A multi-layer early warning-based unmanned aerial vehicle swarm detection method and system

Abstract

The application relates to the technical field of unmanned aerial vehicle swarm detection, and discloses an unmanned aerial vehicle swarm detection method and system based on multilayer early warning, which comprises the following steps: collecting radar signal and photoelectric signal data and unmanned aerial vehicle speed, generating initial data by unified mapping according to preset state transition and observation matrix, making prediction-residual correction by combining preset noise parameters to obtain a dynamic trajectory, smoothing the trajectory, statistically analyzing the trajectory according to height granularity, adaptively clustering to form layered height intervals, identifying dense clusters according to distance radius and core points when the number of layers exceeds a threshold, extracting cluster centers and density gradients, constructing a probability inference structure according to the above and the speed, fusing prior formation, wind disturbance and delay, outputting a dynamic threat level, calling a pre-generated heat map to position a high-threat area when the threat exceeds a threshold, combining airspace safety constraints and speed to predict a future trajectory, updating a situation, adaptively configuring radar beams and coverage, and outputting real-time detection data. The method can realize reliable airspace management.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle swarm detection, and in particular to an unmanned aerial vehicle swarm detection method and system based on multi-layer early warning. BACKGROUND

[0002] At present, the application range of unmanned aerial vehicle swarm technology is continuously expanding in the fields of military, civil and commercial, and its high mobility, group cooperation and complex dynamic characteristics put forward higher requirements for airspace safety management. In order to cope with the detection and prevention and control requirements of low-altitude unmanned aerial vehicle group, the existing system relies on multiple types of intelligent sensors such as radar, photoelectric equipment and radio monitoring to realize the detection, tracking and identification of targets. However, due to the rapid adjustment of formation shape and large range maneuvering of the swarm in the three-dimensional airspace, the traditional detection system faces serious challenges in coverage range, response speed and dynamic adaptability.

[0003] In one prior art, fixed detection points or single type of intelligent sensors are usually used for airspace monitoring, and by continuously collecting information such as the bearing, distance and speed of the target, the position of the unmanned aerial vehicle is estimated and threat level determination is performed. This kind of scheme can realize basic monitoring tasks in single machine or low density target environment, but when the target is a high density unmanned aerial vehicle swarm, due to signal shielding, echo overlapping and measurement error accumulation, blind areas are easy to appear, resulting in part of the target cannot be continuously tracked, and the threat situation update lags behind.

[0004] The prior art is difficult to accurately identify the threat in a complex dynamic environment. SUMMARY

[0005] The present application provides an unmanned aerial vehicle swarm detection method and system based on multi-layer early warning, to solve the problem that the prior art is difficult to obtain accurate three-dimensional position information of the swarm in real time in a complex dynamic environment and accurately identify the threat by layer combining height and density.

[0006] In a first aspect, to solve the above technical problems, the present application provides an unmanned aerial vehicle swarm detection method based on multi-layer early warning, comprising:

[0007] Obtaining multi-source sensor data of radar signals and photoelectric signals and unmanned aerial vehicle speed;

[0008] Mapping the multi-source sensor data to a unified spatial coordinate system based on a preset state transition matrix and a preset observation matrix to obtain an initial data set;

[0009] Correcting the position according to the initial data set and a preset measurement noise parameter set to obtain dynamic trajectory data excluding measurement bias;

[0010] Fitting and outlier rejection processing are performed on three-dimensional positions in the dynamic trajectory data, segmented statistics are performed according to a preset height granularity, and height statistical features are extracted, clustering is performed according to the height statistical features, and layered height intervals are obtained;

[0011] When the number of unmanned aerial vehicles in the layered height interval exceeds a preset density threshold, a high-density aggregation area is identified, and a cluster center position and a density gradient are obtained;

[0012] According to the cluster center position, the density gradient, and the unmanned aerial vehicle speed, a preset prior queue shape template, a preset wind speed disturbance information, and a preset communication delay parameter are fused, a multi-dimensional Bayesian network is constructed, and potential threat behaviors of the unmanned aerial vehicle swarm are inferred according to the multi-dimensional Bayesian network, and a dynamic threat level is obtained;

[0013] When the dynamic threat level is higher than a preset threat threshold, a pre-generated global threat heat map is called, and the layered height interval and the density gradient are combined to identify and determine a high-threat area;

[0014] According to the high-threat area, a preset airspace safety constraint condition and the unmanned aerial vehicle speed are fused, and a future dynamic trajectory of the swarm is predicted, and airspace situation update data is obtained;

[0015] According to the airspace situation update data, radar beam parameters are configured and coverage is reconstructed, and real-time unmanned aerial vehicle detection data is obtained.

[0016] Preferably, the multi-source sensor data is mapped to a unified spatial coordinate system based on a preset state transition matrix and a preset observation matrix to obtain an initial data set, including:

[0017] The distance, azimuth angle, and pitch angle in the multi-source sensor data are subjected to denoising and smoothing processing, and are standardized to obtain a radar measurement data set;

[0018] The pixel coordinates of the photoelectric imaging in the multi-source sensor data are subjected to lens distortion correction, and the contour and texture features of the target are extracted to obtain a photoelectric image data set;

[0019] The radar measurement data set and the photoelectric image data set are time-synchronized, and if the time difference between the radar measurement data set and the photoelectric image data set is greater than a preset time difference threshold, the missing time data is filled in to obtain a synchronous fusion input data set;

[0020] The synchronous fusion input data set is input into the preset state transition matrix and the preset observation matrix, and a fusion state vector is calculated;

[0021] Based on the fusion state vector, the coordinate system of the multi-source sensor data is converted into a global spatial coordinate system to obtain an initial data set of the unmanned aerial vehicle swarm.

[0022] Preferably, the step of correcting the position according to the initial data set and a preset measurement noise parameter set to obtain dynamic trajectory data excluding measurement bias comprises:

[0023] extracting an initial state vector containing position and velocity information from the initial data set, and combining a preset state transition matrix to calculate a predicted state vector and a corresponding covariance matrix to obtain a predicted state data set;

[0024] calculating a noise covariance matrix according to the predicted state data set and the initial data set, combined with a preset measurement noise parameter set;

[0025] calculating an observation residual based on the predicted state vector and the initial data set, and calculating a filtering gain using a preset formula;

[0026] when the observation residual is greater than a preset residual threshold, correcting the state vector according to the filtering gain and the covariance matrix to obtain a corrected state data set;

[0027] combining the corrected state data set with a preset state transition matrix to generate dynamic trajectory data, and performing interpolation and smoothing processing on the dynamic trajectory data to remove instantaneous measurement bias to obtain dynamic trajectory data.

[0028] Preferably, the step of fitting and outlier removal processing on the three-dimensional position in the dynamic trajectory data, segmenting and extracting height statistical features according to a preset height granularity, and performing clustering according to the height statistical features to obtain a hierarchical height interval comprises:

[0029] performing smoothing processing on the three-dimensional position values in the dynamic trajectory data, and combining the state vector in the dynamic trajectory data with the corresponding covariance matrix to perform iterative updating to obtain smoothed first trajectory data;

[0030] fitting the first trajectory data based on spline interpolation, and detecting trajectory deviation of trajectory points from a preset interpolation curve;

[0031] when the trajectory deviation is greater than a preset trajectory deviation threshold, removing the corresponding trajectory points to obtain second trajectory data excluding outliers;

[0032] performing height segmentation statistics on the second trajectory data according to a preset height granularity, calculating the mean, variance, kurtosis and density gradient of each height interval to form a height statistical feature set;

[0033] determining a clustering distance threshold according to the mean and density gradient in the height statistical feature set, and performing clustering based on the clustering distance threshold to obtain a hierarchical height interval.

[0034] Preferably, when the number of drones within the hierarchical height range exceeds a preset density threshold, identifying high-density clustering areas and obtaining the cluster center location and density gradient includes:

[0035] When the number of drones in the layered height interval is greater than a preset density threshold, the height interval corresponding to the number of drones greater than the preset density threshold is marked as the height interval to be processed, and the set of height intervals to be processed is obtained.

[0036] For each height interval in the set of height intervals to be processed, density clustering is performed based on a preset distance radius and a preset minimum number of core points to divide high-density clustering regions and obtain clusters;

[0037] Based on the clusters, calculate the center position and density gradient of each cluster to obtain the cluster center position and corresponding density gradient.

[0038] Preferably, the step of constructing a multidimensional Bayesian network based on the cluster center location, the density gradient, and the drone speed, fusing a preset prior formation template, preset wind speed disturbance information, and preset communication delay parameters, and inferring the potential threat behavior of the drone swarm based on the multidimensional Bayesian network to obtain a dynamic threat level includes:

[0039] Based on the cluster center locations, the density gradient, and the drone speed, a multidimensional Bayesian network containing the relationship between location, speed, and density is constructed.

[0040] The multidimensional Bayesian network is fused with a preset prior formation template, preset wind speed disturbance information, and preset communication delay parameters to obtain the updated swarm cooperative formation intention;

[0041] Based on the swarm's cooperative formation intent and combined with preset threat determination rules, the deviation range, speed change degree, and similarity to the expected trajectory of the drone swarm are comprehensively evaluated, and the dynamic threat level is output according to the evaluation results.

[0042] Preferably, when the dynamic threat level is higher than a preset threat threshold, a pre-generated global threat heatmap is invoked, and high-threat areas are identified and determined by combining the hierarchical height range and the density gradient, including:

[0043] When the dynamic threat level is higher than the preset threat threshold, the pre-generated global threat heat map is invoked to divide the hierarchical height range into low-altitude, mid-altitude, and high-altitude layers.

[0044] For the low-altitude layer, the spatial distribution features of slow-moving low-altitude targets are extracted as anchor points for initial clustering to obtain low-altitude threat sub-regions.

[0045] By combining the density gradients corresponding to each layered height interval of the low-altitude threat sub-region with the layered height intervals, a secondary clustering analysis is performed on the UAV clusters in the mid-altitude and high-altitude layers to locate mid- and high-altitude target regions with concentrated density and consistent height.

[0046] The low-altitude threat sub-region and the medium- and high-altitude target region are spatially merged and their boundaries refined to determine the boundary range and spatial coordinate information of the high-threat region, thus obtaining the high-threat region.

[0047] Preferably, the step of predicting the future dynamic trajectory of the swarm based on the high-threat area, integrating preset airspace safety constraints and the speed of the UAV, to obtain airspace situation update data includes:

[0048] Extract the spatial distribution information of the high-threat area and combine it with the speed of the UAV to form an initial situation dataset;

[0049] If the initial situation dataset satisfies the preset airspace safety constraints, the speed of the UAV is iteratively calculated to obtain the trajectory prediction result of the future dynamics of the swarm.

[0050] Based on the trajectory prediction results and preset airspace safety constraints, safe path planning data is generated, situation assessment indicators are calculated, and airspace situation update data is output.

[0051] Preferably, the step of configuring radar beam parameters and reconstructing the coverage area based on the airspace situation update data to obtain real-time UAV detection data includes:

[0052] The trajectories of threat targets in the airspace situation update data are smoothed and matched with a preset penetration intent template to determine the set of penetration targets;

[0053] The azimuth and elevation angles of the target set are calculated to adjust the azimuth and elevation parameters of the radar beam, thus obtaining the beam parameters.

