Malicious unmanned aerial vehicle detection method and system based on anti-fact planning and constraint potential field

Abstract

The invention discloses a malicious unmanned aerial vehicle detection method and system based on anti-factual planning and a constrained potential field, and relates to the technical field of low-altitude airspace and intelligent perception. The method comprises the steps of fusing multi-modal operation data in a monitoring area to obtain state observation information of a detection target; constructing a constraint potential field of the airspace based on flight layout and flight constraint; solving an optimal control sequence and an anti-fact trajectory of the unmanned aerial vehicle flying to a compliance domain under the constraint potential field according to the dynamic observation model; generating a main evidence of the malicious unmanned aerial vehicle based on the optimal control sequence and the anti-fact trajectory, calculating an auxiliary evidence according to the multi-modal operation data, and fusing the main evidence and the auxiliary evidence to generate a malicious score so as to judge whether the detection target is the malicious unmanned aerial vehicle; the judgment condition of appearance or identity similarity is converted into intention and controllability measurement, the judgment evidence effectively avoids interference errors of a pseudo link and sensing abnormity, and the malicious detection efficiency, robustness and accuracy of the unmanned aerial vehicle are improved.

Description

Technical Field

[0001] This application relates to the fields of low-altitude airspace and intelligent sensing technology, specifically to a method and system for detecting malicious drones based on counterfactual programming and constrained potential fields. Background Technology

[0002] The widespread adoption of drones has led to increasingly prominent issues such as airspace conflicts, intrusion into sensitive areas, and close-range flights to critical infrastructure. Traditional security measures typically rely on fixed electronic fences, geographic no-fly databases, or remote ID broadcasts to identify unauthorized intrusions or unmarked flights. However, remote IDs can be disabled, altered, or relayed, and electronic fence rules struggle to cover complex temporary airspace adjustments and fine-grained time window restrictions, resulting in both false positives and false negatives in complex scenarios. More critically, these methods often focus on whether a location has crossed boundaries, failing to address whether the target has the ability or inclination to return to a compliant state, thus failing to reflect behavioral intent.

[0003] To this end, existing technologies propose to use other technologies for UAV detection, such as appearance and signal feature identification methods, including vision-based aircraft identification, passive radio frequency-based signal fingerprinting, and rotor feature matching based on radar micro-Doppler; anomaly detection and trajectory rule engine application, comparing statistics such as speed, acceleration, turning radius, and hovering time with thresholds, or using density estimation methods to determine the degree of "outlier" in manual feature space; and mitigating the vulnerability of single-modality through multimodal fusion. However, these technologies still have the following shortcomings: (1) Identity dependence and closed set assumptions lead to poor adaptability to unknown aircraft models, modified aircraft, and camouflage behavior; (2) Rule and threshold-driven anomaly detection ignores flight dynamics and control accessibility, and cannot judge the intention from a higher level of whether it can return to the compliance domain at a reasonable cost; (3) Evidence expression is mainly based on black box scores, lacking verifiable counterfactual trajectories and constraint violation quantities; (4) Insufficient robustness to link deception and observation degradation. Summary of the Invention

[0004] The purpose of this application is to address the problems in existing low-altitude drone security technologies, such as insufficient adaptability, inability to determine intent, and lack of traceability and interference resistance in the detection process, resulting in poor detection performance and unreliable detection results. This application proposes a malicious drone detection method and system based on counterfactual programming and constrained potential fields. By solving for the optimal control sequence and counterfactual trajectory of the drone returning to the compliance domain under constrained potential fields, the maliciousness of the drone can be determined. This transforms the judgment condition of appearance or identity similarity into a measure of intent and controllability. Furthermore, the detection process provides judgment verification evidence through precise data analysis and calculation, effectively avoiding interference errors from pseudo-links and sensor anomalies, thus improving the efficiency, robustness, and accuracy of malicious drone detection.

[0005] To achieve the above objectives, the technical solutions adopted in the embodiments of this application are as follows: In a first aspect, embodiments of this application provide a method for detecting malicious drones based on counterfactual programming and constrained potential fields, the method comprising: By integrating multimodal operational data within the monitoring area, the status observation information of the monitored targets can be obtained; Constructing a constrained potential field in the airspace based on flight layout and flight constraints; The optimal control sequence and counterfactual trajectory of the UAV flying to the compliance domain under the constraint potential field are obtained based on the dynamic observation model. The system generates primary evidence of malicious drones based on the optimal control sequence and counterfactual trajectory, and calculates auxiliary evidence based on multimodal operational data. The primary and auxiliary evidence are then fused to generate a malicious score to determine whether the detected target is a malicious drone.

