A method and system for airspace safety warning

Abstract

This invention discloses an airspace safety early warning method and system, relating to the field of unmanned aerial vehicle (UAV) flight safety technology. The method includes: acquiring multi-source heterogeneous data from UAVs; preprocessing the multi-source heterogeneous data; constructing an intelligent collision prediction model based on artificial intelligence algorithms, wherein the intelligent collision prediction model performs real-time analysis of the preprocessed data and outputs a collision risk probability; and sending an early warning message to staff when the collision risk probability reaches a preset threshold. This invention solves the problems of incomplete airspace data coverage and difficulty in heterogeneous compatibility, delayed risk identification, and poor collision detection adaptability. It achieves full-domain collaborative integration of multi-source data, pre-judgment of collision risks, improves the accuracy of collision prediction, outputs precise risk probabilities, adapts to multiple targets and complex airspaces, and meets the needs of low-altitude safety management.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) flight safety technology, and more specifically to an airspace safety early warning method and system. Background Technology

[0002] With the rapid development of the low-altitude economy, the number of aircraft in low-altitude airspace has exploded, and airspace flight activities have become increasingly frequent, posing a severe challenge to airspace safety management. Currently, the main pain points in low-altitude airspace management are as follows: First, the airspace risk identification is significantly delayed. Existing management methods rely heavily on traditional means such as manual patrols and post-event tracing, making it difficult to detect violations such as unlicensed flights, flights outside designated areas, and speeding in real time. Verification is often only carried out after an accident occurs, failing to achieve proactive risk prevention. Second, the accuracy of flight collision warnings is insufficient. Low-altitude airspace aircraft are diverse in type and have flexible flight trajectories. Existing monitoring systems often use single data sources (such as relying solely on GPS positioning or localized radar monitoring), resulting in incomplete data coverage and poor spatiotemporal synchronization. This leads to low accuracy in predicting dynamic conflicts involving multiple aircraft, making it difficult to accurately determine the probability of collision risk and the warning time window. The lack of a unified spatial positioning benchmark and efficient data fusion technology further hinders the timeliness and accuracy of warnings, failing to meet the actual needs of low-altitude airspace safety management.

[0003] Therefore, in view of the shortcomings of existing technologies, how to provide an airspace safety early warning method and system is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, the present invention provides an airspace safety early warning method and system. The present invention solves the problems of incomplete airspace data coverage and difficulty in heterogeneous compatibility, delayed risk identification, and poor collision detection adaptation. It realizes the full-domain collaborative integration of multi-source data, predicts collision risks in advance, improves the accuracy of collision prediction, outputs accurate risk probabilities, adapts to multiple targets and complex airspace, and meets the needs of low-altitude safety management.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an airspace safety early warning method, comprising: Acquire multi-source heterogeneous data from drones. The multi-source heterogeneous data is preprocessed; An intelligent collision prediction model is constructed based on artificial intelligence algorithms. The intelligent collision prediction model performs real-time analysis on preprocessed data and outputs the probability of collision risk. When the probability of a collision reaches a preset threshold, an early warning message is sent to staff.

[0006] Preferably, the multi-source heterogeneous data includes UAV body data, airspace dynamic data, and environmental auxiliary data; the UAV body data includes the UAV's BeiDou grid code positioning data, flight plan data, and flight status data. The airspace dynamic data includes ADS-B data and radar data; The environmental support data includes satellite remote sensing data, aerial monitoring data, and ground sensor network data.

[0007] Preferably, the intelligent collision prediction model includes an input layer, a feature fusion layer, a prediction evaluation layer, and an output layer connected in sequence; The prediction and evaluation layer includes a trajectory prediction submodule, a collision detection submodule, and a risk assessment submodule, which are used to perform prediction, detection, and evaluation calculations on the feature fusion data.