[0054] Based on the beam parameters, the scanning frame period is dynamically adjusted to extract the detection data of the core point coverage area and the boundary point coverage area, respectively.

[0055] The detection data is synchronized, filtered, and fused to update the target trajectory database and generate real-time detection data from the UAV.

[0056] Secondly, the present invention provides a drone swarm detection system based on multi-layer early warning, comprising:

[0057] The data acquisition module is used to acquire multi-source sensor data, including radar and photoelectric signals, as well as the speed of the UAV.

[0058] The spatial mapping module is used to map the multi-source sensor data to a unified spatial coordinate system based on a preset state transition matrix and a preset observation matrix to obtain an initial dataset;

[0059] The dynamic trajectory module is used to correct the position based on the initial dataset and the preset measurement noise parameter set to obtain dynamic trajectory data that eliminates measurement deviations;

[0060] The height stratification module is used to fit the three-dimensional position in the dynamic trajectory data and remove outliers, perform segmented statistics according to a preset height granularity and extract height statistical features, and perform clustering based on the height statistical features to obtain the stratified height intervals.

[0061] The density recognition module is used to identify high-density clustering areas and obtain the cluster center location and density gradient when the number of drones in the layered height range exceeds a preset density threshold.

[0062] The dynamic threat module is used to construct a multidimensional Bayesian network by fusing the cluster center location, density gradient, and UAV speed with a preset prior formation template, preset wind speed disturbance information, and preset communication delay parameters, and infer the potential threat behavior of the UAV swarm based on the multidimensional Bayesian network to obtain the dynamic threat level.

[0063] The threat template module is used to call a pre-generated global threat heatmap when the dynamic threat level is higher than a preset threat threshold, and to identify and determine high-threat areas by combining the hierarchical height range and the density gradient.

[0064] The airspace situation module is used to predict the future dynamic trajectory of the swarm based on the high-threat area, by integrating preset airspace security constraints and the speed of the UAV, and to obtain airspace situation update data.

[0065] The detection data module is used to update data based on the airspace situation, configure radar beam parameters and reconstruct the coverage area to obtain real-time detection data from the UAV.

[0066] Thirdly, the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the UAV swarm detection method based on multi-layer early warning as described above.

[0067] Fourthly, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the UAV swarm detection method based on multi-layer early warning as described above.

[0068] Compared with the prior art, the present invention has the following beneficial effects:

[0069] (1) The present invention adopts a chain processing of "multi-source intelligent sensor - unified coordinate - noise constraint - residual correction" to improve the accuracy and robustness of three-dimensional positioning. The radar / photoelectric data is denoised, calibrated and time-aligned, and predicted and corrected under the constraints of preset noise parameters. The position sequence is updated according to the residual weight, so as to reduce the measurement variance before fusion, suppress jump points after fusion, reduce drift and jitter errors, and improve the three-dimensional positioning continuity of fast maneuvering bee swarm.

[0070] (2) This invention uses a hierarchical identification path of "high-granularity statistics - feature extraction - adaptive threshold clustering - density cluster index" to improve the identification accuracy in complex environments. By determining the clustering distance threshold through height statistics, the spatial aggregation intensity is characterized by the density cluster center and density gradient, so as to achieve stable detection of low-altitude high density and mid-to-high-altitude rapid reconstruction of formations, and reduce the risk of threshold mismatch and scene migration.

[0071] (3) This invention constructs a closed loop of "intent inference - global heat map - trajectory prediction - radar adaptation" to improve response speed and detection resource utilization. The multi-dimensional probabilistic structure integrates position, velocity, density and prior parameters to generate formation intent and threat level; when the threshold is exceeded, the heat map focusing area is called, the trajectory is predicted in combination with airspace safety constraints, and the radar beam and scanning cycle are adaptively adjusted to achieve differentiated coverage of high threat directions and altitude layers, reducing detection delay and blind zone rate. Attached Figure Description

[0072] Figure 1 This is a schematic diagram of the process of the UAV swarm detection method based on multi-layer early warning provided in the first embodiment of the present invention;

[0073] Figure 2 This is a schematic diagram of the structure of the UAV swarm detection system based on multi-layer early warning provided in the second embodiment of the present invention. Detailed Implementation

[0074] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0075] Reference Figure 1 The first embodiment of the present invention provides a method for detecting drone swarms based on multi-layer early warning, including the following steps:

[0076] S11 acquires multi-source sensor data, including radar and photoelectric signals, and the speed of the UAV.

[0077] S12, based on the preset state transition matrix and the preset observation matrix, the multi-source sensor data is mapped to a unified spatial coordinate system to obtain the initial dataset;

[0078] S13, Correct the position according to the initial dataset and the preset measurement noise parameter set to obtain dynamic trajectory data with measurement deviation removed;

[0079] S14, Fit the three-dimensional position in the dynamic trajectory data and remove outliers, perform segmented statistics according to the preset height granularity and extract height statistical features, perform clustering according to the height statistical features to obtain hierarchical height intervals;

[0080] S15, when the number of drones in the layered height range exceeds the preset density threshold, identify high-density clustering areas and obtain the cluster center location and density gradient;

[0081] S16. Based on the cluster center location, the density gradient, and the drone speed, a preset prior formation template, preset wind speed disturbance information, and preset communication delay parameters are fused to construct a multidimensional Bayesian network and infer the potential threat behavior of the drone swarm based on the multidimensional Bayesian network to obtain the dynamic threat level.

[0082] S17, When the dynamic threat level is higher than the preset threat threshold, a pre-generated global threat heatmap is invoked, and high-threat areas are identified and determined by combining the hierarchical height range and the density gradient;

[0083] S18. Based on the high-threat area, the preset airspace safety constraints and the speed of the UAV are integrated to predict the future dynamic trajectory of the swarm and obtain airspace situation update data.

[0084] S19, based on the airspace situation update data, configure the radar beam parameters and reconstruct the coverage area to obtain real-time detection data of the UAV.

[0085] In step S11, it is necessary to acquire multi-source sensor data of radar signals and photoelectric signals, as well as the speed of the UAV, including:

[0086] First, the system deploys X-band pulse radar and electro-optical tracking cameras within the operational area, establishing a unified data channel via Ethernet or serial port protocols to achieve synchronous acquisition of data from both types of sensors. The radar transmits linear frequency modulated pulse signals with a fixed 50ms sampling period and receives echo signals. The digital signal processing module performs matched filtering, amplitude detection, and multi-target separation, outputting the distance (calculated in pulse round-trip time, accuracy ±1m), azimuth and elevation angles (obtained through array beamforming and spatial direction inversion, accuracy better than 0.5°), echo amplitude (in dBm), and measurement confidence level for each echo target. The measurement confidence level is calculated as follows: first, the peak amplitude of the echo signal is extracted, and its ratio (SNR) to the mean background noise is calculated; then, the target's positional offset is statistically analyzed over three consecutive frames to determine stability; if the variation is less than 0.2°, it is considered stable; finally, a confidence level (range 0~1) is generated by combining the similarity score between the current echo amplitude and historical echo shapes using a weighted formula.

[0087] The optoelectronic component employs a 1920×1080 resolution visible light CMOS camera with a working frame rate of 30FPS. Images undergo distortion correction and intrinsic / extrinsic parameter calibration using the Zhang Zhengyou calibration method, and are then recorded into the system database according to frame number and timestamp. Target detection utilizes a YOLOv8 network, outputting target category, confidence score (a floating-point number between 0 and 1), and pixel-level bounding boxes (x, y, and width / height pixel values ​​of the top-left corner). Gimbal attitude angles, including pitch, yaw, and roll angles, are calculated using built-in gyroscopes and accelerometers, with an update frequency of 100Hz.

[0088] The UAV's flight speed is calculated using a combination of GNSS (such as BeiDou-2) and a three-axis gyroscope. The horizontal speed is obtained by dividing the distance difference between two consecutive GNSS positioning points by the time interval (unit: m / s, accuracy within 0.5 m / s). The heading angle is obtained from the trajectory tangent direction (accuracy better than 1°). The climb rate is converted from the altitude change rate. The three-dimensional velocity vector consists of these three components. Simultaneously, the system assesses the accuracy level (A to D levels) of the flight speed results based on GNSS reception status, number of satellites, and gyroscope drift error.

[0089] Secondly, the installation position and attitude are calibrated to unify the radar and photoelectric data into the same coordinate system; the intrinsic interference distortion of the photoelectric camera is calibrated; the time is unified through network time synchronization or hardware pulse, and the sampling time is aligned by interpolation or nearest neighbor resampling.

[0090] Next, the system performs preprocessing operations on radar, electro-optical, and velocity data respectively to improve the robustness and consistency of subsequent multi-source fusion. For radar data, firstly, a sliding median filter is used to remove isolated outliers, with the filter window width set to 3-5 frames; secondly, an amplitude threshold method is applied, setting the static noise threshold to remove echo signal intensities lower than the background mean plus twice the standard deviation; finally, a moving average smoothing process is performed on consecutive frames of the target track, with a window size of 5 frames, to suppress the disturbance of echo jitter on the angle measurement results. For electro-optical image data, the system first performs radial distortion correction using camera intrinsic parameters, then unifies the image exposure values ​​to the 0.45-0.55 range, and normalizes the RGB channels by fixing white balance parameters; the target detection boxes are deduplicated using the non-maximum suppression (NMS) method, with the IOU threshold set to 0.5. For UAV velocity data, the system performs a consistency check on the three-dimensional velocity changes between adjacent sampling points. If the velocity variance of three consecutive frames exceeds a preset threshold (e.g., 1.5 m / s²), it is considered to have navigation drift or abrupt change, and the data segment is automatically removed and marked as low-confidence data. Furthermore, all data is uniformly converted to the International System of Units (SI) (meter, second, degree) and accompanied by corresponding observation confidence levels and historical noise statistics to support the confidence-weighted processing of subsequent fusion modules. During quality control, the system constructs a data integrity flag matrix for each moment, marking the availability status of radar data, electro-optical images, and velocity information, as well as the time alignment between the three types of data. If data from a certain source is missing, for example, if an effective frame of an electro-optical image is not obtained due to occlusion, other available data is retained and the integrity flag for that channel is set to "0". For situations where velocity data is temporarily lost or experiences abnormal jumps (e.g., a sudden increase in acceleration exceeding ±5 m / s²), the system uses a linear extrapolation algorithm to estimate the velocity value at the current moment based on the velocity data of the previous two frames. The extrapolated value is also accompanied by a confidence level label, set to 0.7 (lower than 1.0 for regular sampled data). During the fusion calculation, this confidence level is used to reduce the weight of the channel, meaning it only accounts for 70% of the fusion weight. If three or more consecutive frames (including the current frame) fail to align time or a channel is marked as "unavailable," it is considered a "continuous mismatch" state. The system automatically triggers an anomaly alarm and switches the fusion strategy from weighted average mode to "dominant source conservative estimation" mode. At this point, only the remaining reliable sources are used to output the target state estimate, and the current fusion result is marked as "degraded output" for subsequent module identification and processing. This process ensures that the system still has controllable and stable output capabilities even under conditions of partial data source failure or severe asynchrony. Finally, after completing the data integrity flag construction and anomaly identification, the system performs standardization processing on each source data, including field completion, unit unification, and numerical range normalization, removing invalid frames and duplicate data to ensure that the data quality of each channel meets the standards.Subsequently, the five types of data—radar, electro-optical, attitude, velocity, and mass indicators—are synchronized and aligned using timestamps as the primary key, and then stored in a circular buffer. This buffer supports fast retrieval operations based on time indexes or target numbers. The final output is a cleaned, standardized, and time-aligned multi-source raw dataset, which serves as the input for step S12.

