Multi-source sensing airport scene moving target cooperative monitoring and conflict early warning system

Abstract

The invention discloses a multi-source sensing airport scene moving target cooperative monitoring and conflict early warning system, and relates to the technical field of airport scene traffic management, a millimeter wave radar, a laser radar and the like form a heterogeneous network, and the whole area of an airport is synchronously covered through a TSN network; cascading filtering denoising is adopted, GPS time service and Kalman filtering are combined to realize space-time calibration, and a data complementation mechanism is established; based on a multi-modal model classification target, the tracking continuity is kept through an improved algorithm; fusing multiple factors and predicting a path by using an ST-GNN model to adapt to different scenes; constructing a four-dimensional judgment space to evaluate risks in a grading manner, and performing grading response and multi-channel information pushing according to the risks; and the system performance is continuously optimized through federated learning and digital twinborn verification. According to the invention, the method achieves the precise monitoring of the whole area of the airport scene, improves the target recognition and tracking stability, has the self-optimization capability, balances the operation safety and efficiency, and supports the high-quality development of aviation logistics and general aviation services.

Description

Technical Field

[0001] This invention relates to the field of airport surface traffic management technology, and in particular to a multi-source sensing collaborative monitoring and conflict early warning system for moving targets on the airport surface. Background Technology

[0002] With the rapid development of the modern aviation industry, airport surfaces, as the core hub of air transport, integrate multiple functions such as flight takeoffs and landings, air logistics transfers, and general aviation production services. The activity of these targets is characterized by diversity, high density, and high dynamics. Aircraft, cargo trailers, refueling trucks, baggage transfer vehicles, and personnel operate simultaneously on runways, taxiways, aprons, and cargo areas. The intersecting trajectories of these targets and the tight operational rhythm place stringent demands on the safety and efficiency of surface traffic. Traditional airport surface surveillance systems often rely on single-sensor devices. Millimeter-wave radar is susceptible to clutter interference, high-definition cameras suffer from reduced recognition capabilities in low-visibility environments such as rain, fog, and nighttime, ADS-B systems cannot cover ground vehicles and personnel without onboard equipment, and UWB positioning is only suitable for indoor or small areas. This single-sensor approach struggles to achieve comprehensive, all-target coverage, especially in air logistics cargo areas where cargo transfer vehicles frequently move and loading / unloading operations are intensive. Existing equipment struggles to accurately capture the dynamic information of these low-to-medium speed moving targets, easily creating surveillance blind spots.

[0003] The collaborative processing and spatiotemporal consistency of multi-source sensor data are core technological bottlenecks in surface surveillance. In existing systems, differences in acquisition frequencies and coordinate systems among different types of sensors, coupled with a lack of efficient synchronization and calibration mechanisms, lead to spatiotemporal misalignment of data, resulting in significant deviations in the fused target position and status information. Regarding target recognition and tracking, traditional algorithms are mostly based on single-modal data, making it difficult to effectively integrate the spatial features of radar point clouds, the visual features of images, and the identity information of ADS-B. This results in low accuracy in classifying similar-looking targets, and in scenarios involving high-speed target movement, changes in direction, or occlusion, tracking IDs are easily lost, hindering continuous monitoring of the target throughout its entire lifecycle. For small general aviation aircraft, their flexible takeoff and landing and low-speed taxiing characteristics differ significantly from those of civil aircraft. Existing tracking algorithms do not fully consider the motion characteristics of these targets, leading to insufficient tracking stability. Furthermore, in air logistics scenarios, changes in the load status of cargo transport vehicles affect their motion characteristics, and existing systems lack the ability to perceive and adapt to these dynamic features, further reducing tracking accuracy.

