Data collection and preparation: collecting image data and laser point cloud data from optical cameras and 3D-LiDAR, and performing pre-processing operations on the data;

[0015] Data annotation: Use image and point cloud annotation tools to perform semantic segmentation and annotation on the target area;

[0016] For model training, the preprocessed data is randomly sampled and divided into training set, test set and validation set. The training set is used as the input of the image semantic segmentation model and the point cloud 3D object detection model to obtain the image semantic segmentation model and the point cloud 3D object detection model with converged training accuracy. The model is then tested and tuned using the test set.

[0017] Furthermore, in the extraction of the route, the inspection area and the waterway area are set to meet the inspection constraints, which are specifically expressed as follows:

[0018] Among them, (x′, y′)∈a′ is the two-dimensional projection coordinate of the actual UAV flight route; (x, y)∈a is the two-dimensional projection coordinate of the preset route. D is the width of the channel, d is the width of the inspection area, and δ is the distance between the route and the inspection area. Route a is expressed as: A a x+B a y+c a =0, the center line of the inspection area is represented by: A o x+B o y+c o =0.

[0019] Furthermore, the reinforcement learning model of the signal analysis layer constructs a total loss function, and the specific operations are as follows:

[0020] Strategy update constraint function: In the process of optimizing the target, in order to limit the strategy update range, the PPO clipping objective function is introduced, which is expressed as:

[0021] L CLIP (θ)=E t [min(r t (θ)A t ,clip(r t (θ),1-∈,1+∈)A t )]

[0022] A represents the probability ratio of the new and old strategies to select actions in the same state, t represents the advantage function;

[0023] Mean square error optimization of the value function: Update the value function V(s) by minimizing the mean square error t ), expressed as:

[0024] L VF (θ)=E t [(V(s t )-R t ) 2 ]

[0025] V(s t ;θ), the value function prediction of the current state; cumulative returns;

[0026] Policy entropy regularization optimization: Policy entropy regularization, expressed as: L ENT (θ)=E t [H(πθ(s t ))], which can prevent the strategy from converging to the local optimum too early;

[0027] Combining the policy loss, value function loss and entropy regularization term, the total loss function is formed:

[0028] L(θ)=E t [L CLIP (θ)-c1L VF (θ)+c2L ENT (θ)], c1, c2 represent weight coefficients to balance policy optimization, value function update and entropy regularization.

[0029] Furthermore, the signal analysis layer adopts the PPO training algorithm, and the specific operations are as follows:

[0030] Use the current policy to interact with the environment, sampling the current value s in the state space, reward function, and action space t , a t , r t ;

[0031] Evaluate the current action a t In state s t The performance is good or bad, and the advantage function A t Bootstrapping Strategy π θ improve;

[0032] Compare the new and old strategies for the current action a t The probability of selection r t (θ);

[0033] If updated, the amplitude is guaranteed to be within the range of (1-∈, 1+∈), otherwise, it will be punished;

[0034] Update the value function and use the loss function L VF (θ)=E t [(V(s t )-R t ) 2 ], optimize the value function V(s t );

[0035] Perform multiple rounds of iterations to optimize strategy π θ , until convergence.

[0036] Furthermore, the task decision layer adopts the SAC training algorithm, which increases the policy entropy value and adds the policy randomness weight to the reward function to form the objective function:

[0037]

[0038] Among them, E s～D,a～π [logπ φ (a|s t )], represents the entropy of the strategy, and α is the entropy coefficient, which controls the balance between entropy and reward.

[0039] The above embodiment has the following advantages or beneficial effects:

[0040] (1) The virtual simulation environment provides a variety of data samples for anti-deception training of UAVs;

[0041] (2) Simulate various deception methods and adopt adversarial strategy training to improve the strategy accuracy of the model;

[0042] (3) Adopt hierarchical reinforcement learning to achieve effective decision learning through hierarchical decision goals and improve the decision-making ability of the decision model.