[0006] In this scheme, multimodal data fusion is used to comprehensively estimate the current operating state of the target from multiple dimensions, avoiding the failure of a single sensor in scenarios such as occlusion and electromagnetic interference, and ensuring the basic accuracy of detection. Furthermore, the constraint potential field is used as a constraint condition for the law-abiding intent to solve the minimum cost control sequence, i.e. the optimal control sequence, and the corresponding counterfactual trajectory of the target returning from the current state to the compliance domain, which indirectly reflects the degree of deviation of the target from compliance. Combined with auxiliary evidence, it forms a malicious score of the target, reducing the interference risk of single evidence and improving the reliability of the detection results.

[0007] Preferably, the method of obtaining state observation information of the detected target from multimodal operational data within the fusion monitoring area includes: Based on the time-source synchronized multi-source monitoring equipment, the operating parameters and initial state estimation of the detected target are collected through a time sliding window to obtain multi-modal operating data, including at least the spatial position of the detected target, the three-axis velocity in the spatial coordinates, the heading angle, and the heading angular velocity; A Kalman filter is used to smooth the multimodal operational data to obtain an estimate of the current state of the detected target. The operation data of each mode is convolved according to the encoder to obtain state observation values; The state estimate and the state observation are used as state observation information.

[0008] Preferably, the constraint potential field in the airspace constructed based on flight deployment and flight constraints includes: An airspace constraint scenario graph containing nodes and edges is generated based on the electronic fence base map of the monitoring area, temporary airspace notices, facility buffer zones, maximum flight speed, and minimum flight altitude. Based on the aforementioned airspace constraint scenario diagram, each flight constraint and airspace rule is mapped to a penalty function, and the penalty functions are linearly combined to obtain the airspace constraint potential field.

[0009] Preferably, the penalty function includes a geographical potential function, a velocity constraint potential function, a height constraint function, and a neighborhood buffer potential function; The geographic potential function characterizes the degree of traversal of the flight polygon, and the function value is calculated based on the distance from the current position of the detected target to the center of the safe zone and the radius of the airspace safety boundary. The velocity constraint potential function characterizes the flight velocity constraint in the airspace, and the function value is calculated based on the maximum safe speed threshold and the velocity vector of the detected target. The altitude constraint function represents the flight altitude constraint, and the function value is calculated based on the minimum safe flight altitude and the current altitude of the detected target; The neighborhood buffer potential function characterizes the buffer penalty and is calculated based on the current two-dimensional planar coordinate vector of the detected target, the center and radius of the buffer circle, and the distance between the current position of the detected target and the center of the sensitive target.

[0010] Preferably, the step of obtaining the optimal control sequence and counterfactual trajectory for the UAV to fly to the compliance region under the constraint potential field based on the dynamic observation model includes: A dynamic observation model is constructed based on the aforementioned state observation information and historical compliant flight data. Based on the relevant constraints of the constrained potential field, the cost function is solved to obtain the optimal control sequence for the UAV to fly to the compliance domain with the minimum cost under the dynamic observation model, and the trajectory of flying to the compliance domain with the minimum cost is recorded as the counterfactual trajectory corresponding to the minimum cost.

[0011] Preferably, the step of constructing a dynamic observation model based on the state observation information and historical compliant flight data includes: Based on the state estimate and the control input parameters of the detected target, a structured state space network is used to perform linear state evolution, obtain the linear acceleration and angular velocity residuals of the detected target on the spatial position coordinate axis, and then establish a state evolution function based on the linear acceleration and angular velocity residuals. Simultaneously, the state estimate is projected onto an embedding space of the same dimension as the state observation to perform observation mapping on the state of the detected target; Based on historical compliant flight data, the state evolution function and the mapping observations are calibrated to complete the construction of the dynamic observation model.

[0012] Preferably, the generation of master evidence for a malicious drone based on the optimal control sequence and counterfactual trajectory includes: The reachability gap value is calculated based on the optimal control sequence and the counterfactual trajectory, and is used as the primary evidence to determine the flight malice of the detected target; The primary evidence is positively correlated with the accessibility gap value, representing the degree of difficulty for drones to return to the compliance domain with the intention of compliance.

[0013] Preferably, the step of calculating auxiliary evidence based on multimodal operational data includes: Multimodal fusion embeddings are obtained by weighted fusion of multimodal operational data, and the benign task energy of the detection target is calculated based on the multimodal fusion embeddings. Mahalanobis distance is calculated based on the mean and covariance of benign samples in the multimodal fusion embedding term; The observation-dynamic consistency residual of the detected target is calculated based on the observation mapping of the state estimate and the state observation. The benign task energy, the Mahalanobis distance, and the consistency residual constitute auxiliary evidence for determining the malice of the detected target. When the benign task energy and the Mahalanobis distance are both greater than the corresponding thresholds and the consistency residual increases, it indicates that the state observation of the detected target is incompatible with the physical model, and the detected target is a malicious drone.