[0008] Preferably, the trajectory prediction submodule adopts an LSTM and GRU network architecture, with 128 neurons in the LSTM hidden layer and 64 neurons in the GRU hidden layer, and the activation function is the tanh function; The collision detection submodule integrates OBB bounding box, swirling region modeling and coordinate collision analysis technologies to determine whether the drone has entered the area. The risk assessment submodule uses an SVM model. The collision risk index is input into the trained SVM model, which outputs the collision risk level. The collision risk level is then converted into a collision risk probability through probability calibration.

[0009] Preferably, the collision detection submodule integrates OBB bounding box, spiral region modeling, and coordinate collision analysis technologies, including: For both drones and stationary obstacles, select m boundary points in the reduced 2D coordinates. Calculate the mean and covariance matrices of X and Y coordinates, solve for eigenvalues ​​and eigenvectors to determine the bounding box orientation, establish a new coordinate system, calculate the length, width and center point, and form a bounding box that fits the UAV and obstacle targets. For moving obstacle targets, the center and radius of the loop are determined based on historical trajectories to construct a circular danger zone and determine whether the drone enters it.

[0010] Preferably, the shortest distance between the UAV's OBB bounding box and the stationary obstacle target's OBB bounding box is calculated, and when the shortest distance is less than a first threshold, it is determined to be a potential static collision. Based on the trajectory prediction results, the positional relationship between the two objects is calculated over a period of time in the future. If there is an overlap of the OBB bounding boxes or if the objects enter a circular danger zone, it is determined to be a dynamic collision.

[0011] Preferably, the calculation process for the collision risk index includes: ; in, As weight; The minimum safe distance threshold, This is the actual distance. For the maximum permissible relative speed, This is the actual relative speed. For the maximum permissible attitude angle deviation, This represents the actual attitude angle deviation. Visibility threshold For actual visibility, The wind speed threshold, This represents the actual wind speed.

[0012] Preferably, an airspace safety early warning system includes: The data acquisition module is used to acquire multi-source heterogeneous data from the UAV; The data preprocessing module is used to preprocess the multi-source heterogeneous data; The collision monitoring module is used to build an intelligent collision prediction model based on artificial intelligence algorithms. The intelligent collision prediction model analyzes the preprocessed data in real time and outputs the collision risk probability. The early warning module is used to send warning information to staff when the probability of collision risk reaches a preset threshold.

[0013] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses an airspace safety early warning method and system. The present invention solves the problems of incomplete airspace data coverage and difficulty in heterogeneous compatibility, delayed risk identification, and poor collision detection adaptation. It realizes the full-domain collaborative integration of multi-source data, predicts collision risks in advance, improves the accuracy of collision prediction, outputs accurate risk probabilities, adapts to multiple targets and complex airspaces, and meets the needs of low-altitude safety management. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of an airspace safety early warning method provided by the present invention.

[0016] Figure 2 This is a schematic diagram of an airspace safety early warning system provided by the present invention. Detailed Implementation

[0017] 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.

[0018] This invention discloses an airspace safety early warning method, such as... Figure 1 As shown, it includes: Acquire multi-source heterogeneous data from drones to achieve accurate acquisition of multi-source heterogeneous data from drones; The multi-source heterogeneous data is preprocessed; An intelligent collision prediction model is constructed based on artificial intelligence algorithms. The intelligent collision prediction model performs real-time analysis on preprocessed data and outputs the probability of collision risk. When the probability of a collision reaches a preset threshold, a warning message is sent to staff via SMS, system app push notifications, or other means.

[0019] Specifically, the multi-source heterogeneous data includes UAV body data, airspace dynamic data, and environmental auxiliary data; the UAV body data includes the UAV's BeiDou grid code positioning data, flight plan data, and flight status data. The airspace dynamic data includes ADS-B data and radar data, achieving comprehensive coverage of dynamic data of all aircraft within the airspace. The environmental support data includes satellite remote sensing data, aerial monitoring data, and ground sensor network data. This provides environmental and scenario support for collision risk assessment.