[0004] The scientific rigor of conflict early warning and risk assessment directly impacts airport surface operational safety. Existing systems often use fixed safety distances, failing to adequately consider dynamic factors such as target speed, weather conditions, and road surface conditions. In high-speed target intersections or low-visibility environments, fixed safety distances are either overly conservative, leading to low operational efficiency, or overly lenient, causing safety hazards. Conflict time calculations are based solely on current position and speed, neglecting the uncertainty of target movement trends and failing to adequately predict sudden movements such as rapid acceleration and sharp turns, easily resulting in false or missed warnings. Furthermore, existing systems lack continuous optimization capabilities. As airport operating modes adjust and target types increase, the performance of recognition models and tracking algorithms gradually degrades, and the aging and failure of sensor equipment are not effectively predicted, leading to a decline in the long-term reliability and stability of the system. This makes it difficult to adapt to the high-density, high-dynamic operational demands of modern airports, failing to provide accurate and timely decision support for airport dispatchers, and hindering the improvement of air logistics efficiency and general aviation production service quality. Summary of the Invention

[0005] The present invention proposes a multi-source sensing collaborative monitoring and conflict early warning system for moving targets on the airport surface to solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a multi-source sensing airport surface moving target collaborative monitoring and conflict early warning system, comprising the following modules: The multi-source sensor data acquisition module consists of a heterogeneous sensor network composed of millimeter-wave radar, lidar, high-definition camera, ADS-B receiver, UWB positioning module and meteorological sensor; all sensors are synchronized through a time-sensitive network. The data preprocessing and spatiotemporal calibration module uses a cascaded filtering architecture to process the raw data. Radar data is denoised and clutter is removed by detection, and camera images are denoised and illumination compensation is used to correct abnormal areas. Multi-source data are synchronized in time through GPS timing and local crystal oscillator calibration. Spatial calibration is completed based on the airport's three-dimensional coordinate system, an external parameter matrix is ​​established, and coordinate deviations are corrected in real time. The active target collaborative identification and tracking module classifies targets based on a multimodal Transformer model, taking input data and outputting target categories; it adopts a multi-target tracking framework combining improved particle filtering and Hungarian algorithm, assigning unique IDs to targets; a high-dynamic tracking mode is enabled during aircraft takeoff and landing to handle target motion at speeds ≥300km / h; The dynamic path prediction module integrates historical target trajectories, airport control rules, and flight schedules to generate predicted paths. It adopts a spatiotemporal graph neural network model, taking parameters from the past 30 seconds as input and outputting path predictions for the next 10 to 60 seconds. It also introduces a scene adaptive factor to dynamically adjust the prediction step size and frequency. The conflict risk assessment module constructs a four-dimensional conflict determination space and sets the minimum safe distance under different scenarios; it calculates the conflict time based on relative motion vectors, classifies five risk levels, and obtains a weighted risk value by combining distance and time risks. The early warning information generation and push module generates graded response strategies based on risk levels, and gradually upgrades the processing from system recording to the linkage braking system. The system's self-optimization module constructs a closed-loop iterative mechanism, generates a performance report every 24 hours, updates the model using a federated learning framework, establishes a sensor health assessment model, and triggers maintenance prompts; it also introduces digital twin verification and correction parameters, and conducts monthly system-wide calibration to ensure stable operation.

[0007] Furthermore, it also includes a sub-module for refined conflict risk assessment, which calculates the conflict probability distribution of the target within the predicted time period through integration, using the formula: Where P is the conflict probability. To predict the start time, To predict the termination time, Let be the function representing the real-time distance between targets as a function of time. For minimum safe distance, This represents the standard deviation of the distance fluctuation.

[0008] Furthermore, it also includes a dynamic path prediction optimization submodule, which introduces a path prediction model that corrects the rate of change of target motion acceleration. The formula for the rate of change of acceleration is: ,in For the rate of change of acceleration, Let be the target acceleration at time t. for The target acceleration at all times, The sampling time interval is set to 0.1s.

[0009] Furthermore, the multi-source sensor data acquisition module adopts a layered deployment and redundancy design; eight millimeter-wave radars are symmetrically deployed 50 meters on both sides of the runway centerline at the runway level to form 360° coverage without blind spots, with the radar beam elevation angle set at 3° to simultaneously detect low-altitude aircraft and ground targets; one sensor node is set every 200 meters at the taxiway level; the apron level is grouped according to parking positions, with each group of 4-6 parking positions configured with one UWB positioning base station and three ADS-B receivers, with the base stations deployed at the corners of the parking positions to form triangulation positioning; at the cargo logistics level, UWB tags and fixed readers are deployed at the entrances and exits of the cargo station and along the conveyor belt; critical position sensors adopt a 1+1 hot backup.