[0014] Preferably, the step of fusing primary and secondary evidence to generate a malicious score to determine whether the detected target is a malicious drone includes: The primary evidence and the auxiliary evidence are fused based on dynamic weights to generate a malicious score; When the malicious score is greater than or equal to the malicious determination threshold, the detected target is determined to be a malicious drone; When the malicious score is less than the malicious judgment threshold and is within the threshold fluctuation range, a high-risk warning is triggered for the detected target.

[0015] Secondly, embodiments of this application provide a malicious drone detection system based on counterfactual programming and constrained potential fields, comprising: The detection module is used to fuse multimodal operational data within the monitoring area to obtain state observation information of the detected targets; The constraint module is used to construct the constraint potential field in the airspace based on flight layout and flight constraints; The calculation module is used to obtain the optimal control sequence and counterfactual trajectory of the UAV flying to the compliance domain under the constraint potential field based on the dynamic observation model; The judgment module is used to generate primary evidence of malicious drones based on the optimal control sequence and counterfactual trajectory, and to calculate auxiliary evidence based on multimodal operation data. The primary evidence and auxiliary evidence are fused to generate a malicious score to determine whether the detected target is a malicious drone.

[0016] The beneficial effects of this application are: 1. By constructing a constraint potential field, discrete scenario rules such as no-fly zones, speed / altitude restrictions, and critical facility buffer zones are transformed into continuous and computable mathematical constraints, making the constraints more continuous and overcoming the problem that fixed flight rules are difficult to adapt to different areas and scenarios, thereby improving the scenario adaptability of the detection method; at the same time, it provides a quantitative basis for subsequent malicious judgment of the detected target, making the solution of counterfactual trajectories smoother and more in line with physical laws, and improving the rationality of intent inference; 2. By calculating the minimum cost and counterfactual trajectory (i.e. the optimal path for the detection target to return to the compliance domain), it is possible to quantify whether the detection target is willing / compliant, realize the mathematical judgment of intent to reduce false positives and false negatives, and at the same time make the detection process traceable; 3. By integrating primary and secondary evidence, the risk of interference from single evidence is reduced. The mutual corroboration of evidence can effectively identify link spoofing such as replay, relay, and fake GNSS, thereby improving the reliability and robustness of the detection results. Attached Figure Description

[0017] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0018] Figure 1 A flowchart of a malicious drone detection method based on counterfactual programming and constrained potential field provided in this application embodiment.

[0019] Figure 2 The flowchart of step S1 in the malicious drone detection method based on counterfactual programming and constrained potential field provided in the embodiments of this application is shown.

[0020] Figure 3 This is a flowchart of step S2 in the malicious drone detection method based on counterfactual programming and constrained potential field provided in the embodiments of this application.

[0021] Figure 4 A schematic diagram of a malicious drone detection system module based on counterfactual programming and constrained potential field provided in an embodiment of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely one preferred embodiment of this application and are only used to explain this application. They do not limit the scope of protection of this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] Example 1: As Figure 1 As shown, a malicious drone detection method based on counterfactual programming and constrained potential fields includes steps S1-S4, wherein: S1. Obtain status observation information of the detection target by integrating multimodal operation data within the monitoring area.

[0024] As an optional implementation method, such as Figure 2 As shown, step S1 specifically includes: S11. Based on the time-source synchronized multi-dimensional monitoring equipment, the operating parameters of the detected target are collected and the initial state is estimated through a time sliding window to obtain multi-modal operating data, including at least the spatial position of the detected target, the three-axis velocity in the spatial coordinates, the heading angle and the heading angular velocity; S12. Use a Kalman filter to filter and smooth the multimodal operation data to obtain the current state estimate of the detected target; S13. Perform convolution processing on the operation data of each mode according to the encoder to obtain state observation values; The state estimate and the state observation are used as state observation information.

[0025] S2. Construct the constraint potential field in the airspace based on flight layout and flight constraints.

[0026] Specifically, the time-source synchronized multi-source monitoring equipment includes millimeter-wave radar, passive radio frequency receiver, and photoelectric (EO / IR) camera; millimeter-wave radar, passive radio frequency and photoelectric observations are collected within a fixed sliding window, and the current state estimate of the detected target and the observation embedding vector are obtained after time synchronization and filtering fusion.