[0091] In step S12, the multi-source sensor data needs to be mapped to a unified spatial coordinate system based on a preset state transition matrix and a preset observation matrix to obtain an initial dataset, including:

[0092] The range, azimuth, and elevation angles in the multi-source sensor data are denoised and smoothed, and then standardized to obtain a radar measurement dataset.

[0093] Lens distortion correction is performed on the pixel coordinates of photoelectric imaging in the multi-source sensor data, and the contour and texture features of the target are extracted to obtain a photoelectric image dataset.

[0094] The radar measurement dataset and the photoelectric image dataset are synchronized in time. If the time difference between the radar measurement dataset and the photoelectric image dataset is greater than a preset time difference threshold, the missing data is filled in to obtain a synchronized fusion input dataset.

[0095] The synchronous fusion input dataset is input into a preset state transition matrix and a preset observation matrix to calculate the fusion state vector;

[0096] Based on the fused state vector, the coordinate system of the multi-source sensor data is converted into a global spatial coordinate system to obtain the initial dataset of the UAV swarm.

[0097] In one specific embodiment, sliding denoising and small window smoothing are performed on the range, azimuth, and elevation angles in the radar measurements to remove abrupt jumps and isolated spikes, unify the units to meters and degrees, and retain the timestamps and measurement confidence scores to form the radar measurement dataset; lens distortion and geometric correction are performed on the photoelectric image stream, camera intrinsic and extrinsic parameters are loaded, target detection is completed, and the category, confidence score, and pixel coordinates are output; at the same time, the gimbal attitude and timestamp are recorded, and contour, corner, and texture features are combined and deduplicated using nonmaximum suppression to form the photoelectric image dataset.

[0098] In one specific embodiment, the system first pairs the position and velocity information collected by the radar and photoelectric sensors respectively, using a unified time stamp as a reference. When the time difference between the two data sources does not exceed a preset time difference threshold (e.g., 50ms), the system directly aligns and uses them; if the time difference exceeds the threshold, linear interpolation or nearest neighbor resampling is used to fill in the data with large time differences, while labeling the corresponding interpolation source and compensation weight, thus obtaining a time-aligned and fully labeled synchronous fusion input dataset. Subsequently, the system inputs this synchronous fusion input dataset into a preset state transition matrix and observation matrix for state updates. The state transition matrix is ​​set to 6×6 dimensions, describing the motion changes of the target in continuous time steps in inertial space, with a typical structure including discrete transition forms between position, velocity, and acceleration; the observation matrix is ​​set as a projection relationship matrix used to project radar and photoelectric measurements (i.e., measured position, velocity, or angle) onto the system state space, and is set to a sparse matrix of 3×6 or 4×6 depending on the measurement dimensions of each sensor. The system makes predictions based on the state vector of the previous time step using the state transition matrix. It then combines the radar and photoelectric measurement data of the current time step to calculate the residual between the prediction and the observation. Finally, it performs weighted corrections using a set of measurement noise parameters (including the variance and covariance obtained from the calibration of each sensor) to update the current fused state vector.

[0099] Subsequently, based on the velocity and position information in the fused state vector, and combined with known sensor extrinsic parameters (such as the relative attitude transformation matrix between radar and camera) and platform attitude information, the system uses rigid body transformation calculations to uniformly map the results in the radar coordinate system and camera coordinate system to the global spatial coordinate system. Finally, an initial dataset record is generated for each identified target. This record contains at least the following fields: 1) a unified time stamp for subsequent time alignment; 2) the target's position and velocity in the global coordinate system; 3) a source weight, determined by the type of sensor data, the interpolation compensation ratio, and its corresponding measurement confidence level (e.g., high weight for raw data, such as 0.9, and low weight for interpolated data, such as 0.4); 4) a synchronization method flag, indicating whether the record was obtained through direct alignment, interpolation compensation, or resampling; and 5) a quality flag, determined based on the degree of time alignment, the size of the observation residual, and the confidence level of the corrected state. For example, if the observation residual is within one time the noise standard deviation and the interpolation ratio is less than 20%, it is marked as "high quality"; otherwise, it is marked as "low quality" or "requires review." These fields together form a unified initial dataset for subsequent tracking, identification, and decision processing.

[0100] In one specific embodiment, both the millimeter-wave radar and the visible light gimbal output target detection results at a frequency of twenty frames per second. To ensure data alignment accuracy, the system sets a preset time difference threshold of twenty milliseconds. During the data preprocessing stage, the radar range sequence is denoised by sliding three frames, and the angle data is smoothed using a five-point moving average to reduce jitter interference. During target screening, if the confidence score of the detection box is lower than 0.3 (a normalized value between 0 and 1), it is considered unreliable and is discarded. Subsequently, the system performs pairing processing on the two types of sensor data based on a unified time stamp: when the time difference between adjacent frames is within fifteen milliseconds, they are directly aligned; if the time difference reaches thirty milliseconds, linear interpolation is performed on the photoelectric data to complete it, and the source weight of the data in that frame is reduced to 50% of the original value. In the fusion processing stage, the system constructs a state prediction process based on a constant velocity motion model. The state quantity includes a six-dimensional vector of position and velocity, and the update step size is set to fifty milliseconds. The state prediction result at each moment is combined with the current observation value and the corresponding measurement noise weight to perform residual calculation and weighted correction, resulting in the fused target position and velocity output. Finally, based on the gimbal attitude information from external participants, the radar and photoelectric results are uniformly transformed into a global spatial coordinate system and processed continuously for approximately ten seconds to generate a stable and reliable initial dataset record. Each record includes at least: timestamp, global position, global velocity, source weight, synchronization method, and quality flag. In a specific embodiment, the output frame rate of both the millimeter-wave radar and the visible light gimbal is 20 frames per second. In typical urban low-altitude operation scenarios, the preset time synchronization threshold is set to 20 milliseconds, which can be relaxed to 30 milliseconds considering network jitter. If hardware time synchronization is used and the system is stabilized after testing, it can be tightened to 10 milliseconds. Samples with time differences exceeding the preset threshold have their confidence weight reduced during fusion. Distance information is denoised using a three-frame median filter, and angle data is smoothed using a five-point sliding window average. If the timestamp difference between radar and photoelectric data is within 15 milliseconds, they are directly paired; if it is between 15 and 30 milliseconds, the photoelectric data is linearly interpolated and the fusion weight is reduced by 20%. When the confidence of a detection box is below 0.3, it is directly discarded and does not participate in subsequent fusion.

[0101] State prediction employs a short-window constant velocity motion model, with state variables being a six-dimensional vector of three-dimensional position and velocity. The step size is fifty milliseconds, and the predicted state is corrected using radar and photoelectric observations. When the velocity direction change angle exceeds 45 degrees or the vertical acceleration exceeds 3 m / s² within three consecutive frames, the system switches to a constant acceleration model, and the change time period is recorded for anomaly analysis.

[0102] The preset observation matrix includes: millimeter-wave radar uses spherical coordinates, and obtains the three-dimensional spatial position through extrinsic parameter transformation; photoelectric data calculates the line-of-sight direction by combining pixel coordinates with extrinsic and extrinsic parameter matrices and gimbal attitude transformation, and obtains the spatial point position by combining radar distance or stereo vision depth. The weight allocation is dynamically adjusted according to the environment: when the visible light visibility is less than 100 meters, the weight of the photoelectric channel is reduced to 50% of the original value; if the radar signal echo confidence is less than 0.5 or there is obstruction, the photoelectric data is given priority. Multi-frame triangulation compensation is enabled when necessary, and the triggering conditions include: the spatial deviation between radar and photoelectric in the current frame exceeds 0.5 meters, and the target movement speed within two frames is less than 1 m / s; the triangulation processing constructs the parallax based on the center point of the photoelectric detection frame and the gimbal attitude within five consecutive frames and solves the three-dimensional intersection point.

[0103] After global coordinate transformation, the final result forms a stable tracking sequence that lasts for ten consecutive seconds, which is used to construct the initial fusion dataset.

[0104] In step S13, the position needs to be corrected based on the initial dataset and the preset measurement noise parameter set to obtain dynamic trajectory data after removing measurement deviations, including:

[0105] An initial state vector containing position and velocity information is extracted from the initial dataset, and a predicted state vector and its corresponding covariance matrix are calculated by combining the preset state transition matrix to obtain the predicted state dataset.

[0106] Based on the predicted state dataset and the initial dataset, a noise covariance matrix is ​​calculated using a preset set of measurement noise parameters.

[0107] The observation residuals are calculated based on the predicted state vector and the initial dataset, and the filter gain is calculated using a preset formula.

[0108] When the observed residual is greater than a preset residual threshold, the state vector is corrected according to the filter gain and the covariance matrix to obtain the corrected state dataset;

[0109] The corrected state dataset is combined with a preset state transition matrix to generate dynamic trajectory data. Interpolation and smoothing are then performed on the dynamic trajectory data to remove instantaneous measurement deviations, thus obtaining the dynamic trajectory data.

[0110] In a specific embodiment, firstly, position and velocity are extracted from the initial dataset in order of target number and time to form an initial state vector containing position and velocity components; then, the state at adjacent time moments is extrapolated by the preset state transition matrix to obtain the predicted position and predicted velocity at the next time moment, and uncertainties such as predicted position variance, velocity variance and covariance are generated to form a predicted state dataset.

[0111] Subsequently, the predicted state is paired with the observation data at the same time, and the preset measurement noise parameter set corresponding to the sensor channel is called to calculate the measurement covariance matrix of dimensions such as distance, azimuth angle, and pixel positioning, which are used for subsequent observation residual weighting.

[0112] In one specific embodiment, the system calculates a residual vector based on the differences in position and velocity between the current predicted trajectory and the real-time observation data. The absolute value of the residual represents the degree of deviation between the current predicted state and the actual observation. Subsequently, the system calculates a filter gain factor based on the covariance matrix of the previous state and the covariance matrix of the current observation data, combined with the linear least mean square criterion. This factor is used to determine the contribution ratio of the predicted state and the observed value in this round of fusion, and its calculation process is based on the matrix weight allocation formula. When the calculated residual value is less than a preset threshold, the system performs weighted fusion of the predicted position and the observed position according to the filter gain factor, and synchronously updates the velocity component according to the position correction result, thereby outputting a new state vector. The spatial error threshold can be set to within 1.5 meters, and the velocity error threshold can be set to within 0.3 meters / second, with the specific values ​​determined based on the nominal accuracy of the sensor and the real-time control requirements of the system. If the residual exceeds the above thresholds, the system determines that there is a sudden change or interference in the current observation. At this time, the weight factor of the observation data is set to zero, or the state vectors of the previous effective time and the next time are used for linear interpolation to generate alternative observations to participate in the update. Subsequently, the system updates the state covariance matrix, including the covariance terms of the position and velocity components, based on the corrected state vector and the magnitude of the current error change, as the input basis for the next prediction cycle.