[0010] Furthermore, the data preprocessing and spatiotemporal calibration module introduces deep learning-assisted calibration, constructs a CNN-based sensor deviation prediction model, inputs historical sensor deviation data, meteorological parameters, and equipment runtime, and outputs real-time deviation compensation values; time synchronization adopts a master-slave clock architecture, with slave devices synchronizing via the PTPv2 protocol; spatial calibration establishes a dynamic error map, dividing the airport into 10×10 meter grids, with each grid storing the average calibration error of that area.

[0011] Furthermore, the collaborative target identification and tracking module employs federated distillation technology to optimize the identification model. A general feature extractor is generated by the central server, and each sensor node trains a category classifier locally. The classifier knowledge is then distilled into the central model. The tracking algorithm incorporates target intent recognition, predicting the target's next action by analyzing the target's historical trajectory and airport control instructions. For small targets such as drones, micro-Doppler feature analysis is enabled to extract the Doppler frequency shift of rotating parts from radar echoes. In air logistics target tracking, cargo electronic tag information is linked.

[0012] Furthermore, the dynamic path prediction module, combined with the airport's digital twin system, integrates real-time sensor data with virtual scenes to simulate target movement trends in virtual space, and the prediction results are presented through 3D visualization. The path planning algorithm incorporates ecological driving factors, adjusting the suggested speed for ground vehicles based on conflict risks; for flight taxiing paths, it optimizes acceleration / deceleration timing by combining engine thrust characteristics and taxiway gradient. The prediction model has self-learning capabilities; when the deviation between the actual trajectory and the predicted trajectory exceeds a threshold three times consecutively, it automatically triggers fine-tuning of model parameters, updating weights using online gradient descent to adapt to changes in airport operation modes.

[0013] Furthermore, the conflict risk assessment module establishes a dynamic safety distance model, where the safety distance adaptively adjusts based on target speed, weather conditions, and road surface conditions, as shown in the formula: ,in For dynamic safety distance, Here, v is the velocity coefficient, and v is the target velocity. Here, w is the meteorological coefficient, and w is the wind speed. denoted as pavement coefficient, and r as pavement slope.

[0014] Furthermore, the early warning information generation and push module has augmented reality visualization capabilities. When dispatchers wear AR glasses, early warning information is overlaid on the actual scene through virtual icons. For ground support vehicles, the vehicle terminal integrates voice interaction, supporting the query of early warning details or confirmation of avoidance operations through voice commands. An early warning effect feedback mechanism is established to record the target response behavior after each early warning, calculate the early warning effectiveness rate, and automatically adjust the early warning method or content when the effectiveness rate is less than 80%.

[0015] Furthermore, the system's self-optimization module employs blockchain technology to ensure data credibility. Each sensor node acts as a blockchain node, encrypting the collected data before storing it on the chain, and using hash value verification to confirm data integrity. Equipment health monitoring incorporates remaining lifespan prediction, analyzing historical sensor fault data, operating temperature, and vibration intensity to predict remaining lifespan using an LSTM neural network. System performance evaluation introduces a comprehensive efficiency index, calculated using the following formula: Where E is the comprehensive performance index, A is the identification accuracy, T is the tracking continuity, W is the early warning accuracy, and F is the equipment fault-free operation time. Deep optimization is initiated when E < 80.

[0016] Compared with existing technologies, the beneficial effects of this invention are: At the sensor acquisition level, a layered deployment network is constructed using a variety of heterogeneous devices, including millimeter-wave radar, lidar, and high-definition cameras. Combined with redundancy design and data complementarity mechanisms, this achieves full coverage of airport runways, taxiways, aprons, and cargo areas, adapting to complex weather conditions such as rain, fog, and nighttime, eliminating blind spots in traditional systems. Various sensors achieve high-precision synchronization through a time-sensitive network, ensuring spatiotemporal consistency of multi-source data and laying a solid foundation for subsequent data processing. In particular, it can accurately capture dynamic information of different types of targets, such as air cargo vehicles and general aviation small aircraft, meeting diverse surveillance needs.

[0017] In target recognition and tracking, the multi-modal Transformer model, which integrates multi-source features, significantly improves the accuracy and speed of target classification. It can effectively distinguish between different types of ground vehicles, aircraft, and personnel, and maintain the unique ID of each target continuously. The improved multi-target tracking framework combines a target motion feature library with intent recognition technology, adapting to different motion states such as high speed, low speed, and change of direction. Even in scenarios with target occlusion or variable load, it can maintain stable tracking, achieving full-chain traceability of "vehicle-cargo-flight" and providing accurate data support for aviation logistics management. At the same time, the tracking strategy is optimized for the motion characteristics of general aviation small aircraft, improving the tracking stability of such flexible targets.