[0027] Furthermore, the multimodal operational data characterizes the state vector of the detected target, expressed as: ; in, ( ) is the three-dimensional position of the UAV (target detection), ( This is used to detect the three-axis velocity of a target along the x, y, and z axes. It is the heading angle. It represents the heading angular velocity, and T represents the matrix.

[0028] In some embodiments, the radar outputs the target range-velocity pair and micro-Doppler spectral blocks, the radio frequency link outputs estimates of the angle of arrival and time difference of arrival, and the electro-optical camera outputs the image plane coordinates and depth inference obtained through multi-target tracking. Through geometric triangulation and physics-based extrinsic parameter calibration, the three-mode observations are jointly solved into an initial state estimate. Using a Kalman filter to Filtering and smoothing to obtain the current state Each mode is then compressed into an observation embedding using a fixed-structure encoder (convolution + state space layer). . and As the input state and observation representation for subsequent steps and processes, that is, state observation information.

[0029] In this embodiment, the data collection by multiple sensors (monitoring devices) achieves complementary advantages, avoids the failure of single sensors in scenarios such as obstruction and electromagnetic interference, and reduces the impact of environmental interference on state estimation. At the same time, different sensors are adapted to different low-altitude scenarios, and multi-source data fusion can cover a wider range of application scenarios, thereby solving the problem of insufficient scenario adaptability of existing technologies.

[0030] As an optional implementation method, such as Figure 3 As shown, step S2 specifically includes: S21. Generate an airspace constraint scenario graph containing nodes and edges based on the electronic fence base map of the monitoring area, temporary airspace notices, facility buffer zones, maximum flight speed, and minimum flight altitude. S22. Based on the airspace constraint scenario diagram, each flight constraint and airspace rule is mapped to a penalty function, and the penalty functions are linearly combined to obtain the airspace constraint potential field.

[0031] As an optional implementation, the penalty function includes a geographic potential function, a velocity constraint potential function, a height constraint function, and a neighborhood buffer potential function; The geographic potential function characterizes the degree of traversal of the flight polygon, and the function value is calculated based on the distance from the current position of the detected target to the center of the safe zone and the radius of the airspace safety boundary. The velocity constraint potential function characterizes the flight velocity constraint in the airspace, and the function value is calculated based on the maximum safe speed threshold and the velocity vector of the detected target. The altitude constraint function represents the flight altitude constraint, and the function value is calculated based on the minimum safe flight altitude and the current altitude of the detected target; The neighborhood buffer potential function characterizes the buffer penalty and is calculated based on the current two-dimensional planar coordinate vector of the detected target, the center and radius of the buffer circle, and the distance between the current position of the detected target and the center of the sensitive target.

[0032] Specifically, based on the airspace constraint scenario diagram, various flight and deployment rules in the airspace are mapped to continuously differentiable penalty functions and linearly combined to obtain the constraint potential field. The constraint potential field is represented as follows: ; ; ; ; ; in, The value represents the geographic potential function, and soft distance is used to measure the extent of crossing of the no-fly polygon. This represents the Euclidean distance from the current location to the center of the safe zone. Indicates the radius of the safety boundary; The value of the velocity constraint potential function represents the constraint velocity. The L2 norm of the drone's velocity vector. Indicates the maximum safe speed threshold; This represents the altitude constraint function value, used to constrain the flight altitude of the drone. For the minimum safe flight altitude, This represents the drone's height on the z-axis, which is its current height. The value of the neighborhood buffer potential function represents the buffer penalty. This represents the current two-dimensional planar coordinate vector of the drone. For the first The center and radius of the buffer circle Indicates the current position of the drone and the number of... Euclidean distance between the centers of the sensitive targets; , , and These are the weights of each function value, as shown in this embodiment. , , , .

[0033] It should be noted that the aforementioned constraint potential field is directly used as a component of the objective function in subsequent optimization, ensuring that the rule constraints are incorporated into the calculation in the form of energy. By unifying regulations such as geographical no-fly zones, speed / altitude restrictions, and critical infrastructure buffers into continuous and differentiable potential functions, these regulations can be directly incorporated into the optimization objective and can be quickly adapted to temporary airspace and fine-grained rule changes. This transforms discrete rules into computable energy, making counterfactual programming "optimizable, interpretable, and portable" at the regulatory level.