[0020] Specifically, the BeiDou grid code positioning data includes latitude and longitude, altitude, positioning reliability, and flight attitude data; the flight attitude data includes roll angle, pitch angle, and yaw angle. The flight plan data includes preset route coordinate sequences, flight altitude ranges, mission duration, payload information, no-fly zone avoidance list, and emergency return-to-base plan data; The drone's status data includes vibration amplitude during flight, remaining battery power, motor operating status, and battery temperature data.

[0021] The collection of BeiDou grid code positioning data includes: the UAV is equipped with a dual-mode positioning system of BeiDou and GPS to collect BeiDou grid code positioning data in real time, including parameters such as latitude and longitude, altitude, positioning reliability, and flight attitude. The positioning system has an anti-interference algorithm built in to resist low-altitude electromagnetic interference and ensure positioning stability in complex environments.

[0022] The UAV flight plan data is imported into the airspace BIM (Building Information Modeling) dynamic model. The flight plan data collection includes: preset flight path coordinate sequences, flight altitude ranges (minimum / maximum flight altitudes), mission duration, payload information, no-fly zone avoidance list, and emergency return-to-base plan. The flight plan data is stored in XML format, maintaining consistency with the coordinate system of the airspace BIM dynamic model, and supports real-time access and modification.

[0023] The collection of drone status data includes: carrying vibration sensors, power sensors, motor speed sensors, and battery temperature sensors to collect data such as vibration amplitude, remaining power, motor operating status, and battery temperature in real time during flight, in order to determine the impact of the drone's own status on collision risk.

[0024] Specifically, broadcast data from large UAVs in the airspace is received through the ADS-B ground station. The ADS-B data includes parameters such as aircraft identification code, real-time position, flight altitude, cruise speed, heading angle, and rate of climb.

[0025] By integrating data from air traffic control secondary radar and low-altitude surveillance radar, and targeting low-altitude, slow-moving, and small targets and obstacles, FMCW (Frequency Modulated Continuous Wave) radar technology is used to supplement the coverage blind spots of the ADS-B system. The radar data includes the target obstacle's distance, azimuth angle, and radial velocity.

[0026] Specifically, high-resolution optical satellites and synthetic aperture radar (SAR) satellites are used to acquire static airspace information, including airspace restricted area boundaries (airport clear zones, military control zones), topographic data (mountain slopes, urban building heights), and surface vegetation cover, which are then transmitted to ground data centers via satellite communication links.

[0027] A swarm of multi-rotor UAVs is used as aerial monitoring nodes, equipped with visible light cameras, thermal imagers, and lidar to collect dynamic airspace environmental data: visible light cameras are used to identify low-altitude floating objects (balloons, kites, etc.), thermal imagers are used to detect abnormal heating of aircraft engines, and lidar is used to construct a three-dimensional terrain model of the airspace.

[0028] Ground-based sensors, including wind speed sensors, visibility sensors, and precipitation sensors, are deployed at key airspace nodes (around airports, on top of urban high-rise buildings, and at mountain passes). The data is transmitted via LoRa wireless communication technology and used to correct environmental impact factors in aircraft trajectory prediction.

[0029] Specifically, the preprocessing of the multi-source heterogeneous data includes: Data standardization processing: Heterogeneous data such as BeiDou positioning data, ADS-B data, radar data, and sensor data are uniformly converted into JSON format; all spatial data are converted into the WGS-84 coordinate system, and coordinate deviations from different data sources are corrected through coordinate transformation algorithms to ensure that the position data of UAVs and obstacles are comparable under the same coordinate system; For image data acquired by lidar and visible light cameras, a cubic polynomial of univariate form is used to construct pixel distance. A fitting model to the actual distance L; ; Wherein, A, B, C, and D are coefficients calibrated through samples, with a fitting error ≤2%, achieving accurate conversion of the target size in the image to the actual size.