[0113] In one specific embodiment, the system concatenates the corrected state vectors output at each time point in chronological order to form a preliminary dynamic trajectory sequence. Each state vector contains numerical components of position and velocity, where position is composed of three-axis coordinates, and velocity is calculated from the first-order difference of position over time. To handle gaps in the trajectory, the system uses linear interpolation to complete the missing position and velocity data. The interpolation operation calculates intermediate states at equal intervals according to two adjacent valid time points as boundaries, ensuring trajectory continuity. Subsequently, the system performs sliding weighted smoothing based on a fixed time window (e.g., 5 sampling periods). This processing is applied to the three-dimensional position and velocity components, with weights determined by the residual amplitude corresponding to each time point; the smaller the residual, the higher the weight. Through smoothing, position fluctuations caused by short-term observation anomalies are effectively suppressed, generating a stable and continuous trajectory curve. The timestamp field in the dynamic trajectory data comes from the system time records of each state update cycle; the smoothed 3D position and velocity are derived from the weighted smoothing calculation results mentioned above; the residual magnitude is the magnitude of the difference between the prediction and observation at each time point; the weight allocation ratio records the respective proportions of the predicted and observed values ​​in the fusion calculation at that time point, and this ratio is extracted and normalized based on the Kalman filter gain vector; the data source flag is used to identify whether the value at that time point is an observed value, a predicted value, or an interpolated value, and can be set as a flag bit (e.g., 0 for prediction, 1 for observation, 2 for interpolation); the quality level is jointly determined by the current residual magnitude and velocity change magnitude. If the residual and velocity change continuously and stably, it is marked as high quality; otherwise, it is marked as medium or low quality, so as to be used for trajectory confidence screening later. The final generated dynamic trajectory data structure is in time series form, recording all the above fields one by one, and is in a standard format for subsequent high-level hierarchical and density clustering processing.

[0114] It should be noted that, in a specific embodiment, the system first extracts the three-dimensional position and velocity information of the target in chronological order based on the target number and timestamp, forming an initial six-dimensional state vector [x,y,z,vx,vy,vz]. Then, using the assumption of constant velocity motion, a state transition relationship is constructed. By extrapolating the state vector at each moment to the next moment, the corresponding predicted position and velocity are obtained. The prediction results are also combined with the degree of historical perturbation to generate uncertainty measures of position and velocity, including prediction variance and covariance, for subsequent residual calculation and gain correction. When the sampling period is Δt, the state transition matrix takes the form:

[0115]

[0116] This setup enables the calculation of the next moment's position based on the current velocity, ensuring the continuity of the predicted trajectory over a short period. The output of this process serves as a predicted state dataset, providing a predictive benchmark for subsequent residual calculations.

[0117] In a specific embodiment, for example, extrapolating the above formula when Δt=0.1s can maintain trajectory continuity within a short window.

[0118] Subsequently, the predicted state is paired with the observations at the same time, and the preset measurement noise parameter set is called according to the sensor channel to construct the measurement covariance matrix R. Example parameters are: distance noise standard deviation 0.5m, azimuth 0.2°, elevation 0.2°, and photoelectric pixel positioning noise equivalent to 0.8 pixels. R is used for subsequent differential weighting of the observation residuals.

[0119] Next, the position and velocity residuals are calculated, and the gain K is obtained according to the prediction uncertainty and R using a preset gain formula to achieve the confidence distribution between the predicted and observed values: when the accuracy is comparable, K usually falls between 0.4 and 0.6; when the observation is significantly better than the prediction, K can be increased to above 0.7. The position and velocity are weighted and corrected according to K, and the corrected state vector is output while the covariance is updated synchronously.

[0120] The residuals at each time step are compared with a preset threshold (example threshold 1 m). If the residuals do not exceed the threshold, normal fusion is performed with the gain; if they exceed the threshold, the observation is downweighted to 0.2 or removed, and replaced with the predicted value or interpolation from the adjacent time step to suppress the influence of abnormal observations on the trajectory. The corrected state and covariance are used as the initial values ​​for the prediction at the next time step.

[0121] Finally, all corrected states are concatenated over time to generate dynamic trajectory data; missing samples are imputed by time interpolation, and a three-point weighted sliding smoothing method is used to reduce short-term jitter. The output dynamic trajectory data includes timestamps, smoothed 3D positions, velocity, residual amplitude, weight allocation ratio, data source identifiers, and quality levels, and can be directly used for subsequent high-level stratification and density clustering.

[0122] In step S14, the three-dimensional positions in the dynamic trajectory data need to be fitted and outlier removed. The data is then segmented and statistically analyzed according to a preset height granularity, and height statistical features are extracted. Clustering is performed based on these height statistical features to obtain hierarchical height intervals, including:

[0123] The three-dimensional position values ​​in the dynamic trajectory data are smoothed, and the state vector and the corresponding covariance matrix in the dynamic trajectory data are combined with the state vector in the dynamic trajectory data for iterative update to obtain the smoothed first trajectory data.

[0124] The trajectory data is fitted using spline interpolation, and the trajectory deviation between the trajectory points and the preset interpolation curve is detected.

[0125] When the trajectory deviation is greater than the preset trajectory deviation threshold, the corresponding trajectory point is removed to obtain the second trajectory data after removing the abnormal points;

[0126] The second trajectory data is segmented and statistically analyzed according to a preset height granularity. The mean, variance, kurtosis and density gradient of each height interval are calculated to form a height statistical feature set.

[0127] The clustering distance threshold is determined based on the mean and density gradient of the height statistical feature set, and clustering is performed based on the clustering distance threshold to obtain the hierarchical height interval.

[0128] In one specific embodiment, the system performs covariance-weighted smoothing on the 3D positions in dynamic trajectory data. Specifically, it reads the state vector and covariance matrix corresponding to each trajectory point sequentially, extracts the diagonal covariance elements of the 3D position, and normalizes them to ensure that the position covariance of each trajectory point is under a uniform dimension. Then, it assigns smoothing weights based on the normalized covariance values, using an inverse proportional function for calculation; that is, the larger the covariance value, the smaller the corresponding weight. For example, the maximum covariance corresponds to a weight of 0.2, the minimum covariance to a weight of 1.0, and other weights are linearly interpolated within this range. Based on these weights, the system first processes the position coordinates of each trajectory point sequentially from the trajectory starting point. In each step, the current point's position value is linearly combined with the smoothing result of the previous point. The combined weight is determined by the covariance weight of the current position and the result of the previous step, yielding the first forward smoothing result. Subsequently, starting from the trajectory endpoint in reverse order, each trajectory point is processed in the same way to generate a backward smoothing result. Finally, the forward and backward smoothing results are averaged at each trajectory point to obtain a smoothed trajectory after bidirectional fusion. This method not only reduces the risk of cumulative unidirectional bias but also ensures the stability and consistency of position data throughout the entire time period. In the processed first trajectory data, the original timestamp, velocity component, and quality level fields remain unchanged, ensuring direct connection with subsequent velocity calculation and quality control steps without additional conversion. In a specific embodiment, the system first extracts the three-dimensional coordinate point sequence from the first trajectory data and establishes a continuous position path in chronological order. Subsequently, a cubic spline interpolation method is used to connect these discrete trajectory points into a smooth curve to construct the fitted trajectory. For each original trajectory point, the system projects vertically from that point onto the fitted curve and calculates the straight-line distance between that point and the corresponding projection point on the fitted curve in three-dimensional space, obtaining the trajectory deviation at that point. This deviation reflects the degree of deviation of the actual trajectory from the fitted path. If the trajectory deviation value of a point is greater than 1.5 meters, exceeding the preset trajectory deviation threshold, the system marks the point as abnormal and removes it from the trajectory data. The system will perform linear interpolation between adjacent normal points to fill in the missing points and restore the integrity of the trajectory, thus obtaining the second trajectory data after removing the anomalies.

[0129] In one specific embodiment, based on the image feature dataset acquired above, the system first performs a contour matching operation to filter out regions where there may be fitting deviations. This process is performed by calculating the spatial positional difference between the feature contour of the current image and the contour of the standard fitting image. The system uses the relative position, directional offset, and degree of overlap of the contours as the judgment criteria to identify regions with positional offsets, angular rotations, or boundary misalignments. Based on this, the system further uses key point registration to correct each candidate deviation region. Specifically, this includes extracting key corner points and edge points within the region and comparing them with reference points at the same location in the standard image using template matching to determine the degree of geometric deformation in translation, rotation, or scaling dimensions. If the region has quantifiable geometric deviations, the system marks it as a region to be corrected and calculates a correction vector for each deviation region. Finally, the correction vectors of all regions to be corrected are fused to generate a complete fitting correction displacement map. This map serves as the input basis for subsequent motor drive signal adjustments, ensuring that deviation regions can be fitted and adjusted according to the expected trajectory, achieving precise correction. The entire process does not require formula descriptions; instead, it constructs a concrete and executable image deviation correction path through an engineering workflow of contour comparison and key point correction. In a specific embodiment, a clustering distance threshold is determined based on height statistical characteristics: when the mean height difference between adjacent intervals is less than one meter and the absolute value of the density gradient is less than five percent, they are considered to be in the same layer and merged; otherwise, they are retained as independent height layers. Upon completion, the layered height intervals are output, recording the layer number, upper and lower boundary heights, average height, point density, and temporal coverage.

[0130] In a specific embodiment, the method for obtaining the preset trajectory deviation threshold is as follows: select an interference-free historical trajectory, calculate the distance distribution according to cubic splines, remove the long tail above the 95th percentile, calculate the mean and standard deviation of the remaining samples, and use the mean plus twice the standard deviation as the threshold; for example, if the mean is 0.8 meters and the standard deviation is 0.35 meters, then the threshold is 1.5 meters, which is used to remove outliers while retaining normal short-term deviations.

[0131] In a specific embodiment, the zero to twelve meter range is divided into six intervals with a two-meter particle size and the above rules are applied: the zero to two meter, two to four meter, and four to six meter ranges meet the merging conditions and form the first height layer; the six to eight meter range is listed separately as the second height layer due to the large density gradient change; the eight to ten meter range and the ten to twelve meter range are merged into the third height layer, and the output includes the layer number, boundary height, average height, and point density.

[0132] In step S15, when the number of drones within the hierarchical height range exceeds a preset density threshold, it is necessary to identify high-density clustering areas and obtain the cluster center location and density gradient, including:

[0133] When the number of drones in the layered height interval is greater than a preset density threshold, the height interval corresponding to the number of drones greater than the preset density threshold is marked as the height interval to be processed, and the set of height intervals to be processed is obtained.

[0134] For each height interval in the set of height intervals to be processed, density clustering is performed based on a preset distance radius and a preset minimum number of core points to divide high-density clustering regions and obtain clusters;

[0135] Based on the clusters, calculate the center position and density gradient of each cluster to obtain the cluster center position and corresponding density gradient.