[0034] In this embodiment, by constructing an airspace constraint scenario graph, personalized rules for different regions can be dynamically incorporated, such as airport no-fly zones, school buffer zones, and urban height restrictions. This allows for adaptation to new scenarios without modifying the core algorithm, addressing the problem of existing technologies where fixed rules are difficult to adapt to changing scenarios. The constraint potential field transforms "compliance" into "the cost of distance from the compliance domain," indirectly reflecting the degree to which the drone deviates from compliance, overcoming the limitations of relying solely on thresholds to determine compliance and failing to correlate intent. Furthermore, the differentiability of the constraints ensures that they can be embedded in subsequent control sequence optimization, making the solution to counterfactual trajectories smoother and more consistent with physical laws, thus improving the rationality of intent deduction.

[0035] S3. Based on the dynamic observation model, obtain the optimal control sequence and counterfactual trajectory for the UAV to fly to the compliance domain under the constraint potential field.

[0036] As an optional implementation, step S3 specifically includes: S31. Construct a dynamic observation model based on the aforementioned state observation information and historical compliant flight data; S32. Solve the cost function based on the relevant constraints of the constrained potential field to obtain the optimal control sequence for the UAV to fly to the compliance domain with the minimum cost under the dynamic observation model, and record the trajectory of flying to the compliance domain with the minimum cost as the counterfactual trajectory corresponding to the minimum cost.

[0037] As an optional implementation, step S31 specifically includes: Based on the state estimate and the control input parameters of the detected target, a structured state space network is used to perform linear state evolution, obtain the linear acceleration and angular velocity residuals of the detected target on the spatial position coordinate axis, and then establish a state evolution function based on the linear acceleration and angular velocity residuals. Simultaneously, the state estimate is projected onto an embedding space of the same dimension as the state observation to perform observation mapping on the state of the detected target; Based on historical compliant flight data, the state evolution function and the mapping observations are calibrated to complete the construction of the dynamic observation model.

[0038] Specifically, a discrete-time state-space dynamics in residual form is established based on the state estimate and the control input parameters of the detected target (UAV): ; in Indicates the drone at a certain time The state vector at time, This represents the control input of the drone at time t. The parameterized state evolution function is implemented by a structured state-space network with time recursion characteristics: first, for Latent variables are obtained through linear projection, and then subjected to a single-layer SSM (length and...). After matching and activation function processing, the acceleration and angular velocity residuals are finally read linearly. The specific formula is as follows: Represents linear acceleration along the x-axis. Represents linear acceleration along the y-axis. Represents the linear acceleration along the z-axis. The residual is the angular velocity; the observation mapping uses... ,in, For a two-layer perceptron, the state is projected onto the... Embedding space of the same dimension; Δ is the time step, Δ=0.2 s.

[0039] Among them, SSM stands for Discrete-Time State-Space Model, which describes the dynamic relationship between the internal state and input / output variables of a system based on time series.

[0040] Furthermore, using historical compliant flight datasets The dynamics and observations are simultaneously calibrated using the least squares criterion to obtain a dynamic-observation model, which is represented as follows: ; in, For residual fitting terms, For observation consistency, After training, only minimal linear drift correction is allowed during runtime on the last layer of the observation mapping to accommodate minor variations in illumination and electromagnetic environment, and to adjust dynamic parameters. Keep it fixed to ensure the stability of extrapolation.

[0041] In some embodiments, the fixed planning time domain H=15, from the current state Starting with conditional diffusion sampling, the control sequence is solved. Each control vector in the control sequence This includes thrust control, pitch rate control, roll rate control, and yaw rate control. The optimal control sequence is obtained by solving the cost function, ensuring the trajectory is optimized under the world model. To return to the compliance domain at the minimum cost, the cost function is expressed as follows: ; Here, J represents the cost of returning to the compliance domain.

[0042] Specifically, this makes the trajectory under the dynamic observation model unfold... To return to the compliance domain at minimal cost, dynamic constraints must be satisfied: initial value To achieve real-time performance, a single-path, two-stage solution is adopted: In the first stage, based on the conditional noise prior of the "law-abiding intent family," eight candidate control sequences are generated using a fixed number of diffusion sampling steps (e.g., 8 times); in the second stage, the candidate with the lowest cost is selected as the initial value, and the results are processed accordingly. Perform a fixed 10-step gradient descent refinement to obtain the optimal solution. and minimum cost Simultaneously record the corresponding counterfactual trajectory .

[0043] In this embodiment, residual modeling ensures the extrapolation force for different wind fields, loads and attitude disturbances. Combined with small-scale online correction, it maintains stability while taking into account cross-scenario generalization. It can distinguish between "seemingly abnormal" appearances and "physically unreachable" causal mechanisms, reduce false alarms caused by environmental noise, and ensure the reliability and accuracy of malicious drone detection results.