[0030] Data denoising and anomaly correction: For satellite remote sensing images and UAV visible light images, a 3×3 neighborhood mean filter is used. The center pixel value is replaced by the average value of the eight neighboring pixels around the pixel to remove salt-and-pepper noise. For thermal imaging images, a median filter is used. The neighboring pixel values ​​are sorted and the median value is taken to preserve image edge details and remove isolated bright spot noise. Kalman filtering was applied to BeiDou positioning data, velocity data, wind speed data, etc., to establish state equations and observation equations; The state equation is expressed as ; The observation equation is expressed as ; Where A is the state transition matrix, H is the observation matrix, and W... k and V k The noise is Gaussian white noise. The optimal estimate is dynamically calculated to remove random noise caused by electromagnetic interference.

[0031] Images acquired by visible light cameras are converted to grayscale, Gaussian blurred, and subjected to morphological opening and closing operations to remove image noise caused by ambient light and fog, laying the foundation for subsequent detection.

[0032] Data dimensionality reduction and feature extraction processing: PCA Feature Extraction: For the 3D coordinate data (X, Y, Z) acquired by LiDAR, the PCA dimensionality reduction algorithm is used. Standardize the 3D coordinate data: ; in, The mean, For variance; Calculate the correlation coefficient matrix of standardized data: ; Solve for the characteristic equation of the correlation coefficient matrix: ; By selecting the eigenvectors corresponding to the two largest eigenvalues, the three-dimensional data is reduced to two-dimensional data, simplifying the subsequent computational complexity.

[0033] Image feature extraction: Extracting security risk-related features from preprocessed satellite and UAV images. Texture features: Gray-level co-occurrence matrix (GLCM) is used to calculate contrast and correlation at a distance of d=1 pixel and orientations of 0° / 45° / 90° / 135°, in order to identify terrain undulations and building cluster distribution; The formula for calculating contrast is as follows: ; The formula for calculating correlation is as follows: ; The target edges are extracted by Gaussian filtering (σ=1.0), Sobel operator to calculate horizontal / vertical gradients, nonmaximum suppression, and double threshold detection. The edge point sequence is converted into complex form and discrete Fourier transform is performed. The first M sequences are taken as Fourier descriptors to describe the shape features of the aircraft and obstacles. Each center pixel is compared with its eight neighboring pixels, and the frequency of different LBP values ​​is counted to identify changes in the surface texture of the aircraft (such as damage to the drone propeller).

[0034] Temporal feature extraction: Extract statistical features (mean, variance, standard deviation) and rate of change features from the aircraft's position and velocity time-series data. ), frequency domain characteristics (converted to frequency domain signal through FFT, and the average power spectrum is calculated) are used to capture the motion trend of the aircraft.

[0035] Specifically, the intelligent collision prediction model includes an input layer, a feature fusion layer, a prediction evaluation layer, and an output layer connected in sequence; The prediction and evaluation layer includes a trajectory prediction submodule, a collision detection submodule, and a risk assessment submodule, which are used to perform prediction, detection, and evaluation calculations on the feature fusion data.

[0036] Specifically, the input layer receives preprocessed multi-source feature data.

[0037] The feature fusion layer uses an attention mechanism to weight and fuse features from different sources, dynamically assigning weights and outputting a fused feature vector.

[0038] The prediction and evaluation layer includes a trajectory prediction submodule, a collision detection submodule, and a risk assessment submodule, realizing the entire process of prediction, detection, and evaluation calculation.

[0039] The output layer outputs the collision risk probability, estimated collision time, and collision location to enable real-time decision-making.

[0040] Specifically, the trajectory prediction submodule adopts an LSTM and GRU network architecture, with 128 neurons in the LSTM hidden layer and 64 neurons in the GRU hidden layer, and the activation function is the tanh function. The collision detection submodule integrates OBB bounding box, swirling region modeling and coordinate collision analysis technologies to determine whether the drone has entered the area. The risk assessment submodule uses an SVM model. The collision risk index is input into the trained SVM model, which outputs the collision risk level. The collision risk level is then converted into a collision risk probability through probability calibration.

