A method for actively identifying interference and deception in inspection drones

Abstract

This invention provides a method for actively identifying interference and deception by an inspection drone. The method includes acquiring real-time feedback data from various sensors of the drone to obtain first data; constructing a virtual anti-deception training scenario for the drone and training an intelligent anti-deception strategy model using a hierarchical reinforcement learning training mode; quantitatively deploying the intelligent anti-deception strategy model onto the drone's onboard computer; using the preprocessed first data as input to the intelligent anti-deception strategy model to actively identify interference and deception, and outputting and executing flight decisions in real time. The virtual simulation environment provides diverse data samples for the drone's anti-deception training; it simulates various deception methods and employs adversarial strategy training to improve the model's strategy accuracy; and it uses hierarchical reinforcement learning, through hierarchical decision objectives, to achieve effective decision learning and improve the decision-making ability of the model.

Description

Technical Field

[0001] This application relates to the field of drone applications, and in particular to a method for inspection drones to actively identify interference and deception. Background Technology

[0002] In recent years, drone technology has been rapidly advancing towards intelligent and multi-scenario applications. Relying on high-precision navigation, AI visual recognition, and multi-sensor fusion technology, it has demonstrated significant advantages in the field of infrastructure inspection. In railway inspection, drones can quickly scan tracks, bridges, and slopes along the line, accurately identifying cracks and potential settlement hazards, with efficiency several times higher than manual labor. In power grid inspection, drone swarms can autonomously complete the detection of insulator damage and conductor strand breakage on hundreds of kilometers of high-voltage lines, significantly reducing the risks of high-altitude operations. In highway scenarios, drones monitor road surface defects in real time through 3D modeling and simultaneously analyze traffic flow data to assist in maintenance decisions. In river inspection, drones equipped with spectrometers and thermal imaging equipment can track water pollution sources around the clock and assess the risk of dam leakage.

[0003] However, when drones operate in complex environments, their GPS navigation, wireless communication, and sensor systems have significant security vulnerabilities. In railway inspections, if a drone encounters GPS signal spoofing or electromagnetic interference, it may deviate from the inspection track and collide with high-speed trains or overhead contact lines, causing derailments and economic losses. In power grid inspections, directional radio frequency interference can cause drone swarm navigation to fail, leading to collisions with high-voltage towers and causing regional power outages. In highway patrol scenarios, attackers could hijack aircraft and guide them onto the road to collide with normally traveling vehicles, causing traffic accidents. During river monitoring, malicious interference could force drones 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 anti-jamming and anti-spoofing technologies for drones have formed a multi-layered protection system. At the navigation level, multi-mode fusion positioning using GPS / BeiDou / inertial navigation / visual SLAM, combined with anti-spoofing algorithms, effectively resists conventional signal spoofing. The communication link employs frequency hopping spread spectrum technology and an AI spectrum sensing system, greatly enhancing its anti-interception capabilities.

[0005] However, defensive frequency modulation through active interference signals cannot effectively resist various forms of interference and deception; for example, a correct deception signal can easily bypass signal detection. Therefore, in the execution of tasks in multiple scenarios, there is still a set of anti-interference systems that actively identify deception to improve the level of drone deception resistance and risk prevention. Summary of the Invention

[0006] In view of the above-mentioned problems, the present invention provides a method for actively identifying interference and deception by inspection drones. The method uses reinforcement learning to construct a virtual simulation environment and train the drone's anti-deception strategy to cope with complex and ever-changing 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 actively identifying interference and deception by an inspection drone, comprising the following:

[0008] The first data is obtained by acquiring real-time feedback data from various sensors of the drone; the first data includes image data and point cloud data.

[0009] A virtual anti-deception training scenario for drones is constructed, and an intelligent anti-deception strategy model is trained using a hierarchical reinforcement learning training mode; the hierarchical reinforcement learning includes a signal analysis layer and a task decision layer.