[0044] It should be noted that the intention-conditional diffusion planner first generates candidate control sequences under the prior knowledge of law-abiding intentions, and then performs differentiability under the constraints of the dynamic observation model and potential field, stably obtaining the minimum cost and the optimal counterfactual trajectory. This approach combines the advantages of global exploration and local convergence, and can still provide solutions of controllable quality under real-time constraints, significantly reducing misjudgments and omissions.

[0045] S4. Generate primary evidence of malicious drones based on the optimal control sequence and counterfactual trajectory, and calculate auxiliary evidence based on multimodal operation data. Combine the primary evidence and auxiliary evidence to generate a malicious score to determine whether the detected target is a malicious drone.

[0046] As an optional implementation, the generation of master evidence for a malicious drone based on the optimal control sequence and counterfactual trajectory includes: S41. Calculate the reachability gap value based on the optimal control sequence and the counterfactual trajectory, as the primary evidence for judging the flight malice of the detected target; S42. The primary evidence is positively correlated with the accessibility gap value, representing the degree of difficulty for the drone to return to the compliance domain with the intention of compliance.

[0047] Specifically, based on minimum cost The counterfactual trajectory defines the reachability gap, expressed as follows: ; in, For the accessibility gap value, The infeasibility penalty constant, If any step of the counterfactual trajectory is still within the positive region of the constraint potential (i.e., infeasible), the indicator function takes the value 1. Directly equal to ,otherwise Equal to minimum cost ; The higher the value, the greater the difficulty in "returning to compliance with the intent to abide by the law," thus serving as primary evidence of malice.

[0048] As an optional implementation, the calculation of auxiliary evidence based on multimodal operational data includes: S43. Obtain multimodal fusion embeddings from multimodal operational data through weighted fusion, and calculate the benign task energy of the detection target based on the multimodal fusion embeddings; S44. Calculate Mahalanobis distance based on the mean and covariance of benign samples in multimodal fusion embedding terms; S45. Calculate the observation-dynamic consistency residual of the detected target based on the observation mapping of the state estimate and the state observation; S46. The benign task energy, the Mahalanobis distance, and the consistency residual constitute auxiliary evidence for determining the malice of the detected target. When the benign task energy and the Mahalanobis distance are both greater than the corresponding thresholds and the consistency residual increases, it indicates that the state observation of the detected target is incompatible with the physical model, and the detected target is a malicious drone.

[0049] Specifically, the multimodal fusion embedding term is obtained by gating and weighting the radar, radio frequency, and optoelectronic paths. The open set energy and prototype distance are calculated based on the multimodal fusion embedding term to obtain the benign task energy and Mahalanobis distance of the detected target. This includes obtaining a two-dimensional vector based on the logits vector output by the benign task discriminator. Therefore, based on the vector The benign task energy is obtained, and in this embodiment, the benign task energy corresponds to the benign or malignant nature of the detection target; the mean value of each cluster is pre-calculated for benign samples. With covariance Calculate the Mahalanobis distance to the nearest prototype; The energy of a benign task is represented as follows: ; ; Where E represents the energy of a benign task. This is the Mahalanobis distance.

[0050] Furthermore, based on observation mapping Calculate the observation-dynamic consistency residuals, which are expressed as follows: ; E, and This constitutes complementary evidence against spoofing and link deception, namely: when and At the same time, it is too large and During the ascent, a warning is issued indicating incompatibility between the observations and the physical model, supporting the determination of malicious intent.

[0051] It should be noted that among the auxiliary evidence, benign task energy is used to measure "whether it resembles a known benign task", prototype Mahalanobis distance is used to measure deviation from benign clusters, and the consistency residual between observation and dynamics is used to measure the deviation between sensing and physics. The three, together with the main evidence (accessibility gap), corroborate each other and can effectively identify link spoofing such as replay, relay, and pseudo GNSS, and automatically reduce weight when observations degrade, thereby improving overall robustness.

[0052] As an optional implementation, the step of fusing primary and secondary evidence to generate a malicious score to determine whether the detected target is a malicious drone includes: The primary evidence and the auxiliary evidence are fused based on dynamic weights to generate a malicious score; S47. When the malicious score is greater than or equal to the malicious determination threshold, the detection target is determined to be a malicious drone. S48. When the malicious score is less than the malicious judgment threshold and is within the threshold fluctuation range, a high-risk warning is triggered for the detection target.

[0053] Specifically, the main evidence Supporting evidence Malicious scoring based on dynamic weight fusion is represented as follows: ; Where M represents malicious rating, , and The weights are dynamic, determined by the real-time sliding window variance, which is expressed as follows: ; in, Indicates the confidence level. , As weight, , The original score for the evidence, including Three types, This represents the variance operator, used to measure the volatility of the evidence within the transition window. That is, the evidence was in the past A sequence of moments, It is the stability constant.