[0136] In a specific embodiment, after obtaining the layered altitude intervals output in step S14, the system, using each altitude interval as a unit, combines real-time UAV flight data to count the number of UAVs currently in that altitude range within each interval. Specifically, during the counting process, the system iterates through the flight records of all operating UAVs, determining the current altitude value of each UAV. If the altitude value falls between the upper and lower limits of a certain altitude interval, the UAV is counted in the corresponding interval's count. By accumulating this process, the system can obtain the number of UAVs in each layered interval. This number is then compared with a preset density threshold to determine if there is excessive density or abnormal airspace resource occupancy in that altitude range, providing a basis for subsequent airspace reallocation or conflict avoidance processing. This statistical method does not rely on complex calculations; it only requires range judgment based on the altitude field to complete quickly, offering advantages such as strong real-time performance and wide applicability. When the number in a certain interval exceeds the threshold, that interval is marked as a pending altitude interval and added to the pending set. For each altitude range to be processed, extract the three-dimensional coordinates of all UAVs within the range, perform density clustering (such as DBSCAN) according to the preset distance radius and the preset minimum number of core points, expand to form a high-density clustering region in the neighborhood of the core point, and output one or more clusters.

[0137] In one specific embodiment, after clustering is completed, the system calculates the center location and density gradient for each cluster. The center location is calculated using a weighted average of the coordinates within the cluster, with weights allocated based on measurement confidence or positioning accuracy. The density gradient is defined as the rate of change of unit volume density within a fixed neighborhood over time. The resulting "cluster center-density gradient" pair serves as the input for subsequent threat identification and response decisions.

[0138] In one specific embodiment, the preset density threshold is determined based on airspace traffic statistics and minimum safety intervals: the average number of similar airspaces in the corresponding altitude range in history is statistically analyzed and multiplied by a safety redundancy coefficient to obtain the threshold. For example, if the average is six aircraft and the redundancy coefficient is 1.5, then the threshold is set to nine aircraft; if this number is exceeded, the clustering process begins.

[0139] In one specific embodiment, the preset distance radius is obtained by combining a "static safety radius" and a "dynamic safety extension." The static component is given by the aircraft dimensions and the minimum clearance, while the dynamic component is the product of the maximum cruising speed and the average control delay. The two components are added together and rounded according to the environment to obtain the cluster radius. For example, if the aircraft diameter is one meter, the minimum clearance distance is ten meters, the maximum speed is fifteen meters per second, and the control delay is 0.5 seconds, then the radius is approximately 1 meter + 10 meters + 15 × 0.5 = 18 meters.

[0140] In one specific embodiment, the preset minimum number of core points is selected through historical data simulation: using the cluster profile coefficient and the proportion of noise points as indicators, a stable range for different numbers of core points is searched, and the minimum value that maintains a high profile coefficient and a low noise proportion is taken as the parameter. For example, when the optimal stability is obtained with a core point count of three and the noise point proportion is less than five percent, the minimum number of core points is set to three.

[0141] In step S16, a multidimensional Bayesian network needs to be constructed by fusing the cluster center location, density gradient, and UAV speed with a preset prior formation template, preset wind speed disturbance information, and preset communication delay parameters. Based on this multidimensional Bayesian network, the potential threat behavior of the UAV swarm is inferred to obtain the dynamic threat level, including:

[0142] Based on the cluster center locations, the density gradient, and the drone speed, a multidimensional Bayesian network containing the relationship between location, speed, and density is constructed.

[0143] The multidimensional Bayesian network is fused with a preset prior formation template, preset wind speed disturbance information, and preset communication delay parameters to obtain the updated swarm cooperative formation intention;

[0144] Based on the swarm's cooperative formation intent and combined with preset threat determination rules, the deviation range, speed change degree, and similarity to the expected trajectory of the drone swarm are comprehensively evaluated, and the dynamic threat level is output according to the evaluation results.

[0145] In a specific embodiment, the system constructs a multidimensional Bayesian network containing the relationship between position, velocity, and density based on the cluster center positions, density gradients, and real-time velocities of each UAV obtained in step S15. To facilitate real-time inference, the continuous quantity intervals are discretized (velocity consistency is divided into high / medium / low, density gradient into rising / stable / falling, and distance from cluster center to restricted area into near / medium / far). Network nodes at least include formation compactness, velocity consistency, density change trend, degree of proximity to restricted area, and top-level cooperative formation intention. The causal relationship of the edges is preset as follows: increased density and improved velocity consistency jointly increase the probability of "cooperative contraction," and the continuous approach of cluster centers to restricted areas increases the probability of "approaching sensitive targets." The system injects observational evidence according to the update cycle and iteratively updates the posterior, with the number of iterations and time budget limited in the configuration to meet millisecond- to hundred-millisecond real-time requirements.

[0146] In one specific embodiment, the multidimensional Bayesian network is fused with three types of preset information to update the cooperative formation intention: first, a preset prior formation template (column, wedge, triangle, encirclement, etc.), providing expected features such as neighboring aircraft spacing, heading difference, speed difference, and inter-layer altitude difference; second, preset wind speed disturbance information, which aligns and compensates for speed direction deviations based on meteorological measurements to avoid misjudging deviations induced by crosswinds / gusts as tactical contraction; and third, preset communication delay parameters, which attenuate neighboring aircraft synchronization requirements according to link delay to prevent short-term speed inconsistencies caused by delay from being misjudged as formation disbandment. The fusion process first generates conditional probability priors based on the template, then corrects the relevant edge weights based on wind speed and delay, and outputs the cooperative formation intention and its confidence level.

[0147] In one specific embodiment, the system assigns a dynamic threat level based on the updated cooperative formation intent and pre-defined threat determination rules. The rules integrate three categories of quantitative indicators: deviation magnitude (cluster center approach speed, minimum distance to the preset boundary, and angle deviation from the controlled flight path), speed mutation degree (instantaneous increase in average speed, short-window change in speed variance, and acceleration persistence), and similarity to the expected trajectory (temporal / morphological similarity to prior templates, historical threat samples, and normal cruise samples). The system calculates a comprehensive threat score using a weighted average method based on the configured weights corresponding to the three indicators: deviation magnitude, speed mutation degree, and trajectory similarity. The score range is 0 to 1. Specifically, the system pre-defines the weight ratio for each indicator (e.g., deviation magnitude 0.4, speed mutation 0.35, trajectory similarity 0.25), and normalizes each indicator before participating in the weighted calculation to obtain a single score. Based on this score, the system sets three grading thresholds: a score below 0.35 is considered a low-level threat, between 0.35 and 0.7 is a medium-level threat, and above 0.7 is a high-level threat. To avoid frequent level switching caused by small fluctuations in ratings, the system introduces a hysteresis judgment strategy: a level change is only triggered when the rating level crosses two segments (such as jumping from low to high) and continues for more than a set time threshold (such as 300 seconds), ensuring the stability and reliability of the level division.

[0148] In a specific embodiment, reproducible operational results are given: within a certain altitude layer, the cluster center continuously approaches the confined area at a rate of approximately six meters per second over thirty seconds; the density gradient changes from stable to increasing and significantly increases within ten seconds; velocity consistency improves from moderate to high, and the velocity difference between adjacent aircraft continues to converge. The crosswind at ground level is approximately three meters per second, and the system deducts the small directional sway caused by it; the communication delay is eighty milliseconds, and the system relaxes synchronization constraints. After fusion, the confidence level of the intent to "coordinate contraction and approach sensitive targets" is significantly increased; the minimum boundary distance of the cluster center is lower than the safety upper limit, the velocity mutation exceeds the middle threshold, the temporal similarity with historical attack samples is high, the comprehensive score exceeds the high-level threshold, a high-level dynamic threat is output, and global heatmap focusing, path prediction, and adaptive radar beam configuration are triggered.

[0149] In a specific embodiment, the preset prior formation template is derived from historical formation data and training rule library, covering five categories: column, wedge, triangle, encirclement and barrier; for example, in the column cruise phase, the distance between adjacent aircraft is 10 to 15 meters, the heading difference is no more than 10 degrees, and the speed difference is no more than 1 meter per second; in the wedge engagement phase, the wingspan ratio is 1:1 to 1:1.5, the height difference between layers is no more than 5 meters, and the speed difference is no more than 1.5 meters per second; the template provides scaling rules according to the mission phase (the distance is shortened by 20% in the assembly phase and enlarged by 30% in the withdrawal phase) and holding time thresholds, which are loaded as priors before operation and the matching degree is calculated in real time.

[0150] In a specific embodiment, the preset wind speed disturbance information is jointly provided by ground meteorological stations, airspace meteorological models, and airborne wind field calculations, and a classification and compensation strategy is defined: light wind less than 3 m / s, light wind 3 to 8 m / s, strong wind 8 to 12 m / s, and gusts greater than 12 m / s; under light / light wind conditions, only the heading deviation is relaxed to a tolerance of 5 degrees / 10 degrees; under strong / gust conditions, additional tolerances are added to speed consistency and spacing determination, and wind shear correction is introduced to the lower atmosphere (adding a heading tolerance of 0.5 degrees and a speed tolerance of 0.2 m / s for every 10 meters of height difference); the wind field is updated every 5 to 10 seconds, and the most recent reliable value is used and weighted down when measurements are missing.

[0151] In one specific embodiment, the preset communication delay parameter obtains the round-trip delay through link detection and classifies it into levels: no more than 50 milliseconds is normal, 50 to 120 milliseconds is elevated, 120 to 250 milliseconds is degraded, and more than 250 milliseconds is critical. The normal level maintains standard synchronization and an evaluation window of three to five seconds. The elevated level relaxes the speed / heading consistency threshold by one level and expands the evaluation window to five to eight seconds. The degraded level introduces hysteresis in the consistency determination and reduces the sensitivity to instantaneous density increases. The critical level enters a conservative mode and only assigns high weight to behaviors that continuously approach sensitive areas. Each level has a hysteresis time of no less than ten seconds for entry / exit.

[0152] In a specific embodiment, the preset threat determination rule adopts a sub-item scoring and grading threshold: when the minimum distance is less than 500 meters and the approach speed is greater than 5 meters per second and lasts for more than 10 seconds, or when the density gradient continues to rise and the angle between the approach direction and the restricted area is less than 30 degrees, the comprehensive score enters the high level; if two of the above conditions are met but the duration is insufficient or the angle is between 30 and 45 degrees, it is recorded as the medium level; if only a single item has a short-term anomaly (less than 5 seconds), it is recorded as the low level; both the scoring and grading are set with a minimum holding time of not less than 15 seconds to avoid frequent level switching.

[0153] In step S17, when the dynamic threat level is higher than a preset threat threshold, a pre-generated global threat heatmap needs to be invoked, and high-threat areas need to be identified and determined by combining the hierarchical height range and the density gradient, including:

[0154] When the dynamic threat level is higher than the preset threat threshold, the pre-generated global threat heat map is invoked to divide the hierarchical height range into low-altitude, mid-altitude, and high-altitude layers.

[0155] For the low-altitude layer, the spatial distribution features of slow-moving low-altitude targets are extracted as anchor points for initial clustering to obtain low-altitude threat sub-regions.