[0010] Among them, a reinforcement learning framework is constructed, which includes a state space with UAV pose state, channel state and interference mode as elements, a reward function with positioning error, communication interruption and survival metric as reward and punishment elements, and an action space with path replanning, communication relay request and switching visual SLAM assistance as elements.

[0011] The aforementioned intelligent anti-deception strategy model was quantitatively deployed onto the UAV's onboard computer.

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

[0013] Furthermore, the training process for extracting flight routes from the model training of the first dataset is as follows:

[0014] Data collection and preparation: Acquire image data and laser point cloud data from optical cameras and 3D-LiDAR, and perform preprocessing operations on the data;

[0015] Data annotation involves using image and point cloud annotation tools to perform semantic segmentation and annotation of the target region.

[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 input to 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 convergent training accuracy. Then, the model is tested and optimized 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, specifically as follows:

[0018] Where (x′,y′)∈a′ are the two-dimensional projected coordinates of the actual UAV flight path; (x,y)∈a are the two-dimensional projected coordinates of the preset flight path. D is the width of the flight path, d is the width of the inspection area, and δ is the distance between the flight path and the inspection area. The flight path a is represented as: A a x+B a y+c a =0, the center line of the inspection area is represented as: 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, specifically as follows:

[0020] Policy Update Constraint Function: To limit the policy update magnitude during the optimization process, a PPO clipping objective function is introduced, 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 ratio of the probability of choosing an action under the same state using the new and old strategies. t Represents the dominance function;

[0023] Optimization of the mean square error of the value function: updating the value function V(s) by minimizing the mean square error. t ), represented as:

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

[0025] V(s t ;θ), predict the value of the current state using a function; Cumulative returns;

[0026] Policy entropy regularization optimization: Policy entropy regularization, denoted as: L ENT (θ)=E t [H(πθ(s t This can prevent the strategy from converging to a 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, optimized using a balancing strategy, value function update, and entropy regularization.

[0029] Furthermore, the signal analysis layer employs the PPO training algorithm, specifically as follows:

[0030] Interact with the environment using the current policy, and sample 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 of the following functions is evaluated, and the advantages and disadvantages are determined by the dominance function A. t Guiding Strategy π θ improve;

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

[0033] If an update is performed, the magnitude is guaranteed to be within the range of (1-∈, 1+∈); otherwise, a penalty is imposed.

[0034] The update value function uses the loss function L. VF (θ)=E t [(V(s t )-R t ) 2 ], optimize the value function V(s) t );

[0035] Perform multiple iterations to optimize strategy π θ Until it converges.

[0036] Furthermore, the task decision layer employs the SAC training algorithm, which adds a policy entropy value and incorporates policy randomness weights to the reward function to construct 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 implementation scheme has the following advantages or beneficial effects:

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

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

[0042] (3) Hierarchical reinforcement learning is adopted to achieve effective decision learning through hierarchical decision objectives and improve the decision-making ability of the decision model.

[0043] Other features and advantages of this application will be set forth in the following description and will be apparent in part from the description or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0045] Figure 1 This is a flowchart illustrating a method for actively identifying interference and deception by an inspection drone, provided by the present invention.

[0046] Figure 2 This is a schematic diagram of the flight path analysis of the strip-shaped inspection area provided in the embodiments of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0048] Example 1

[0049] like Figure 1 The method shown is a way for an inspection drone to actively identify interference and deception, including the following:

[0050] The first data is obtained by acquiring real-time feedback data from various sensors of the drone; the first data includes image data and point cloud data.

[0051] The acquired sensor data includes flight path deviation, which is combined with image data and point cloud data to make judgments about the current environment; the model training process for image data and point cloud data is as follows:

[0052] Data collection and preparation: Drones, equipped with optical cameras and 3D-LiDAR, fly along a fixed flight path over a narrow area to collect image data and laser point cloud data. An image and point cloud database for the inspection scenario is constructed by collecting data at different times, under different weather conditions, and along different flight paths. Based on this database, data cleaning is performed. Invalid data, such as overexposed or underexposed images, and sparse point cloud data irrelevant to the inspection area, are removed. Through data cleaning, clean image data is obtained, characterized by containing high-quality, cleaned inspection targets, as well as some occluded or blurry low-quality data.