[0018] In terms of conflict early warning and system optimization, a dynamic safety distance model and a four-dimensional conflict determination space are constructed. Combining target priority and motion uncertainty, a refined assessment of conflict risk is achieved. The tiered early warning response strategy aligns with actual airport operations, ensuring operational safety while avoiding excessive intervention that could impact efficiency. Early warning information is pushed through multiple channels, combined with augmented reality visualization and voice interaction to improve human-machine collaboration efficiency, ensuring that dispatchers and target operators receive critical information promptly and take effective avoidance measures. The system self-optimization module, through federated learning, digital twin verification, and equipment health monitoring, enables continuous iteration of model algorithms and early prediction of equipment failures, ensuring the long-term reliability and stability of the system and adapting to dynamic changes in airport operation modes. Overall, this invention significantly improves the comprehensiveness and accuracy of airport surface surveillance, optimizes the scientific rigor and timeliness of conflict early warning, and enhances airport surface operational efficiency and safety, providing strong technical support for efficient air logistics transfer and high-quality development of general aviation production services. Attached Figure Description

[0019] Figure 1 This is a schematic block diagram of the multi-source sensing airport surface moving target collaborative monitoring and conflict early warning system proposed in this invention; Figure 2 A comparison chart of target recognition accuracy under different operating scenarios; Figure 3 A comparison chart of early warning accuracy and false alarm rate for different risk levels; Figure 4 This is a graph showing the trend of the system's overall efficiency index over time. Detailed Implementation

[0020] 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.