[0041] Specifically, the input data for the trajectory prediction submodule consists of preprocessed time-series data of the aircraft over the past 30 seconds (position, velocity, attitude, etc.), environmental data (wind speed, visibility, etc.), and flight plan data. The LSTM submodule controls the flow of information through input gates, forget gates, and output gates to capture the cruise trend of the aircraft, while the GRU submodule simplifies the calculation through update gates and reset gates to capture speed changes caused by gusts. Output the predicted trajectory coordinates (X, Y, Z) for the three future time nodes t1 (10s), t2 (30s), and t3 (60s). The X, Y, and Z coordinates of the predicted trajectory are optimized sequentially: if modifying the X coordinate increases the collision probability, the X coordinate is modified in the opposite direction; if the collision probability decreases, the Y coordinate is optimized, and so on, until the collision probability converges, ensuring the accuracy of the predicted trajectory.

[0042] Specifically, the collision detection submodule integrates OBB bounding box, spiral region modeling, and coordinate collision analysis technologies, including: For both drones and stationary obstacles, select m boundary points in the reduced 2D coordinates. Calculate the mean and covariance matrices of X and Y coordinates, solve for eigenvalues ​​and eigenvectors to determine the bounding box orientation, establish a new coordinate system, calculate the length, width and center point, and form a bounding box that fits the UAV and obstacle targets. For moving obstacle targets, the center and radius of the loop are determined based on historical trajectories to construct a circular danger zone and determine whether the drone enters it.

[0043] Specifically, the shortest distance between the drone's OBB bounding box and the stationary obstacle target's OBB bounding box is calculated. When the shortest distance is less than a first threshold, it is determined to be a potential static collision. Based on the trajectory prediction results, the positional relationship between the two objects is calculated over a period of time in the future. If there is an overlap of the OBB bounding boxes or if the objects enter a circular danger zone, it is determined to be a dynamic collision.

[0044] Specifically, an OBB directed bounding box is generated, and for UAVs and stationary obstacle targets, m=50 boundary points are randomly selected in the dimensionality-reduced two-dimensional coordinates; Calculate the average values ​​of the X and Y coordinates and construct the covariance matrix. ; Solve for the eigenvalues ​​and eigenvectors of the covariance matrix, and select the direction corresponding to the largest eigenvalue as the direction of the OBB bounding box; Establish a new coordinate system, calculate the length, width, and center point of the OBB bounding box, and form a bounding box that fits the shape of the target.

[0045] For targets with unstable motion, such as kites and balloons, the center and radius of rotation are determined based on historical trajectory data. The center of rotation is the geometric center of the historical trajectory, and the radius of rotation is the maximum distance of the historical trajectory from the center plus a safety margin. A circular rotation area is constructed to determine whether the drone has entered a dangerous area.

[0046] Specifically, the collision detection logic includes: Static collision detection: Calculate the shortest distance between the UAV's OBB bounding box and the OBB bounding box of a stationary obstacle target. When the shortest distance is less than a first threshold, it is determined to be a potential static collision. Dynamic collision judgment: Based on the trajectory prediction results, calculate the positional relationship between the two in the future. If there is an overlap of the OBB bounding boxes or the OBBs enter the circular danger zone, it is judged as a dynamic collision. Specifically, the calculation process for the collision risk index includes: ; in, As weight; The minimum safe distance threshold, This is the actual distance. For the maximum permissible relative speed, This is the actual relative speed. For the maximum permissible attitude angle deviation, This represents the actual attitude angle deviation. Visibility threshold For actual visibility, The wind speed threshold, This represents the actual wind speed.

[0047] Specifically, based on historical collision event data, the critical collision risk index for different combinations of aircraft and obstacle zone targets is determined, serving as the basis for adjusting the warning threshold.