[0054] In this embodiment, the malicious intent determination threshold is fixed at 1. ,when ≥ Output a "malicious drone" determination at the time; when < but When the value approaches the threshold, output "High Risk Concern".

[0055] In this embodiment, for each determination, a structured evidence package is generated, which is represented as follows: ; It also includes a threshold (M) for normalization and determination. ⋆ (Quantilance parameters), the evidence package is written to the append-only log with a timestamp index and digitally signed to ensure that it can be verified and cannot be tampered with afterward. The on-site interface can directly render counterfactual trajectories and potential field heatmaps to help on-duty personnel understand "how to control it if it is benign".

[0056] It should be noted that the system outputs the current state, optimal control, counterfactual trajectory, minimum cost, reachability gap, and various statistical and threshold curves simultaneously with the judgment. By adopting append-only storage and cryptographic signature, the mechanism transforms "why it is malicious" from a black-box score into a verifiable trajectory and quantifiable evidence of "how it is not malicious", which facilitates law enforcement evidence collection and compliance review, and improves the efficiency and reliability of malicious drone detection.

[0057] Example 2, as Figure 4 As shown in the embodiments of this application, a malicious drone detection system based on counterfactual programming and constrained potential fields is also provided. The system includes: The detection module is used to fuse multimodal operational data within the monitoring area to obtain state observation information of the detected targets; The constraint module is used to construct the constraint potential field in the airspace based on flight layout and flight constraints; The calculation module is used to obtain the optimal control sequence and counterfactual trajectory of the UAV flying to the compliance domain under the constraint potential field based on the dynamic observation model; The judgment module is used to generate primary evidence of malicious drones based on the optimal control sequence and counterfactual trajectory, and to calculate auxiliary evidence based on multimodal operation data. The primary evidence and auxiliary evidence are fused to generate a malicious score to determine whether the detected target is a malicious drone.

[0058] In this embodiment, malicious scoring is obtained by constraining potential field, dynamic observation model, evidence judgment process and fusion weight, which gets rid of the dependence on model library and remote identity, and has reliable robustness against unknown models, shell replacement and appearance disguise, significantly reducing operation and maintenance costs.

[0059] The above-described embodiments are preferred embodiments of this application and are not intended to limit the specific scope of this application. The scope of this application includes but is not limited to the specific embodiments described above. All equivalent changes made in accordance with the shape, structure, and method of this application are within the protection scope of this application.

Claims

1. A method for malicious UAV detection based on counterfactual planning and constraint potential field, characterized in that: The method comprises the following steps: Fusion monitoring area multi-modal operation data acquisition detection target state observation information; Constructing a constraint potential field of airspace based on flight planning and flight constraints; According to the dynamics observation model, the optimal control sequence and counterfactual trajectory of the UAV flying to the compliance region under the constraint potential field are calculated; Based on the optimal control sequence and counterfactual trajectory, the main evidence of the malicious UAV is generated, and the auxiliary evidence is calculated based on the multi-modal operation data, and the malicious score is generated by fusing the main evidence and the auxiliary evidence to judge whether the detection target is a malicious UAV.

2. The counterfactual plan and constraint potential field based malicious drone detection method of claim 1, wherein: The method comprises the following steps: According to the time source synchronous multi-element monitoring device, the running parameters of the detection target are collected and the initial state estimation is performed through the time sliding window to obtain the multi-modal operation data, including at least the spatial position, the three-axis speed on the spatial coordinates, the heading angle and the heading angle speed of the detection target; The multi-modal operation data is filtered and smoothed by using a Kalman filter to obtain the current state estimation value of the detection target; The state observation value is obtained by convolution processing of each modal operation data according to the encoder. The state estimation value and the state observation value are used as the state observation information.

3. The counterfactual plan and constraint potential field based malicious drone detection method of claim 1, wherein: The method comprises the following steps: Based on the electronic fence base map of the monitoring area, the temporary airspace announcement, the facility buffer zone, the maximum flight speed and the minimum flight height, an airspace constraint scene graph containing nodes and edges is generated; Based on the airspace constraint scene graph, each flight constraint and airspace rule is mapped into a penalty function, and the penalty functions are linearly combined to obtain the constraint potential field of the airspace.