[0053] Data annotation involves labeling the cleaned data using image and point cloud annotation tools. The image portion performs semantic segmentation and annotation of the inspection target area; the point cloud portion performs 3D target annotation on each frame of point cloud data.

[0054] Model Training: After data annotation, corresponding label and location data files for image data and point cloud data are obtained. These two types of data are separated and randomly sampled to form training, testing, and validation sets. Image semantic segmentation and point cloud 3D object detection models are used to train the data respectively, and the results are tested and optimized using the test set. Through model pruning and quantization to compress the model size, a model and weights that balance accuracy and computational power are finally obtained.

[0055] In practice, the following steps are required for model deployment: The trained model is deployed to the UAV's onboard computer, which acquires real-time point cloud and image data via Ethernet and USB interfaces. After the model loads and initializes the weight data, the preprocessed point cloud and images are directly input into the model. The model outputs the target segmentation results 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 3D bounding box data (x,y,z,l,w,h,θ) of the target region relative to the LiDAR coordinate system, and calculate the plane equation formed by the center line and center point of the target region based on the detection result data.

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

[0058] A virtual anti-deception training scenario for drones is constructed, and an intelligent anti-deception strategy model is trained using a hierarchical reinforcement learning training mode; the hierarchical reinforcement learning includes a signal analysis layer and a task decision layer.

[0059] Among them, a reinforcement learning framework is constructed, which includes a state space with UAV pose state, channel state and interference mode as elements, a reward function with positioning error, communication interruption and survival metric as reward and punishment elements, and an action space with path replanning, communication relay request and switching visual SLAM assistance as elements.

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

[0061] Digital modeling of UAVs and their environment can be performed using ANSYS to build high-precision 3D dynamic models; 3D spatial propagation models can be constructed based on ray tracing methods, and the surrounding environment can be modeled in 3D. It can also connect to 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 up a periodic interference model with an adjustable duty cycle from 10% to 60%, and its mathematical model is expressed as: T j : Pulse period; τ is pulse width; duty cycle D = τ / T j .

[0063] Random impulse interference: A k τ: Amplitude of the k-th pulse; k f is the pulse width; j The center frequency of the interference; φ k : Random phase.

[0064] Alternatively, intelligent game interference models can be built using algorithms such as Q-learning.

[0065] Meanwhile, to build an intelligent anti-deception strategy training environment, a layered reinforcement learning framework must be constructed, including a signal processing layer and a task decision layer. Based on the decision objectives of each layer, a reinforcement learning framework is constructed and the corresponding decision model is trained.

[0066] Design a reinforcement learning framework: Define a state space, including the UAV pose (x, y, z, roll, pitch, yaw), i.e., three-dimensional coordinates (x, y, z), roll, pitch, and heading; specific indicators include flight path deviation, etc.; channel state (SINR, BER), specific indicators include signal-to-noise ratio, signal distortion, etc.; interference type probability distribution, etc., which serve as elements for constructing the state space.

[0067] The system designs a reward function, penalties for positioning errors and communication interruptions, and sets a measurement standard, specifically -0.5 / mm, where each millimeter of positioning error will result in a penalty with a weight of 0.5. It also rewards survival time or steps, specifically -2 / step, meaning each incorrect decision step will result in a penalty with a weight of 2. The system also constructs an action space by switching to visual SLAM assistance.