[0156] By combining the density gradients corresponding to each layered height interval of the low-altitude threat sub-region with the layered height intervals, a secondary clustering analysis is performed on the UAV clusters in the mid-altitude and high-altitude layers to locate mid- and high-altitude target regions with concentrated density and consistent height.

[0157] The low-altitude threat sub-region and the medium- and high-altitude target region are spatially merged and their boundaries refined to determine the boundary range and spatial coordinate information of the high-threat region, thus obtaining the high-threat region.

[0158] In one specific embodiment, when the dynamic threat level exceeds the preset threat threshold, the system invokes the pre-generated global threat heatmap. The heatmap uses the frequency of target occurrence, velocity consistency, and local density within a fixed time window as weights, is grid-encoded, and uses color depth to represent threat strength. To facilitate hierarchical identification, the system maps the hierarchical height ranges to low-altitude, mid-altitude, and high-altitude layers (e.g., low-altitude no higher than 80 meters, mid-altitude 80 to 200 meters, and high-altitude higher than 200 meters). In urban canyons or ultra-low-altitude corridor scenarios, the low-altitude upper limit can be lowered. The spatial mapping process is considered complete only after all grid cells have completed threat intensity calculations and all target flight altitudes have been successfully assigned to the mapped layers. Subsequently, a binding index relationship is established between each heatmap grid and its corresponding height layer. After mapping, an index is created for the heatmap grid and its height layer, supporting hierarchical clustering and boundary processing.

[0159] In one specific embodiment, the system first performs anchor-point-based initial clustering in the low-altitude layer to quickly capture low-altitude slow-moving threats. Low-altitude slow-moving targets are determined using a dual-threshold strategy: targets are marked as candidate anchor points if their altitude is in the low-altitude layer, their ground speed is below a preset slow speed threshold (e.g., four meters per second), and their trajectory curvature and speed fluctuations are relatively small within a short window. The system then searches for neighborhood grid density and speed consistency centered on these candidate anchor points, completes initial clustering according to a preset neighborhood radius and minimum cluster size, and obtains low-altitude threat sub-regions. Only stable sub-regions that persist for at least a minimum retention time are retained, and attributes such as center coordinates, temporal stability, average altitude, and area are recorded as subsequent inputs.

[0160] In one specific embodiment, the system combines the density gradients corresponding to the low-altitude threat sub-regions and each layered height interval to perform secondary clustering analysis on the mid-altitude and high-altitude layers to identify mid-to-high-altitude target regions with concentrated density and high consistency. The secondary clustering follows the principles of "density priority, intra-layer consistency, and inter-layer correlation": dense clusters within a layer are identified based on a density threshold and the minimum number of cores, requiring that the height variance within a cluster does not exceed a preset consistency threshold; mid-to-high-altitude dense clusters that spatially overlap with or are nearly directionally aligned with the low-altitude threat sub-regions are prioritized as key candidates. To enhance robustness, the clustering radius can adaptively adjust with the density gradient: it is moderately widened when density increases rapidly to merge fragmented sub-groups, and tightened when changes are gradual to avoid excessive merging. In one specific embodiment, the clustering radius is adaptively adjusted based on the rate of change of the threat density gradient in each height layer. If the rate of increase in target density within a unit height interval exceeds 20% / 10 meters (i.e., a density slope greater than 0.02 / m), the system expands the cluster radius of the current layer to 1.5 times the original set value to merge sparsely distributed but highly uniform small groups. Conversely, if the rate of change in density is less than 5% / 10 meters (i.e., a density slope less than 0.005 / m), the cluster radius is reduced to 0.8 times the original set value to enhance the identification accuracy of local clusters. This adjustment process ensures that the clustering algorithm avoids missing high-density small groups and also prevents the merging of unrelated targets due to excessively large radii.

[0161] In one specific embodiment, the system performs spatial merging and boundary refinement on the low-altitude threat sub-regions and mid-to-high-altitude candidate regions, producing high-threat regions that can be directly used for command and control. Spatial merging first determines connectivity: regions with a horizontal distance less than a preset merging distance and an inter-layer height difference less than a preset inter-layer threshold are merged into the same threat entity. Boundary refinement then employs a combination of contour smoothing and concave polygon correction to remove jagged edges and isolated protrusions, while retaining tactical features such as narrow corridors. Finally, the boundaries are output as polygons or grid sets, and each high-threat region is accompanied by metadata such as center location, coverage area, main direction, altitude range, average density, and stability, serving as direct input for path prediction and radar beam reconfiguration.

[0162] In a specific embodiment, reproducible parameters and procedures are provided: When the dynamic threat level reaches a high level, a heatmap is invoked within a ten-second time window, with a grid resolution of ten meters and a low-altitude upper limit of eighty meters; the low-altitude slow speed threshold is four meters per second, the initial clustering neighborhood radius is thirty meters, the minimum cluster size is five aircraft, and the minimum holding time is five seconds. Under this setting, two stable sub-regions are identified in the low-altitude layer, and their centers and stability are recorded; subsequently, secondary clustering is performed in the mid-altitude and high-altitude layers according to the density thresholds of each layer and the minimum number of cores, and the radius of rapidly increasing density regions is automatically widened by twenty percent to avoid fragmentation; finally, a mid-altitude dense cluster that is consistent with the direction of the low-altitude sub-regions and horizontally overlaps is obtained. After merging, continuous high-threat regions are determined through boundary smoothing and concave correction, and the boundary coordinates, main direction, altitude range, and average density are output.

[0163] In one specific embodiment, the preset threat threshold is calibrated using historical samples: labeled threats and normal samples are collected, and a comprehensive score is calculated based on deviation magnitude, velocity mutation, trajectory similarity, density gradient, and approach direction. A threshold that balances false alarms and false negatives is selected. The threshold is divided into three levels (e.g., 60% / 80% are medium / high thresholds) and a five-minute hysteresis and a minimum hold time of no less than 15 seconds are set to suppress jitter. The triggering conditions are increased in strong wind or high latency scenarios and decreased in sensitive airspace or during major events. The threshold version is updated with configuration and seasonal baseline changes, and the effective time and evaluation report are recorded.

[0164] In one specific embodiment, the pre-generated global threat heatmap is constructed offline as a multi-resolution base map when the system is idle: the monitoring area is divided into 10-meter grids and three-layer heights, and no-fly zones and buffer zones, key target areas, terrain masking codes, and normal flight path baselines are loaded; the frequency of grid-level target occurrences, speed consistency, density gradient, and proximity are statistically analyzed on the data of the most recent seven days in different time periods, and baseline heat values ​​are formed according to weighted templates and then spatiotemporally smoothed to generate a heatmap library grouped by time period / weather / airspace configuration. When running online, the system calls the corresponding base map according to the current time period and configuration, and every five seconds, it overlays and corrects the real-time density and speed consistency increments within the 30-second sliding window, outputting the current global heatmap, including grid weights, main direction, height layer labels, and quality indicators, along with version numbers and construction parameters, for quick access in subsequent layered identification and resource scheduling.

[0165] In step S18, based on the high-threat area, it is necessary to integrate preset airspace safety constraints with the speed of the UAV to predict the future dynamic trajectory of the swarm and obtain airspace situation update data, including:

[0166] Extract the spatial distribution information of the high-threat area and combine it with the speed of the UAV to form an initial situation dataset;

[0167] If the initial situation dataset satisfies the preset airspace safety constraints, the speed of the UAV is iteratively calculated to obtain the trajectory prediction result of the future dynamics of the swarm.

[0168] Based on the trajectory prediction results and preset airspace safety constraints, safe path planning data is generated, situation assessment indicators are calculated, and airspace situation update data is output.

[0169] In one specific embodiment, the system forms an initial situational awareness dataset based on the spatial distribution information of the high-threat area and the current UAV speed. The high-threat area is represented by a grid or polygon, including boundary coordinates, coverage area, main direction, upper and lower altitude limits, average density, and stability. For each UAV, a timestamp, global position, ground speed, heading, rate of climb, speed confidence, and source identifier are provided. The system uses a unified time reference for alignment and coordinate unification, removes duplicate targets, extrapolates missing speeds or headings using the two most recent frames and adjusts their weights, and outputs a complete record containing "target identifier—position—speed—altitude layer—quality indicator" as prediction input.

[0170] In one specific embodiment, the system performs airspace safety constraint verification on the initial situation dataset. These constraints include no-fly zones and buffer zones, permissible corridor altitudes, minimum safe distances, maximum ground speed and maximum rate of climb, and minimum distances to control boundaries. Taking urban low-altitude airspace as an example, the minimum safe distance can be set at 50 meters, the maximum ground speed at 15 meters per second, the corridor altitude at 80 to 150 meters, and the buffer zone for key targets at 200 meters. Samples violating any constraint are preferentially pruned to the nearest legal altitude or speed; if still not satisfied, they are marked as unusable and removed, and an alarm entry is generated. The verified data serves as the prediction input set.

[0171] In one specific embodiment, the system performs trajectory prediction processing on UAV flight sample data that has passed preliminary screening and verification. This processing uses a fixed-length time window (e.g., 1520 seconds) as the prediction period and sets the prediction step size to 0.5 seconds to ensure fine temporal granularity. For each UAV, the system simultaneously generates 5 to 8 candidate trajectories based on its current position, velocity vector, and attitude information, combined with historical trajectory change trends, to cover possible minor maneuvers and wind disturbance deviations. The wind disturbance correction term is calculated based on wind speed and direction data provided in the current meteorological database (updated every 10 seconds), and the predicted position is corrected at each step through velocity vector superposition processing. The impact of communication delay is modeled by setting a delay hold-time parameter (e.g., 200ms), during which the current command state of the UAV remains unchanged. The system calculates all candidate trajectories for each prediction step based on the following three scoring indicators: (1) whether the safe distance from obstacles and nearby aircraft meets the minimum separation requirement (e.g., more than 15 meters), (2) whether it conforms to the current group's intended path direction (e.g., heading deviation less than 10 degrees), and (3) whether the trajectory's speed and altitude are consistent with the current cluster center's average value (deviation within ±5%). Candidate trajectories with a score below 0.6 are directly eliminated, and trajectories with a similarity higher than 90% are merged through vector distance calculation, outputting 12 representative prediction trajectories. Finally, based on the high-confidence prediction trajectories of all UAVs, the system calculates the future flight envelope boundary at the group level, the predicted cluster center path curve, and the target distribution ratio in each altitude region, providing input data for subsequent situation assessment and path optimization.

[0172] In one specific embodiment, the system generates safe path planning data and calculates situation assessment indicators based on prediction results and airspace safety constraints, forming airspace situation update data. Path planning prioritizes avoiding high-threat areas and buffer zones, selecting turning points and altitude layer switching points while meeting minimum spacing and allowable altitude requirements, and outputs identifiable waypoint sequences and speed commands. Situation assessment indicators include at least threat proximity (the closest distance and approach speed between the future envelope and the high-threat boundary), path deviation (spatial and temporal deviation relative to the established route), airspace congestion (the number of predicted targets per unit volume and the rate of change), conflict predictions (the logarithm of possible spacing violations within the future time window), and feasibility confidence (given by a combination of data quality and model consistency). The system packages and publishes "waypoints and command sets, assessment indicators, effective time windows, version numbers, and data quality flags" for resource scheduling and adaptive radar beam calling.