[0048] Calculate the volatility of the risk index within the target time before the collision: ; Where T1 is the collision response interval, which is based on historical data statistics; When volatility is ≥ Z0 (mean + 1 standard deviation), it is considered a rapid increase in risk, and the early warning response time should be shortened.

[0049] The first node whose risk index volatility changes from less than Z0 to greater than Z0 within the target time period is identified as a suspected collision node. The actual indicator data (distance and speed) of this node are extracted, and the corresponding standard threshold is updated to achieve dynamic adaptation of the model.

[0050] Specifically, based on the probability of collision risk, three levels of warning thresholds are set to adapt to different handling needs: Low-risk warning: The drone is in stable flight status, there is a potential collision risk but there is sufficient time to deal with it, and the warning threshold can be dynamically adjusted according to the airspace congestion level; Medium-risk warning: The distance between the drone and the target is rapidly decreasing or the relative speed exceeds the limit, and the flight path needs to be adjusted in time; High-risk warning: A collision may occur in the short term (≤10s), requiring emergency response (return to base, hover, or avoidance).

[0051] Specifically, this also includes: constructing a three-dimensional dynamic model of the airspace, integrating satellite remote sensing terrain data, three-dimensional building models, and the real-time position of the aircraft, and intuitively displaying collision risk areas (red = high risk, yellow = medium risk, blue = low risk), predicted trajectories (dashed lines), and turning areas (semi-transparent circles) in the model.

[0052] Specifically, the management platform displays the distance change curve, speed change curve, and risk index change curve between the drone and the target, and marks the warning trigger points and suggested handling points to help staff analyze risk trends; Staff can manually adjust warning thresholds, modify flight plans, and issue avoidance instructions via an app or platform, with real-time feedback on the execution status after the instructions are issued.

[0053] This invention, for the first time, integrates trajectory prediction, OBB bounding box, and turning area to construct an airspace safety early warning system covering the entire process; it integrates collision risk index, volatility analysis, and suspicious node updates to achieve dynamic adjustment of early warning thresholds, adapting to different airspace scenarios and different aircraft types; and it constructs a dynamically updated three-dimensional airspace model to achieve a visual presentation of collision risks, improving the intuitiveness of early warnings and the efficiency of handling.

[0054] By combining LSTM+GRU trajectory prediction with OBB collision detection, the accuracy of collision prediction is improved, meeting the early warning requirements of high-speed low-altitude aircraft. It is adaptable to complex airspace scenarios such as urban areas, mountainous areas, and coastal areas, and has strong scalability. It supports continuous optimization of airspace management and improves the overall airspace safety level.

[0055] In one specific embodiment of the present invention, data on different airspace collision accidents, simulated collision data, and normal flight data from the past 5 years are collected, covering different airspace scenarios (urban, mountainous, coastal) and different aircraft types (drones, paragliders). A combination of manual and automatic annotation is used. The time, location, and risk level of collision events are manually annotated, while trajectory features and environmental parameters are automatically annotated. The dataset is divided into 70% training set, 15% validation set, and 15% test set. Stratified sampling is used to ensure a balanced proportion of data in each scenario and to avoid model bias.

[0056] The model parameters are initialized, and the model training process is as follows: Mini-batch stochastic gradient descent (SGD) is used for optimization. The initial learning rate is 0.001, which decays to 0.1 every 10 epochs. The momentum is 0.9 and the weight decay is 0.0005. The trajectory prediction loss uses mean squared error (MSE), the target detection loss uses CIoU loss, and the risk assessment loss uses cross-entropy loss. The total loss is the weighted sum of the three.

[0057] In one specific embodiment of the present invention, an airspace safety early warning system, such as... Figure 2 As shown, it includes: The data acquisition module is used to acquire multi-source heterogeneous data from the UAV; The data preprocessing module is used to preprocess the multi-source heterogeneous data; The collision monitoring module is used to build an intelligent collision prediction model based on artificial intelligence algorithms. The intelligent collision prediction model analyzes the preprocessed data in real time and outputs the collision risk probability. The early warning module is used to send warning information to staff when the probability of collision risk reaches a preset threshold.