[0021] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0022] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0023] Reference Figures 1 to 4 A multi-source sensing collaborative monitoring and conflict early warning system for moving targets on the airport surface, comprising the following modules: The multi-source sensor data acquisition module comprises a heterogeneous sensor network consisting of millimeter-wave radar, lidar, high-definition cameras, an ADS-B receiver, a UWB positioning module, and meteorological sensors. The millimeter-wave radar is deployed at both ends of the runway and taxiway intersections, with a detection range of 50-800 meters, an angular resolution of 0.5°, and a refresh rate of 20Hz. The lidar is deployed at the edge of the apron, with a point cloud density of 200 points / m², a ranging accuracy of ±2cm, and is suitable for rainy, foggy, and low-visibility environments. The high-definition camera uses a dual-lens combination of 360° panoramic and close-up lenses, with the panoramic lens covering a radius of 100 meters. The camera lens supports 20x optical zoom, 50fps frame rate, and infrared night vision; the ADS-B receiver consists of a ground station and a vehicle-mounted mobile station. The ground station has a reception radius of 10 kilometers, and the mobile station is deployed with the support vehicle, with an update frequency of 2Hz; the UWB positioning module forms a cellular network in the terminal building and cargo area, with a positioning error of ±10cm, and supports simultaneous tracking of 1000+ targets; the meteorological sensor collects visibility, wind speed, and precipitation intensity data in real time, with a sampling frequency of 1Hz; all sensors are synchronized through a Time-Sensitive Network (TSN), with a synchronization error of ≤5ms; The data preprocessing and spatiotemporal calibration module uses a cascaded filtering architecture to process raw data. Radar data undergoes wavelet threshold denoising (soft threshold function, threshold λ=0.8σ) and constant false alarm rate (CFAR, window size 16×16) to remove clutter interference. Camera images are denoised using an improved BM3D algorithm, combined with an illumination compensation model to correct backlight and shadow areas. Multi-source data time synchronization uses a combination of GPS timing and local crystal oscillator calibration, with GPS second pulse synchronization accuracy ±1μs and a local calibration cycle of 100ms. Spatial calibration is based on the airport's three-dimensional coordinate system (X-axis along the runway direction, Y-axis perpendicular to the runway, Z-axis vertically upward). A sensor extrinsic parameter matrix is ​​established by calibrating targets, and Kalman filtering is used to correct coordinate deviations in real time, with static calibration error ≤0.1m and dynamic tracking calibration error ≤0.3m. For sensor failure, a data complementarity mechanism is constructed. When the lidar fails, the weight of the millimeter-wave radar is automatically increased to 0.7, and the weight of the camera is increased to 0.3. The collaborative target identification and tracking module classifies targets based on a multimodal Transformer model. It takes radar point cloud data, image features, and ADS-B identification information as input and outputs target categories (aircraft, cargo vehicles, refueling trucks, baggage carts, personnel). The identification confidence level is ≥98%, and the identification speed is ≤50ms. The tracking algorithm employs a multi-target tracking framework combining improved particle filtering and the Hungarian algorithm, assigning a unique ID to each target with an ID retention rate ≥99%. A high-dynamic tracking mode is activated during aircraft takeoff and landing, with an extended Kalman filter used in the state estimator to handle target motion at speeds ≥300km / h. For cargo trailers in air logistics, the module additionally identifies load status (empty / heavy / overloaded) through lidar point cloud volume calculation, with an error ≤5%. A target motion feature library is established, storing parameters such as maximum speed, acceleration, and turning radius for different target types, used for abnormal motion detection. The dynamic path prediction module integrates historical target trajectories, airport control rules, and flight schedules to generate predicted paths. It adopts a spatiotemporal graph neural network (ST-GNN) model, taking into input trajectory points from the past 30 seconds (sampling interval 0.5s), current speed, acceleration, and environmental constraints (restricted areas, construction zones), and outputs path predictions for the next 10-60 seconds. The short-term (10s) prediction deviation is ≤0.5 meters, and the long-term (60s) prediction deviation is ≤3 meters. For small aircraft in general aviation, vertical take-off and landing constraints are added to the prediction model. For vehicles in the cargo area, the dwell time and driving path are corrected by combining the loading and unloading plans of the logistics scheduling system. A scenario adaptive factor is introduced, shortening the prediction step to 10 seconds during busy periods (peak flight take-off and landing times) and extending it to 60 seconds during idle periods. The prediction frequency is dynamically adjusted according to the target speed, with a prediction frequency of 10Hz when the speed is ≥50km / h and 5Hz when the speed is low. The conflict risk assessment module constructs a four-dimensional conflict determination space (x, y, z, t) and calculates the spatiotemporal intersection between targets. Minimum safe distances between aircraft are defined as follows: runway ≥ 100 meters, taxiway ≥ 50 meters, apron ≥ 20 meters; aircraft and vehicle ≥ 30 meters; vehicle and personnel ≥ 10 meters. Conflict time of conflict (TTC) is calculated based on relative motion vectors, using the formula TTC = d / v_rel, where d is the current distance and v_rel is the relative velocity vector magnitude. Risk levels are divided into five levels: Level 1 (no risk) TTC > 60s, Level 2 (concern) 30s < TTC ≤ 60s, Level 3 (alert) 15s < TTC ≤ 30s, Level 4 (warning) 5s < TTC ≤ 15s, and Level 5 (emergency) TTC ≤ 5s. Target priority coefficients are introduced: 1.0 for flights, 0.9 for emergency vehicles, 0.7 for cargo vehicles, 0.5 for ordinary vehicles, and 0.8 for personnel. The comprehensive risk value is a weighted sum of distance and time risks. The early warning information generation and push module generates a tiered response strategy based on the risk level. Level 1 is only recorded by the system; Level 2 is pushed to the dispatch center's large screen; Level 3 sends text prompts to relevant personnel through the dispatch terminal; Level 4 triggers an audible and visual alarm on the vehicle terminal and displays an avoidance path; Level 5 links the target braking system (the vehicle automatically decelerates to 2 km / h, and the aircraft engine limits thrust). The early warning information includes the conflict target ID, current location (latitude and longitude ±0.0001°), estimated conflict time, and suggested avoidance direction (e.g., "left turn, deviation 30 meters"). The push channels support 5G slice communication (latency ≤100ms), airport dedicated communication network, and Beidou short message (areas without public network coverage), covering the entire scenario of runways, cargo areas, and terminals. For international flights, it supports bilingual (Chinese and English) early warnings, with voice prompts using real-time TTS synthesis and a response latency ≤0.3s. The system's self-optimization module constructs a closed-loop iterative mechanism, collecting system operation indicators (identification accuracy, tracking continuity, early warning accuracy, and false alarm rate) in real time and generating a performance report every 24 hours. A federated learning framework is used for model updates, with participating nodes including the airport dispatch center, airlines, and ground support units. Local model training uses the Adam optimizer (initial learning rate 0.001), and federated aggregation uses a weighted average (weights are positively correlated with the amount of data per node). A sensor health assessment model is established, calculating a health index (0-100) by analyzing signal-to-noise ratio, data integrity, and response latency. An index < 60 automatically triggers a maintenance prompt. Digital twin verification is introduced, comparing actual operating data with virtual scenario simulation results to correct prediction model parameters. A full system calibration is performed monthly to maintain long-term stable and efficient system operation.