[0173] In a specific embodiment, the parameters and results are shown as follows: the high-threat area is a 600 x 400 meter polygon, mainly oriented northeast, with an altitude of 80 to 150 meters; the average ground speed of the swarm is 5 meters per second, the minimum spacing is 50 meters, the buffer zone is 200 meters, the prediction time window is 15 seconds, the step size is 0.5 seconds, and the number of candidate trajectories is 800 per drone. After verification, 30 drones entered the prediction; the cluster center moved approximately 90 meters northeast within 15 seconds, and the proportion of low-altitude layer increased from 60% to 70%. The path planning detours around the west side of the high-threat area, suggesting an altitude increase from 120 meters to 160 meters, generating eight waypoints and corresponding speed commands. The evaluation shows: the closest distance is 180 meters, the approach speed is 3 meters per second, the path spatial deviation is 35 meters, there are two potential conflicts in the next 10 seconds, and the congestion level is moderate; based on this, the system outputs a "restricted and controllable" situation level, providing a 15-second effective time window and the trigger conditions for the next round of refresh.

[0174] In one specific embodiment, the preset airspace safety constraints include three types of constraints: geometric constraints, motion constraints, and operating environment constraints.

[0175] First, there are spatial geometric constraints. Based on the electronic fence database and standard route layers, the system extracts no-fly zones, restricted-fly zones, and passable corridors, and automatically generates extended buffer zones for specific sensitive ground targets. The system overlays minimum safe flight altitude constraints based on terrain undulations and obstacle heights, and finally, through a priority rule of "no-fly zone first, buffer zone second, corridor last," overlaps and trims each area to generate the usable airspace boundary for the current version, marking its effective time and version number. Second, there are kinematic and spacing constraints. Based on the UAV platform type and mission characteristics, the system sets maximum ground speed, maximum rate of climb, maximum rate of descent, and maximum permissible turning angular velocity for each UAV flight. Simultaneously, the system sets minimum horizontal and minimum vertical spacing; if the flight path is located in a corridor area, additional constraints are imposed to ensure that lateral and altitude deviations do not exceed set thresholds. When multiple UAVs are operating in formation or flying in high-density airspace, the system performs consistency checks on spacing between the same and different levels. The third type is adaptive constraints under operating environment conditions. The system supports automatically switching between temporary notices to air traffic and route usage rules based on the operating time period. When link latency exceeds a preset threshold or positioning accuracy deteriorates, the system automatically switches to conservative flight mode, reducing the maximum flight speed and increasing the minimum interval threshold. In windy environments, the system adjusts the speed limit and tightens interval restrictions based on wind intensity levels under strong wind or gust conditions, and sets a wind shear tolerance coefficient to improve stability. All of the above safety constraints support multi-scenario template configuration and hysteresis strategies. Each constraint records its effective time, triggering conditions, and rollback logic. Any candidate trajectory that violates any constraint will be directly eliminated by the system or undergo boundary pruning and re-evaluation.

[0176] In step S19, it is necessary to update the data based on the airspace situation, configure the radar beam parameters, and reconstruct the coverage area to obtain real-time UAV detection data, including:

[0177] The trajectories of threat targets in the airspace situation update data are smoothed and matched with a preset penetration intent template to determine the set of penetration targets;

[0178] The azimuth and elevation angles of the target set are calculated to adjust the azimuth and elevation parameters of the radar beam, thus obtaining the beam parameters.

[0179] Based on the beam parameters, the scanning frame period is dynamically adjusted to extract the detection data of the core point coverage area and the boundary point coverage area, respectively.

[0180] The detection data is synchronized, filtered, and fused to update the target trajectory database and generate real-time detection data from the UAV.

[0181] In one specific embodiment, the system first reads the trajectories of targets marked as key targets for monitoring from the airspace situation update data, performs forward-backward smoothing on the position and velocity sequences, and downweights low-quality samples. Then, it matches the velocity magnitude, direction changes, and trends approaching restricted areas with a preset penetration intent template; the template covers patterns such as high-speed direct approach, serpentine approach, and formation shrinking approach, and sets minimum duration and minimum similarity thresholds. Targets meeting the thresholds are included in the penetration target set, and the matching score, duration, and confidence level are recorded for subsequent beam scheduling priority.

[0182] In one specific embodiment, the system calculates the relative azimuth and elevation angles for each target in the set of targets. First, the system transforms the position vectors of each target in the global geographic coordinate system to the radar body coordinate system using real-time attitude data from the inertial navigation system and the radar platform. During the coordinate transformation, installation angle extrinsic parameters (including pitch, yaw, and roll angles) are introduced for correction to ensure that the mapped position vectors accurately represent the spatial positions of the targets in the radar coordinate system. Subsequently, the system calculates the relative azimuth (i.e., horizontal angle) and elevation (i.e., elevation angle) based on the position vectors, where the azimuth is the angle projected onto the horizontal plane, and the elevation angle is the vertical angle relative to the horizontal plane. Based on the aforementioned angle data, the beam center position is determined, and the adaptive beamwidth is further calculated based on target density and maneuverability. If the number of targets within a unit solid angle range in a specified sector exceeds a density threshold, and the rate of change of the target's historical trajectory is lower than the threshold, the system classifies it as a "dense, low-maneuverability scenario," narrowing the beam to improve angular resolution. If the target distribution is sparse or the rate of change of acceleration and steering angle in a short period of time exceeds a set threshold, it is identified as a "sparse, high-maneuverability scenario," widening the beam to increase fault tolerance and reduce the probability of tracking loss. In the case of multiple targets, the system sorts all targets according to a comprehensive threat index, which is a weighted combination of threat level (preset classification), radial approach speed (calculated from radar ranging rate of change), and current minimum distance (radar ranging result). After sorting, the system dynamically allocates pulse slots and scanning resources to priority targets. If there are multiple targets with similar threat levels and azimuth differences less than a preset threshold, the system automatically merges them into a single sector and uniformly sets the beam center angle and scanning sector to reduce antenna switching overhead and improve scheduling efficiency. In one specific embodiment, the system dynamically adjusts the scanning frame period and dwell time based on beam parameters and reconstructs the coverage area. The coverage area is divided into core point coverage area and boundary point coverage area: the core area is used for high-frequency continuous tracking, and the boundary area covers the high-threat extension and potential penetration corridors. The scheduling strategy is "high frequency for the core, minimum for the boundary, and full-area inspection": the core has a shorter frame period and a moderately increased dwell time, the boundary is swept with a longer frame period, and the entire area is inspected with a low duty cycle. When the core's target decreases or stabilizes within a set time window, time slots are reclaimed and allocated to the boundary and inspection; when the boundary shows a continuous increase in density or a new penetration mode, its priority is temporarily increased and its frame period is shortened.

[0183] In one specific embodiment, after scanning, the original detection results of the core and boundary coverage are collected separately, and synchronization, filtering, and fusion are performed to update the target trajectory database. Synchronization unifies multi-beam and multi-channel data to the same time reference and coordinate system; filtering first performs constant false alarm threshold and clutter suppression, followed by measurement gating and anomaly removal; fusion adopts measurement-trajectory association and updating: the position and velocity of existing trajectories are corrected, new targets that continuously meet the confirmation conditions are established into tracks, and targets that are continuously lost are either maintained or terminated. The trajectory database maintains a state vector, covariance, source weight, last update time, and threat label for each trajectory.

[0184] In one specific embodiment, the system encapsulates the updated trajectory along with the quality indicators of the current scan into real-time UAV detection data for publication. The published content includes at least the target number, timestamp, global position and altitude, velocity magnitude and direction, trajectory continuity flag, coverage level, penetration matching score, current threat level, data quality flag, and version number. To control bandwidth, only targets with status changes exceeding thresholds or newly confirmed targets are pushed immediately; the rest are packaged at fixed intervals, and a recent incremental cache is retained for repackaging and reconstruction.

[0185] In a specific embodiment, reproducible operational results are presented: When a high-threat area appears in the southeast, the system identifies three high-speed, directly approaching trajectories as penetration targets within a five-second time window; the azimuth angle is calculated to be 200-220 degrees and the elevation angle to be 6-10 degrees, and beam centers for two overlapping sectors are set accordingly; the frame period in the core area is shortened from one second to 0.5 seconds, the boundary is maintained at one second, and the overall inspection is three seconds. After synchronization, filtering, and fusion, the speed jitter of the three trajectories is significantly reduced and remains continuously locked; subsequently, the density at the outer edge of the boundary increases, the priority of the two trajectories triggering the boundary is increased, and the boundary frame period is temporarily tightened to 0.75 seconds. The entire trajectory library is continuous and without breaks.

[0186] In one specific embodiment, the preset penetration intent template is a priori pattern library of typical penetration behaviors, parameterized by motion and formation characteristics, covering high-speed direct approach, serpentine approach, formation contraction approach, low-altitude obstacle-alongside penetration, and rapid ascent crossing, etc. For each type, criteria such as velocity threshold, acceleration duration, track curvature, pointing angle of relatively sensitive areas, minimum duration, minimum number of sample points, formation compactness, inter-layer altitude difference, and speed consistency are provided. At the same time, wind field and communication delay tolerance corrections are configured to avoid misjudging weather or link jitter as penetration. The template is managed offline by statistical analysis of historical threat samples and training data, and is versioned. The similarity and confidence of each trajectory are calculated online. If the similarity exceeds the threshold, it is marked as a penetration candidate, and the hold time and matching score are output for direct use in dynamic threat assessment and beam scheduling.

[0187] In summary, this invention provides a method and system for detecting drone swarms based on multi-layer early warning, which solves the problem that existing technologies are unable to obtain accurate three-dimensional location information of swarms in real time in complex dynamic environments and to accurately identify layered threats by combining height and density.

[0188] Reference Figure 2 The second embodiment of the present invention provides a drone swarm detection system based on multi-layer early warning, comprising:

[0189] The data acquisition module is used to acquire multi-source sensor data, including radar and photoelectric signals, as well as the speed of the UAV.

[0190] The spatial mapping module is used to map the multi-source sensor data to a unified spatial coordinate system based on a preset state transition matrix and a preset observation matrix to obtain an initial dataset;

[0191] The dynamic trajectory module is used to correct the position based on the initial dataset and the preset measurement noise parameter set to obtain dynamic trajectory data that eliminates measurement deviations;

[0192] The height stratification module is used to fit the three-dimensional position in the dynamic trajectory data and remove outliers, perform segmented statistics according to a preset height granularity and extract height statistical features, and perform clustering based on the height statistical features to obtain the stratified height intervals.

[0193] The density recognition module is used to identify high-density clustering areas and obtain the cluster center location and density gradient when the number of drones in the layered height range exceeds a preset density threshold.

[0194] The dynamic threat module is used to construct a multidimensional Bayesian network by fusing the cluster center location, density gradient, and UAV speed with a preset prior formation template, preset wind speed disturbance information, and preset communication delay parameters, and infer the potential threat behavior of the UAV swarm based on the multidimensional Bayesian network to obtain the dynamic threat level.

[0195] The threat template module is used to call a pre-generated global threat heatmap when the dynamic threat level is higher than a preset threat threshold, and to identify and determine high-threat areas by combining the hierarchical height range and the density gradient.