[0058] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0059] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for airspace safety early warning, characterized in that, include: Acquire multi-source heterogeneous data from drones; The multi-source heterogeneous data is preprocessed; An intelligent collision prediction model is constructed based on artificial intelligence algorithms. The intelligent collision prediction model performs real-time analysis on preprocessed data and outputs the probability of collision risk. When the probability of a collision reaches a preset threshold, an early warning message is sent to staff.

2. The airspace safety early warning method according to claim 1, characterized in that, The multi-source heterogeneous data includes UAV body data, airspace dynamic data, and environmental auxiliary data; the UAV body data includes UAV BeiDou grid code positioning data, flight plan data, and flight status data. The airspace dynamic data includes ADS-B data and radar data; The environmental support data includes satellite remote sensing data, aerial monitoring data, and ground sensor network data.

3. The airspace safety early warning method according to claim 1, characterized in that, The intelligent collision prediction model includes an input layer, a feature fusion layer, a prediction evaluation layer, and an output layer connected in sequence. The prediction and evaluation layer includes a trajectory prediction submodule, a collision detection submodule, and a risk assessment submodule, which are used to perform prediction, detection, and evaluation calculations on the feature fusion data.

4. The airspace safety early warning method according to claim 3, characterized in that, The trajectory prediction submodule adopts an LSTM and GRU network architecture, with 128 neurons in the LSTM hidden layer and 64 neurons in the GRU hidden layer, and the activation function is the tanh function. The collision detection submodule integrates OBB bounding box, swirling region modeling and coordinate collision analysis technologies to determine whether the drone has entered the area. The risk assessment submodule uses an SVM model. The collision risk index is input into the trained SVM model, which outputs the collision risk level. The collision risk level is then converted into a collision risk probability through probability calibration.

5. The airspace safety early warning method according to claim 4, characterized in that, The collision detection submodule integrates OBB bounding box, spiral region modeling, and coordinate collision analysis technologies, including: For drones and stationary obstacle targets, select m boundary points in the dimensionality-reduced two-dimensional coordinates; Calculate the mean and covariance matrices of X and Y coordinates, solve for eigenvalues ​​and eigenvectors to determine the bounding box orientation, establish a new coordinate system, calculate the length, width and center point, and form a bounding box that fits the UAV and obstacle targets. For moving obstacle targets, the center and radius of the loop are determined based on historical trajectories to construct a circular danger zone and determine whether the drone enters it.

6. The airspace safety early warning method according to claim 5, characterized in that, Calculate the shortest distance between the drone's OBB bounding box and the OBB bounding box of a stationary obstacle target. When the shortest distance is less than a first threshold, it is determined to be a potential static collision. Based on the trajectory prediction results, the positional relationship between the two objects is calculated over a period of time in the future. If there is an overlap of the OBB bounding boxes or if the objects enter a circular danger zone, it is determined to be a dynamic collision.

7. The airspace safety early warning method according to claim 4, characterized in that, The calculation process for the collision risk index includes: ； in, As weight; The minimum safe distance threshold, This is the actual distance. For the maximum permissible relative speed, This is the actual relative speed. For the maximum permissible attitude angle deviation, This represents the actual attitude angle deviation. Visibility threshold For actual visibility, The wind speed threshold, This represents the actual wind speed.

8. An airspace safety early warning system, employing the airspace safety early warning method according to any one of claims 1-7, characterized in that, include: The data acquisition module is used to acquire multi-source heterogeneous data from the UAV; The data preprocessing module is used to preprocess the multi-source heterogeneous data; The collision monitoring module is used to build an intelligent collision prediction model based on artificial intelligence algorithms. The intelligent collision prediction model analyzes the preprocessed data in real time and outputs the collision risk probability. The early warning module is used to send warning information to staff when the probability of collision risk reaches a preset threshold.

Images