[0043] Other features and advantages of the present application will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present application. The purposes and other advantages of the present application can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0045] Figure 1 A schematic flow chart of a method for proactively identifying interference and deception using an inspection drone provided by the present invention;

[0046] Figure 2 This is a schematic diagram of the route analysis of the strip-shaped inspection area in the embodiment provided by the present invention. DETAILED DESCRIPTION

[0047] To make the objectives, technical solutions, and advantages of this application more clearly understood, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0048] Example 1

[0049] like Figure 1 A method for proactively identifying jamming and deception by an inspection drone is shown, including the following:

[0050] Acquire real-time feedback data from each sensor of the drone to obtain first data; wherein the first data includes image data and point cloud data;

[0051] The acquired sensor data includes the direction of the route, which is combined with image data and point cloud data to determine the current environment. The model training process for image data and point cloud data is as follows:

[0052] For data collection and preparation, drones equipped with optical cameras and 3D-LiDAR sensors fly at fixed headings over narrow, elongated areas to collect image data and laser point cloud data. By collecting data at different times, weather conditions, and flight paths, a database of images and point clouds for inspection scenarios is constructed. This database is then cleaned to remove invalid data, such as overexposed and dark image data and sparse, point cloud data irrelevant to the inspection area. This cleansing process yields clean image data, characterized by a mixture of high-quality inspection targets and low-quality data such as obscured and blurred data.

[0053] Data annotation: Use image and point cloud annotation tools to annotate the cleaned data. The image part performs semantic segmentation and annotation of the inspection target area; the point cloud part performs 3D object annotation on each frame of point cloud data.

[0054] Model Training: After data annotation, we obtain the corresponding label and location data files for image and point cloud data. We separate these two types of data and randomly sample them, dividing them into training, test, and validation sets. We train the data using an image semantic segmentation model and a point cloud 3D object detection model, respectively, and then perform test and tuning on the test set. Through model pruning and quantization, we reduce the model size, ultimately achieving a model and weights that balance accuracy and computing power.

[0055] In practice, the following steps require model deployment: The trained model is deployed to the drone's onboard computer, which then accesses real-time location cloud and image data via the network and USB ports. The model is initialized with weight data, and the point cloud and image are preprocessed and directly fed into the model. The model then outputs the segmentation results for the object in the image and the location results in the point cloud.

[0056] The data post-processing module mainly includes point cloud detection result post-processing and image segmentation post-processing. The point cloud detection module can output the 3D bounding box data (x, y, z, l, w, h, θ) of the target area relative to the LiDAR coordinate system. Based on the detection result data, it calculates the plane equation formed by the center line and center point of the target area.

[0057] In some implementations, the route map predicted by the semantic segmentation model and the 3D object detection model obtained above is used to construct spatial elements such as the route path in the task decision model in the next stage.

[0058] Construct a virtual anti-deception training scenario for drones, and use a hierarchical reinforcement learning training model to train an intelligent anti-deception strategy model; hierarchical reinforcement learning includes a signal analysis layer and a task decision-making layer.

[0059] Among them, a reinforcement learning framework is constructed for the state space with the drone's posture state, channel state and interference pattern as elements, the reward function with positioning error, communication interruption and survival metric as reward and punishment elements, and the action space with path replanning, communication relay request and switching visual SLAM assistance as elements;

[0060] The intelligent anti-deception strategy model is quantitatively deployed on the drone's onboard computer.

[0061] Digitally model the drone and its environment. ANSYS can be used to build high-precision 3D dynamic models. 3D spatial propagation models based on ray tracing can also be used to model the surrounding environment. This can be connected to tools such as motion simulation tools and control algorithm libraries.

[0062] Design a dynamic interference source injection mechanism, including the construction of models such as periodic interference, random pulse interference, and intelligent game interference. Set a periodic interference model with an adjustable duty cycle from 10% to 60%. Its mathematical model is expressed as: T j : pulse period; τ is pulse width; duty cycle D = τ / T j .

[0063] Random pulse interference: A k : the amplitude of the kth pulse; τ k is the pulse width; f j is the interference center frequency; φ k : Random phase.

[0064] Or intelligent game interference models built with algorithms such as Q-learning.

[0065] At the same time, to build an intelligent anti-deception strategy training environment, it is necessary to construct a layered reinforcement learning framework: including a signal processing layer and a task decision layer. According to the decision goals of each layer, a reinforcement learning framework is constructed and the corresponding decision model is trained.

[0066] Design a reinforcement learning framework: Define the state space, including the drone's position (x, y, z, roll, pitch, yaw), that is, the three-dimensional coordinates (x, y, z), roll, pitch, and heading; specific indicators such as route deviation; channel status (SINR, BER), specifically including indicators such as signal-to-noise ratio and signal distortion; probability distribution of interference types, etc., which serve as elements for constructing the state space.

[0067] A reward function is designed, with penalties for positioning error and communication interruption. A measurement standard is set, which is specifically implemented as -0.5 / mm. Each millimeter of positioning error will result in a penalty of 0.5 proportional weight. The reward is based on survival time or number of steps, which is specifically implemented as -2 / step. That is, each wrong decision step will result in a penalty of 2 proportional weight. The action space is constructed by switching visual SLAM assistance and other methods.