[0196] The airspace situation module is used to predict the future dynamic trajectory of the swarm based on the high-threat area, by integrating preset airspace security constraints and the speed of the UAV, and to obtain airspace situation update data.

[0197] The detection data module is used to update data based on the airspace situation, configure radar beam parameters and reconstruct the coverage area to obtain real-time detection data from the UAV.

[0198] It should be noted that the UAV swarm detection system based on multi-layer early warning provided in this embodiment of the invention is used to execute all the process steps of the UAV swarm detection method based on multi-layer early warning in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0199] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a detection data program. When the processor executes the computer program, it implements the steps described in the various embodiments of the multi-layered early warning-based UAV swarm detection method, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above-described device embodiments, such as the data detection module.

[0200] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0201] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0202] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0203] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0204] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0205] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0206] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for detecting drone swarms based on multi-layer early warning, characterized in that... include: Acquire multi-source sensor data, including radar and photoelectric signals, and the speed of the UAV; Based on the preset state transition matrix and the preset observation matrix, the multi-source sensor data are mapped to a unified spatial coordinate system to obtain the initial dataset; The position is corrected based on the initial dataset and the preset measurement noise parameter set to obtain dynamic trajectory data that eliminates measurement deviations; The three-dimensional position in the dynamic trajectory data is fitted and outlier removal is performed. The data is then segmented and statistically analyzed according to a preset height granularity, and height statistical features are extracted. Clustering is then performed based on the height statistical features to obtain hierarchical height intervals. When the number of drones in the layered height range exceeds the preset density threshold, high-density clustering areas are identified to obtain the cluster center location and density gradient. Based on the cluster center location, density gradient, and UAV speed, a preset prior formation template, preset wind speed disturbance information, and preset communication delay parameters are fused to construct a multidimensional Bayesian network. The potential threat behavior of the UAV swarm is inferred from the multidimensional Bayesian network to obtain the dynamic threat level. When the dynamic threat level is higher than the preset threat threshold, a pre-generated global threat heatmap is invoked, and high-threat areas are identified and determined by combining the hierarchical height range and the density gradient. Based on the high-threat area, the preset airspace safety constraints and the speed of the UAV are combined to predict the future dynamic trajectory of the swarm and obtain airspace situation update data. Based on the updated airspace situation data, the radar beam parameters are configured and the coverage area is reconstructed to obtain real-time UAV detection data.

2. The UAV swarm detection method based on multi-layer early warning according to claim 1, characterized in that, The initial dataset, obtained by mapping the multi-source sensor data to a unified spatial coordinate system based on a preset state transition matrix and a preset observation matrix, includes: The range, azimuth, and elevation angles in the multi-source sensor data are denoised and smoothed, and then standardized to obtain a radar measurement dataset. Lens distortion correction is performed on the pixel coordinates of photoelectric imaging in the multi-source sensor data, and the contour and texture features of the target are extracted to obtain a photoelectric image dataset. The radar measurement dataset and the photoelectric image dataset are synchronized in time. If the time difference between the radar measurement dataset and the photoelectric image dataset is greater than a preset time difference threshold, the missing data is filled in to obtain a synchronized fusion input dataset. The synchronous fusion input dataset is input into a preset state transition matrix and a preset observation matrix to calculate the fusion state vector; Based on the fused state vector, the coordinate system of the multi-source sensor data is converted into a global spatial coordinate system to obtain the initial dataset of the UAV swarm.

3. The UAV swarm detection method based on multi-layer early warning according to claim 1, characterized in that, The step of correcting the position based on the initial dataset and the preset measurement noise parameter set to obtain dynamic trajectory data after removing measurement bias includes: An initial state vector containing position and velocity information is extracted from the initial dataset, and a predicted state vector and its corresponding covariance matrix are calculated by combining the preset state transition matrix to obtain the predicted state dataset. Based on the predicted state dataset and the initial dataset, a noise covariance matrix is ​​calculated using a preset set of measurement noise parameters. The observation residuals are calculated based on the predicted state vector and the initial dataset, and the filter gain is calculated using a preset formula. When the observed residual is greater than a preset residual threshold, the state vector is corrected according to the filter gain and the covariance matrix to obtain the corrected state dataset; The corrected state dataset is combined with a preset state transition matrix to generate dynamic trajectory data. Interpolation and smoothing are then performed on the dynamic trajectory data to remove instantaneous measurement deviations, thus obtaining the dynamic trajectory data.

4. The UAV swarm detection method based on multi-layer early warning according to claim 1, characterized in that, The process involves fitting and outlier removal of the three-dimensional positions in the dynamic trajectory data, segmenting and statistically extracting height statistical features according to a preset height granularity, and performing clustering based on the height statistical features to obtain hierarchical height intervals, including: The three-dimensional position values ​​in the dynamic trajectory data are smoothed, and the state vector and the corresponding covariance matrix in the dynamic trajectory data are combined with the state vector in the dynamic trajectory data for iterative update to obtain the smoothed first trajectory data. The trajectory data is fitted using spline interpolation, and the trajectory deviation between the trajectory points and the preset interpolation curve is detected. When the trajectory deviation is greater than the preset trajectory deviation threshold, the corresponding trajectory point is removed to obtain the second trajectory data after removing the abnormal points; The second trajectory data is segmented and statistically analyzed according to a preset height granularity. The mean, variance, kurtosis and density gradient of each height interval are calculated to form a height statistical feature set. The clustering distance threshold is determined based on the mean and density gradient of the height statistical feature set, and clustering is performed based on the clustering distance threshold to obtain the hierarchical height interval.

5. The UAV swarm detection method based on multi-layer early warning according to claim 1, characterized in that, When the number of drones within the layered height range exceeds a preset density threshold, high-density clustering areas are identified, and the cluster center location and density gradient are obtained, including: When the number of drones in the layered height interval is greater than a preset density threshold, the height interval corresponding to the number of drones greater than the preset density threshold is marked as the height interval to be processed, and the set of height intervals to be processed is obtained. For each height interval in the set of height intervals to be processed, density clustering is performed based on a preset distance radius and a preset minimum number of core points to divide high-density clustering regions and obtain clusters; Based on the clusters, calculate the center position and density gradient of each cluster to obtain the cluster center position and corresponding density gradient.

6. The UAV swarm detection method based on multi-layer early warning according to claim 1, characterized in that, The process involves fusing the cluster center location, density gradient, and UAV speed with a preset prior formation template, preset wind speed disturbance information, and preset communication delay parameters to construct a multidimensional Bayesian network. Based on this multidimensional Bayesian network, the potential threat behavior of the UAV swarm is inferred to obtain a dynamic threat level, including: Based on the cluster center location, the density gradient, and the drone speed, a multidimensional Bayesian network containing the relationship between location, speed, and density is constructed. The multidimensional Bayesian network is fused with a preset prior formation template, preset wind speed disturbance information, and preset communication delay parameters to obtain the updated swarm cooperative formation intention; Based on the swarm's cooperative formation intent and combined with preset threat determination rules, the deviation range, speed change degree, and similarity to the expected trajectory of the drone swarm are comprehensively evaluated, and the dynamic threat level is output according to the evaluation results.

7. The UAV swarm detection method based on multi-layer early warning according to claim 1, characterized in that, When the dynamic threat level exceeds a preset threat threshold, a pre-generated global threat heatmap is invoked, and high-threat areas are identified and determined by combining the hierarchical height range and the density gradient, including: When the dynamic threat level is higher than the preset threat threshold, the pre-generated global threat heatmap is invoked to divide the hierarchical height range into the low-altitude layer, the mid-altitude layer, and the high-altitude layer. For the low-altitude layer, the spatial distribution features of slow-moving low-altitude targets are extracted as anchor points for initial clustering to obtain low-altitude threat sub-regions. By combining the density gradients corresponding to each layered height interval of the low-altitude threat sub-region with the layered height intervals, a secondary clustering analysis is performed on the UAV clusters in the mid-altitude and high-altitude layers to locate mid- and high-altitude target regions with concentrated density and consistent height. The low-altitude threat sub-region and the medium- and high-altitude target region are spatially merged and their boundaries refined to determine the boundary range and spatial coordinate information of the high-threat region, thus obtaining the high-threat region.

8. The UAV swarm detection method based on multi-layer early warning according to claim 1, characterized in that, The process involves, based on the high-threat area, integrating preset airspace safety constraints with the drone speed, predicting the future dynamic trajectory of the swarm, and obtaining airspace situation update data, including: Extract the spatial distribution information of the high-threat area and combine it with the speed of the UAV to form an initial situation dataset; If the initial situation dataset satisfies the preset airspace safety constraints, the speed of the UAV is iteratively calculated to obtain the trajectory prediction result of the future dynamics of the swarm. Based on the trajectory prediction results and preset airspace safety constraints, safe path planning data is generated, situation assessment indicators are calculated, and airspace situation update data is output.

9. The UAV swarm detection method based on multi-layer early warning according to claim 1, characterized in that, The step of updating radar beam parameters and reconstructing coverage area based on the airspace situation update data to obtain real-time UAV detection data includes: The trajectories of threat targets in the airspace situation update data are smoothed and matched with a preset penetration intent template to determine the set of penetration targets; The azimuth and elevation angles of the target set are calculated to adjust the azimuth and elevation parameters of the radar beam, thus obtaining the beam parameters; Based on the beam parameters, the scanning frame period is dynamically adjusted to extract the detection data of the core point coverage area and the boundary point coverage area, respectively. The detection data is synchronized, filtered, and fused to update the target trajectory database and generate real-time detection data from the UAV.

10. A drone swarm detection system based on multi-layer early warning, characterized in that, include: The data acquisition module is used to acquire multi-source sensor data, including radar and photoelectric signals, as well as the speed of the UAV. The spatial mapping module is used to map the multi-source sensor data to a unified spatial coordinate system based on a preset state transition matrix and a preset observation matrix to obtain an initial dataset; The dynamic trajectory module is used to correct the position based on the initial dataset and the preset measurement noise parameter set to obtain dynamic trajectory data that eliminates measurement deviations; The height stratification module is used to fit the three-dimensional position in the dynamic trajectory data and remove outliers, perform segmented statistics according to a preset height granularity and extract height statistical features, and perform clustering based on the height statistical features to obtain the stratified height intervals. The density recognition module is used to identify high-density clustering areas and obtain the cluster center location and density gradient when the number of drones in the layered height range exceeds a preset density threshold. The dynamic threat module is used to construct a multidimensional Bayesian network by fusing the cluster center location, density gradient, and UAV speed with a preset prior formation template, preset wind speed disturbance information, and preset communication delay parameters, and infer the potential threat behavior of the UAV swarm based on the multidimensional Bayesian network to obtain the dynamic threat level. The threat template module is used to call a pre-generated global threat heatmap when the dynamic threat level is higher than a preset threat threshold, and to identify and determine high-threat areas by combining the hierarchical height range and the density gradient. The airspace situation module is used to predict the future dynamic trajectory of the swarm based on the high-threat area, by integrating preset airspace security constraints and the speed of the UAV, and to obtain airspace situation update data. The detection data module is used to update data based on the airspace situation, configure radar beam parameters and reconstruct the coverage area to obtain real-time detection data from the UAV.

Images