4. The malicious UAV detection method based on counterfactual planning and constraint potential field according to claim 3, wherein: The penalty function includes a geographic potential function, a speed constraint potential function, a height constraint function and a neighborhood buffer potential function; The geographic potential function represents the crossing degree of the flight polygon, and the function value is calculated based on the distance from the current position of the detection target to the center of the safety zone and the radius of the airspace safety boundary; The speed constraint potential function represents the flight speed constraint of the airspace, and the function value is calculated based on the maximum safe speed threshold and the speed vector of the detection target; The height constraint function represents the flight height constraint, and the function value is calculated based on the minimum safe flight height and the current height of the detection target; The neighborhood buffer potential function represents the buffer zone penalty, and the function value is calculated based on the current two-dimensional plane coordinate vector of the detection target, the center and radius of the buffer zone circle, and the distance between the current position of the detection target and the center of the sensitive target.

5. The counterfactual plan and constraint potential field based malicious drone detection method of claim 2, wherein: According to the dynamics observation model, the optimal control sequence and counterfactual trajectory of the UAV flying to the compliance region under the constraint potential field are calculated, According to the state observation information, a dynamics observation model is constructed combined with historical compliance flight data; Solve the cost function based on the relevant constraint conditions of the constraint potential field to obtain an optimal control sequence of the UAV flying to the compliance region with the minimum cost under the dynamic observation model, and record the trajectory of flying to the compliance region with the minimum cost as a counterfactual trajectory corresponding to the minimum cost.

6. The counterfactual plan and constraint potential field based malicious drone detection method of claim 5, wherein: The dynamic observation model is constructed according to the state observation information and historical compliance flight data, and includes: Based on the state estimation value and the control input parameter of the detection target, a structured state space network is used for linear state evolution to obtain linear acceleration and angular velocity residuals of the detection target on the spatial position coordinate axis, and then a state evolution function is established according to the linear acceleration and angular velocity residuals; Synchronously, the state estimation value is projected into an embedding space with the same dimension as the state observation value to observe and map the state of the detection target; The state evolution function and the mapped observation are calibrated based on the historical compliance flight data to complete the construction of the dynamic observation model.

7. The counterfactual plan and constraint potential field based malicious drone detection method of claim 5, wherein: The main evidence of the malicious UAV is generated based on the optimal control sequence and the counterfactual trajectory, and includes: The reachability gap value is calculated based on the optimal control sequence and the counterfactual trajectory to serve as the main evidence for judging the flight maliciousness of the detection target; The main evidence is positively correlated with the reachability gap value, representing the difficulty of the UAV returning to the compliance region under the law-abiding intention.

8. The counterfactual plan and constraint potential field based malicious drone detection method of claim 2, wherein: The auxiliary evidence is calculated according to the multi-modal operation data, and includes: The multi-modal fusion embedding item is obtained by weighting and fusing the multi-modal operation data, and the benign task energy of the detection target is calculated according to the multi-modal fusion embedding item; The Mahalanobis distance is calculated based on the mean and covariance of the benign samples in the multi-modal fusion embedding item; The consistency residual of the observation and dynamics of the detection target is calculated based on the observation mapping of the state estimation value and the state observation value; The benign task energy, the Mahalanobis distance, and the consistency residual constitute the auxiliary evidence for judging the maliciousness of the detection target. When the benign task energy and the Mahalanobis distance are greater than the corresponding threshold value at the same time and the consistency residual increases, it represents that the state observation of the detection target is incompatible with the physical model, and the detection target is a malicious UAV.

9. The counterfactual plan and constraint potential field based malicious drone detection method of any one of claims 7 or 8, wherein: The malicious score is generated by fusing the main evidence and the auxiliary evidence to judge whether the detection target is a malicious UAV, and includes: The malicious score is generated by fusing the main evidence and the auxiliary evidence based on dynamic weights; When the malicious score is greater than or equal to the malicious judgment threshold value, it is judged that the detection target is a malicious UAV; When the malicious score is less than the malicious judgment threshold value and within the threshold fluctuation range, a high-risk attention warning of the detection target is triggered.

10. A malicious drone detection system based on counterfactual planning and constraint potential fields, characterized in that: The method is suitable for the malicious UAV detection method based on counterfactual planning and constraint potential field according to any one of claims 1-9, and includes: A detection module is configured to fuse multi-modal operation data in a monitoring area to obtain state observation information of a detection target; A constraint module is configured to construct a constraint potential field of an airspace based on flight layout and flight constraints; A calculation module is configured to obtain an optimal control sequence of the UAV flying to the compliance region under the constraint potential field and a counterfactual trajectory according to a dynamic observation model. The judgment module is used for generating main evidence of the malicious UAV based on the optimal control sequence and the counterfactual trajectory, and calculating auxiliary evidence according to the multi-modal operation data, fusing the main evidence and the auxiliary evidence to generate a malicious score, so as to judge whether the detection target is the malicious UAV.

Images