[0068] The process involves various elements, including the flight path deviation, which are used to assess the current environment in conjunction with image and point cloud data. A high-precision 2D or 3D map of the inspection area needs to be constructed. The flight path file is then directly imported into the map, and the vertical distance between the area and the flight path is calculated. This yields the distance parameter w to the center plane of the target area corresponding to each waypoint. This data is added to the flight path file to form the final UAV flight path.

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

[0070]

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

[0072]

[0073] For ease of calculation, we use (n) here. x ,n y ,n z If we replace the normal vector with n, then the equation of the plane containing the centerline of the target region can be expressed as: n x (Xn)+n y (Yy)+n z (Zz) = 0.

[0074] This allows you to calculate the current distance between the drone and the target area:

[0075]

[0076] Using historical data collected by the drone, a binary classification network is trained to determine whether a rectangle is a target region. The input is the coordinates of four points in the image, and the output is 0 or 1 to determine whether the quadrilateral region meets the characteristics of a strip-shaped target region. For convenience, the calculation is performed with the camera upright on the drone, i.e., the camera plane is perpendicular to the direction of gravity. The equation of the centerline of the circumscribed quadrilateral is easily calculated, as is the distance D from the image center to the centerline of the circumscribed quadrilateral. The angle θ between the centerline and the image border is also calculated. This maps the distance D to the real world... 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 camera focal length, and h is the height of the UAV above the ground. It can be represented by z in point cloud target detection (x,y,z,l,w,h,θ).

[0077]

[0078] The distance from the UAV to the center plane of the inspection target area was calculated using both laser and visual sensors. Two weighting factors, c, were set based on the varying adaptability of the LiDAR and camera in different scenarios along the flight path. L ,c L These are the LiDAR weighting factor and the image weighting factor, respectively. The final fusion distance is obtained as follows:

[0079] Among them, the determination of the weighting factor, c L ,c c This refers to the quality of corresponding image and point cloud data. Features are used here to measure the quality of point clouds and images under different environments. To reduce the computational burden on the onboard device, quality evaluation coefficients are directly integrated into the detection and segmentation networks. During the training phase of the 3D detection network, line and surface feature statistics are performed on each preprocessed frame of point cloud data. Let x be the number of point cloud data features in the i-th frame. i The feature distribution of all point cloud data is statistically analyzed, and then standardized and normalized to obtain the final evaluation coefficient of point cloud quality.

[0080]

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

[0082] Similarly, an image quality evaluation coefficient can be obtained.

[0083] A separate feature evaluation branch is introduced into the network to output the feature evaluation score for each scene. A KL divergence loss function is also introduced to train this branch.

[0084]

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

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

[0087]

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

[0089] δ i =c L δ′ L,i +c c D′ W,i

[0090] In practice, various scenarios are encountered, including strip-shaped areas to be inspected. When inspecting such areas, flight paths are typically set above or to the sides of the strip. For example... Figure 2 As shown, the target area inspected by the drone is white, and the actual flight path during the inspection is green. We approximate the inspection area as a rectangular region, defining inspection paths a and b, with a width of D; the centerline of the inspection area is O, with a width of d; and the distance from the inspection path to the centerline of the area is δ. During normal drone flight, both the drone and the inspection area and flight path must satisfy pre-defined inspection constraints. These constraints prevent the drone from deviating from the flight path and from crossing over the inspection area, and can be expressed as:

[0091]

[0092] Where (x′,y′)∈a′ are the two-dimensional projected coordinates of the actual UAV flight path; (x,y)∈a are the two-dimensional projected coordinates of the preset flight path. D is the width of the flight path, d is the width of the inspection area, and δ is the distance between the flight path and the inspection area. The flight path a is represented as: A a x+B a y+c a =0, the center line of the inspection area is represented as: A o x+B o y+c o =0.

[0093] A reinforcement learning framework is constructed, and the model design of each element including the state space and action space is completed. After constructing the objective loss function, the model is trained using a policy optimization algorithm; among them, the PPO algorithm or the SAC algorithm is used for policy training.