[0068] These factors, including the direction of the route, combine image data and point cloud data to determine the current environment. A high-precision 2D or 3D map of the inspection area is constructed. The route file is directly imported into the map, and the vertical distance between the area and the route is calculated. This determines the distance parameter w to the center plane of the target area corresponding to each waypoint. This data is then added to the route file to form the final drone route.

[0069] The scene is analyzed using the point cloud data and image data obtained from the sensor data. As shown in the figure, the projection of the 3D bounding box on the XY plane, where the coordinates of point O are (x, y). Where AO = OB = 1 / 2, the coordinates of A and B can be obtained as:

[0070]

[0071] Here AB is a point on the bottom edge, then vector AB=(-lcosθ,lsinθ,0), vector The normal vector of the plane AOB can be obtained:

[0072]

[0073] Convenient calculation here uses (n x ,n y ,n z ) instead of the normal vector, the equation of the plane where the center line of the target area is located can be expressed as: x (Xn)+n y (Yy)+n z (Zz)=0.

[0074] Here you can calculate the distance between the current drone and the target area:

[0075]

[0076] Through the data collected by the drone in the past, the data is labeled and trained to train a binary classification network to determine whether the rectangle is the target area. The input is the coordinates of the four points in the image, and the output is 0 or 1 to determine whether the quadrilateral area meets the characteristics of the strip target area. For convenience, the calculation is performed here according to the camera being placed on the drone, that is, the camera phase plane is perpendicular to the direction of gravity. It is easy to calculate the equation of the center line of the long side of the circumscribed quadrilateral, and calculate the distance D from the center of the image to the center line of the circumscribed quadrilateral. At the same time, the angle θ between the center line and the image frame is calculated. In this way, we map the distance D to the real world, D W It is the distance between the center point of the real-world camera and the center line of the target area, f is the focal length of the camera, h is the height of the drone from the ground, and the z representation of (x, y, z, l, w, h, θ) for point cloud target detection.

[0077]

[0078] By calculating the distance between the drone and the center plane of the inspection target area, the distance is obtained by the laser sensor and the visual sensor respectively. According to the different adaptability of the laser radar and camera in different scenarios in the route, two weight factors are set, c L ,c L are the LiDAR weight factor and the image weight factor respectively. The final fusion distance is:

[0079] Among them, the determination of weight factor, c L ,c c It corresponds to the quality of image data and point cloud data. Here, features are used to measure the quality of point clouds and images in different environments. At the same time, in order to reduce the computing pressure of the airborne end, the quality evaluation coefficient is directly integrated into the detection and segmentation network. During the 3D detection network training phase, the line and surface feature statistics are performed on each frame of the preprocessed point cloud. Let the number of point cloud data features of the i-th frame be x i , calculate the characteristic distribution of all point cloud data, perform standardization and normalization processing to obtain the final point cloud quality evaluation coefficient

[0080]

[0081] Where μ is the mean, σ is the variance, and x max ,x min are the maximum and minimum values after normalization.

[0082] The image quality evaluation coefficient can also be obtained

[0083] A separate feature evaluation branch is introduced into the network to output the feature evaluation score for each scenario. At the same time, the KL divergence loss function is introduced to train this branch:

[0084]

[0085] in, is the point cloud 3D object detection network output, and N is the batch size.

[0086] Finally, the feature evaluation scores obtained by the two networks are normalized:

[0087]

[0088] Finally, the distance between the drone and the inspection area at the i-th moment can be obtained as:

[0089] δ i =c L δ′ L,i +c c D′ W,i

[0090] In the specific implementation, various scenarios are encountered, including strip-shaped inspection areas. When inspecting strip-shaped inspection areas, routes are usually set above the strip area or on both sides of the strip area. Figure 2 As shown in the figure, the target area for drone inspection is white, and the actual route of the drone during inspection is green. The inspection area is approximately considered a rectangular area. Here, inspection routes a and b are defined, and the width of the inspection channel is D. The centerline of the inspection area is O, and the width is d. The distance from the inspection route to the centerline of the inspection area is δ. When the drone is flying normally, it should meet the pre-set inspection constraints with the inspection area and the channel area. Here, the constraints that the drone cannot deviate from the route and cannot cross over the inspection area can be expressed as:

[0091]

[0092] Among them, (x′, y′)∈a′ is the two-dimensional projection coordinate of the actual UAV flight route; (x, y)∈a is the two-dimensional projection coordinate of the preset route. D is the width of the channel, d is the width of the inspection area, and δ is the distance between the route and the inspection area. Route a is expressed as: A a x+B a y+c a =0, the center line of the inspection area is represented by: A o x+B o y+c o =0.

[0093] Build a reinforcement learning framework, complete the model design of various elements including state space and action space, build the target loss function, and then use the policy optimization algorithm to train the model; among them, use the PPO algorithm or SAC algorithm for policy training.

[0094] In the signal processing layer, in the process of optimizing the target, in order to limit the strategy update amplitude, the clipping objective function of PPO is introduced, which is expressed as:

[0095] L CLIP (θ)=E t [min(r t (θ)A t ,clip(r t (θ),1-∈,1+∈)A t )]

[0096] A represents the probability ratio of the new and old strategies to select actions in the same state; t represents the advantage function;

[0097] At the same time, the value function is optimized and updated by minimizing the mean square error. t ), expressed as:

[0098] L VF (θ)=E t [(V(s t )-R t ) 2 ]

[0099] V(s t ;θ), the value function prediction of the current state; cumulative returns;

[0100] Policy entropy regularization, expressed as: L ENT (θ)=E t [H(πθ(s t ))], which can prevent the strategy from converging to the local optimum too early.

[0101] Combining the policy loss, value function loss and entropy regularization term, the total loss function is formed:

[0102] L(θ)=E t [L CLIP (θ)-c1L VF (θ)+c2L ENT (θ)], c1, c2 represent weight coefficients to balance policy optimization, value function update and entropy regularization.

[0103] The specific operation process of the PPO algorithm includes the following:

[0104] Use the current policy to interact with the environment, sampling the current value s in the state space, reward function, and action space t , a t , r t ;

[0105] Evaluate the current action a t In state s t The performance is good or bad, and the advantage function A t Bootstrapping Strategy π θ improve;

[0106] Compare the new and old strategies for the current action a t The probability of selection r t (θ);

[0107] If updated, the amplitude is guaranteed to be within the range of (1-∈, 1+∈), otherwise, it will be punished;

[0108] Update the value function and use the loss function L VF (θ)=E t [(V(s t )-R t ) 2 ], optimize the value function V(s t );

[0109] Perform multiple rounds of iterations to optimize strategy π θ , until convergence.

[0110] At the task decision layer, the SAC training algorithm is used for the reinforcement learning model. The specific process includes the following:

[0111] Initialize two sets of value functions Policy function π φ , and the value function V ψ ;

[0112] In the sampling space, the sampling includes a batch of sample values of the parameter values in the action space and the state space (s i ,a i ,r i ,s′ i ); and obtain the action value a in the current action space through the strategy function, and the distribution of the action space is a~π φ (·|s); interact with the environment to obtain the state and reward of the next moment, expressed as (s t ,a t ,r t ,s t+1 ), and saved in the sampling space;

[0113] Update the value function and use the minimum value function loss J V =E[(Vψ (s)-y V ) 2 ], so that y V Approximating the target value:

[0114]

[0115] The policy function is updated to minimize the objective function:

[0116]

[0117] Strategy model compression and quantitative deployment: For large-scale network models, knowledge distillation technology can be used to compress the original teacher training network into a student network model, significantly reducing the number of parameters. FP32 can also be converted to INT8 precision for quantitative inference and deployment on the drone's onboard computer.

[0118] The pre-processed first data is used as the input of the intelligent anti-deception strategy model to actively identify interference deception, output flight decisions in real time and execute them.

[0119] When a drone performs an inspection mission, it first loads the route file onto the onboard computer. The drone then executes the mission and flies along the predetermined route. The drone carries a lidar and camera, continuously collecting data from the inspection target area during flight. The onboard computer preprocesses this data before feeding the point cloud and image data into a point cloud object detection model and an image semantic segmentation model. The model outputs the detected route in real time, obtaining deviations from the predetermined route. This information, combined with environmental information obtained by other sensors, signal-to-noise ratio, and channel status, is then incorporated into an intelligent anti-spoofing strategy model to determine in real time whether the drone is being interfered with or deceived, and intelligently issue decision-making instructions to combat interference and deception.

[0120] Through the above scheme, the following advantages or beneficial effects are achieved:

[0121] (1) The virtual simulation environment provides a variety of data samples for anti-deception training of UAVs;

[0122] (2) Simulate various deception methods and adopt adversarial strategy training to improve the strategy accuracy of the model;

[0123] (3) Adopt hierarchical reinforcement learning to achieve effective decision learning through hierarchical decision goals and improve the decision-making ability of the decision model.