[0094] In the signal processing layer, during the optimization process, to limit the policy update magnitude, a PPO clipping objective function is introduced, 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 ratio of the probability of choosing an action under the same state using the old and new strategies; t Represents the dominance function;

[0097] Simultaneously, the value function is optimized by minimizing the mean square error to update the value function V(s). t ), represented as:

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

[0099] V(s t ;θ), predict the value of the current state using a function; Cumulative returns;

[0100] Policy entropy regularization, denoted as: L ENT (θ)=E t [H(πθ(s t This can prevent the strategy from converging to a 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, optimized using a balancing strategy, value function update, and entropy regularization.

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

[0104] Interact with the environment using the current policy, and sample 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 of the following functions is evaluated, and the advantages and disadvantages are determined by the dominance function A. t Guiding Strategy π θ improve;

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

[0107] If an update is performed, the magnitude is guaranteed to be within the range of (1-∈, 1+∈); otherwise, a penalty is imposed.

[0108] The update value function uses the loss function L. VF (θ)=E t [(V(s t )-R t ) 2 ], optimize the value function V(s) t );

[0109] Perform multiple iterations to optimize strategy π θ Until it converges.

[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 Sum-value function V ψ ;

[0112] In the sampling space, sampling includes a batch of sampled values ​​(s) of parameter values ​​in the action space and state space. i ,a i ,r i ,s′ i The action value 'a' in the current action space is obtained through the policy function, and the action space satisfies the distribution of 'a' to 'π'. φ (·|s); interact with the environment to obtain the state and reward for the next moment, represented as (s t ,a t ,r t ,s t+1 ), and store it in the sampling space;

[0113] Update the value function, using the minimum value function loss J. V =E[(Vψ (s)-y V ) 2 ], making y V Approaching the target value:

[0114]

[0115] Policy function update, minimize objective function:

[0116]

[0117] Strategy model compression and quantization deployment: For large-scale network models, knowledge distillation technology can be used to compress the original teacher training network into the student network model, which can significantly reduce the number of parameters; and FP32 is converted to INT8 precision for quantization inference and deployed to the UAV onboard computer.

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

[0119] When a drone performs an inspection mission, it first loads the flight path file into its onboard computer. The drone then flies along the predetermined route. Carrying a lidar and camera, the drone continuously collects data on the target area during flight. The onboard computer first preprocesses this data, then feeds the point cloud and image data into a point cloud target detection model and an image semantic segmentation model. The model can output the detected flight path in real time, obtaining the deviation value from the predetermined path. This data, along with environmental information obtained from other sensors, as well as signal-to-noise ratio and channel status, participates in an intelligent anti-spoofing strategy model. This model determines in real time whether the drone is being spoofed or interfered with, and intelligently issues decision commands to counteract the interference and spoofing.

[0120] The above solution achieves the following advantages or beneficial effects:

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

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

[0123] (3) Hierarchical reinforcement learning is adopted to achieve effective decision learning through hierarchical decision objectives 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 variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for actively identifying interference and deception in inspection drones, characterized in that, Including the following: Real-time feedback data from various sensors of the UAV is acquired to obtain the first data; the first data includes image data and point cloud data; based on the different adaptability of LiDAR and camera in different scenarios along the flight path, two weighting factors are set. These are the LiDAR weighting factor and the image weighting factor, respectively, and the final fusion distance is obtained: ; A virtual anti-deception training scenario for drones was constructed. A high-precision 3D dynamic model was built using ANSYS, and a 3D spatial propagation model was constructed based on ray tracing. The surrounding environment was modeled in 3D, and motion simulation tools and control algorithm libraries were connected. A dynamic interference source injection mechanism was designed, including periodic interference and random pulse interference. An intelligent game-theoretic interference model constructed using the Q-learning algorithm was developed to form a composite adversarial interference training environment. A periodic interference model with an adjustable duty cycle from 10% to 60% was set up, and its mathematical model is expressed as follows: ; ; Pulse width; duty cycle Random pulse interference: ; The amplitude of the k-th pulse; The pulse width; To interfere with the center frequency; Random phase, using a hierarchical reinforcement learning training mode, to train an intelligent anti-deception strategy model; wherein, the hierarchical reinforcement learning includes a signal analysis layer and a task decision layer; The framework includes a state space consisting of UAV pose state, channel state, interference mode, fusion distance of UAV relative to the inspection area, and flight path deviation index; a reward function consisting of positioning error, communication interruption, and survival metric as reward and punishment elements; a reinforcement learning framework consisting of a motion space consisting of path replanning, communication relay request, and switching visual SLAM assistance. The intelligent anti-deception strategy model is compressed using knowledge distillation technology. After quantizing the inference by converting FP32 to INT8 precision, it is deployed to the UAV's onboard computer. The preprocessed first data, the real-time calculated fusion distance, and the flight path deviation value are used as inputs to the intelligent anti-spoofing strategy model to actively identify interference and spoofing. First, the signal analysis layer actively identifies the type and risk level of interference and spoofing, and then the mission decision layer outputs and executes flight decisions in real time.

2. The method for actively identifying interference and deception by an inspection drone as described in claim 1, characterized in that, The training process for extracting flight routes from the model based on the first data is as follows: Data collection and preparation: Acquire image data and laser point cloud data from optical cameras and 3D-LiDAR, and perform preprocessing operations on the data; Data annotation involves using image and point cloud annotation tools to perform semantic segmentation and annotation of the target region. 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 input to 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 convergent training accuracy. The model is then tested and optimized using the test set.

3. The method for actively identifying interference and deception by an inspection drone as described in 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, specifically as follows: ,in, It is the actual two-dimensional projected coordinates of the drone's flight path; These are the two-dimensional projected coordinates of the preset route. D is the width of the channel, and d is the width of the inspection area. This is the distance from the route to the inspection area. Route 'a' is represented as: The center line of the inspection area is represented as: .

4. The method for actively identifying interference and deception by an inspection drone as described in claim 1, characterized in that, The reinforcement learning model of the signal analysis layer constructs the total loss function, as follows: Policy Update Constraint Function: To limit the policy update magnitude during the optimization process, a PPO clipping objective function is introduced, expressed as: ； This represents the ratio of the probability of choosing an action under the same state using the old and new strategies. Represents the dominance function; Optimization of the mean squared error of the value function: updating the value function by minimizing the mean squared error. , represented as: ； Predicting the value function of the current state; Cumulative returns; Policy entropy regularization optimization: Policy entropy regularization is expressed as: This can prevent the strategy from converging to a local optimum too early; Combining the policy loss, value function loss, and entropy regularization term, the total loss function is formed: , This represents the weighting coefficients, optimized using a balancing strategy, value function updates, and entropy regularization.

5. The method for actively identifying interference and deception by an inspection drone as described in claim 4, characterized in that, The signal analysis layer uses the PPO training algorithm, and the specific operation is as follows: Interact with the environment using the current policy, and sample the current values ​​in the state space, reward function, and action space. , , ; Assess current actions In state The performance of the following functions is evaluated to determine their strengths and weaknesses, and to determine their strengths and weaknesses through the dominance function. Guiding strategy improve; Compare the old and new strategies for the current action Selection probability ; If an update is performed, the magnitude is guaranteed to be within ( Within the specified range, otherwise, penalties will be imposed; Update value function, using loss function Optimize value function ; Perform multiple iterations to optimize the strategy. Until it converges.

6. The method for actively identifying interference and deception by an inspection drone as described in claim 1, characterized in that, The task decision layer uses the SAC training algorithm, and the objective function is constructed by adding policy entropy and policy randomness weights to the reward function: ， in, , representing the entropy of the strategy, This is the entropy coefficient, which controls the balance between entropy and reward.

Images