[0124] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for proactively identifying interference and deception by an inspection drone, characterized in that: These include: Acquire real-time feedback data from each sensor of the drone to obtain first data; wherein the first data includes image data and point cloud data; Construct a virtual anti-deception training scenario for drones and train an intelligent anti-deception strategy model using a layered reinforcement learning training model. The layered reinforcement learning includes a signal analysis layer and a task decision layer. Among them, a reinforcement learning framework is constructed for the state space with the drone's posture state, channel state and interference pattern as elements, the reward function with positioning error, communication interruption and survival metric as reward and punishment elements, and the action space with path replanning, communication relay request and switching visual SLAM assistance as elements; Quantitatively deploy the intelligent anti-deception strategy model on the drone's onboard computer; The pre-processed first data is used as the input of the intelligent anti-deception strategy model to actively identify interference deception, output flight decisions in real time and execute them.

2. The method for actively identifying interference and deception by an inspection drone according to claim 1, characterized in that: The training process of extracting routes by training the model of the first data is as follows: Data collection and preparation: collecting image data and laser point cloud data from optical cameras and 3D-LiDAR, and performing pre-processing operations on the data; Data annotation: Use image and point cloud annotation tools to perform semantic segmentation and annotation on the target area; For model training, the preprocessed data is randomly sampled and divided into training set, test set and validation set. The training set is used as the input of the image semantic segmentation model and the point cloud 3D object detection model to obtain the image semantic segmentation model and the point cloud 3D object detection model with converged training accuracy. The model is then tested and tuned using the test set.

3. The method for actively identifying interference and deception by an inspection drone according to claim 2, characterized in that: In the extraction of the route, the inspection area and the waterway area are set to meet the inspection constraints, which are specifically expressed as follows: Among them, (x ′ ,y ′ )∈a ′ is the two-dimensional projection coordinate of the actual UAV flight route; (x, y)∈a is the two-dimensional projection coordinate of the preset route. D is the width of the channel, d is the width of the inspection area, and δ is the distance between the route and the inspection area. Route a is expressed as: A a x+B a y+c a =0, the center line of the inspection area is represented by: A o x+B o y+c o =0.

4. The method for actively identifying interference and deception by an inspection drone according to claim 1, characterized in that: The reinforcement learning model of the signal analysis layer constructs a total loss function, and the specific operations are as follows: Strategy update constraint function: In the process of optimizing the target, in order to limit the strategy update range, the PPO clipping objective function is introduced, which is expressed as: L CLIP (θ)=E t [min(r t (i)A t ,clip(r t (θ),1-∈,1+∈)A t )] represents the probability ratio of the new and old strategies to select actions in the same state, A t represents the advantage function; Mean square error optimization of the value function: Update the value function V(s) by minimizing the mean square error t ), expressed as: L VF (θ)=E t [(V(s t )-R t ) 2 ] V(s t ;θ) value function prediction of the current state; cumulative returns; Policy entropy regularization optimization: Policy entropy regularization is expressed as: L ENT (θ)=E t [H(πθ(s t ))], which can prevent the strategy from converging to the local optimum too early; Combining the policy loss, value function loss and entropy regularization term, the total loss function is formed: L(θ)=E t [L CLIP (θ)-c1L VF (θ)+c2L ENT (θ)], c1, c2 represent weight coefficients to balance policy optimization, value function update and entropy regularization.

5. The method for actively identifying interference and deception by an inspection drone according to claim 4, characterized in that: The signal analysis layer adopts the PPO training algorithm, and the specific operations are as follows: Use the current policy to interact with the environment, sampling the current value s in the state space, reward function, and action space t , a t , r t ; Evaluate the current action a t In state s t The performance is good or bad, and the advantage function A t Bootstrapping Strategy π θ improve; Compare the new and old strategies for the current action a t The probability of selection r t (θ); If updated, the amplitude is guaranteed to be within the range of (1-∈, 1+∈), otherwise, it will be punished; Update the value function and use the loss function L VF (θ)=E t [(V(s t )-R t ) 2 ], optimize the value function V(s t ); Perform multiple rounds of iterations to optimize strategy π θ , until convergence.

6. The method for actively identifying interference and deception by an inspection drone according to claim 1, characterized in that: The task decision layer adopts the SAC training algorithm, which increases the policy entropy value and adds the policy randomness weight to the reward function to form the objective function: Among them, E s～D,a～π [logπ φ (a|s t )], represents the entropy of the strategy, and α is the entropy coefficient, which controls the balance between entropy and reward.

Images