[0024] This invention also includes a conflict risk refinement assessment submodule, which calculates the conflict probability distribution of the target within the predicted time period through integration, using the following formula: Where P is the conflict probability. To predict the start time, To predict the termination time, Let be the function representing the real-time distance between targets as a function of time. For minimum safe distance, The distance fluctuation standard deviation is used to accurately reflect the temporal distribution characteristics of conflict risk during the target's movement, reducing the probability of misjudgment caused by assessment at a single time point. It is suitable for conflict assessment of high-speed moving targets or targets with complex paths.

[0025] This invention also includes a dynamic path prediction and optimization submodule, which introduces a path prediction model corrected for the rate of change of target motion acceleration. The formula for the rate of change of acceleration is: ,in For the rate of change of acceleration, Let be the target acceleration at time t. for The target acceleration at all times, The sampling time interval is set to 0.1s. This parameter is used to predict sudden changes in the target's motion state (rapid acceleration, rapid deceleration, sharp turn), and adjust the path prediction results in advance to make the predicted path more consistent with the actual motion law of the target and reduce the path prediction deviation caused by sudden changes in motion state.

[0026] In this invention, the multi-source sensor data acquisition module adopts a layered deployment and redundancy design; eight millimeter-wave radars are symmetrically deployed 50 meters on both sides of the runway centerline at the runway level, forming 360° coverage without blind spots, with the radar beam elevation angle set at 3°, simultaneously detecting low-altitude aircraft and ground targets; at the taxiway level, one sensor node is set every 200 meters, each node containing one lidar and two high-definition cameras, with the cameras installed at a height of 6 meters and a depression angle of 30°, covering two adjacent intersections; at the apron level, the aircraft stands are grouped according to parking positions, with each group containing 4-6 aircraft stands. The system is equipped with one UWB positioning base station and three ADS-B receivers. The base station is deployed in the corner of the aircraft position to form a triangulation positioning system. In the freight logistics layer, UWB tags and fixed readers are deployed at the entrance and exit of the freight station and along the conveyor belt. The reader identification distance is ≥5 meters, enabling full tracking of cargo transfer vehicles. All sensors are corrosion-resistant and electromagnetic interference resistant, with an operating temperature of -40℃ to 70℃ and a humidity of 0% to 95% without condensation. The critical position sensors (at both ends of the runway) adopt a 1+1 hot backup, with a fault switching time of ≤100ms.

[0027] In this invention, the data preprocessing and spatiotemporal calibration module introduces deep learning-assisted calibration, constructing a CNN-based sensor deviation prediction model. Inputting historical sensor deviation data, meteorological parameters, and equipment runtime, it outputs real-time deviation compensation values, improving compensation accuracy by 40% compared to traditional methods. Time synchronization employs a master-slave clock architecture, using the airport's UTC master clock as a reference. Slave devices synchronize via the PTPv2 protocol, with a synchronization message transmission interval of 10ms and jitter ≤1ms. Spatial calibration establishes a dynamic error map, dividing the airport into 10×10 meter grids. Each grid stores the average calibration error for that area, used for accuracy compensation in local path planning. In rainy or snowy weather, LiDAR data undergoes additional raindrop point cloud removal processing. By analyzing point cloud reflection intensity and spatial distribution characteristics, rain and fog interference points are removed, achieving an effective data retention rate ≥85%. Under low-light conditions, camera images automatically enable ISP wide dynamic range processing, extending the dynamic range to 120dB to ensure effective target recognition at night.

[0028] In this invention, the collaborative target identification and tracking module employs federated distillation technology to optimize the identification model. A general feature extractor is generated by the central server, and each sensor node trains a category classifier locally. The classifier knowledge is then distilled into the central model, reducing model parameters by 60% and increasing inference speed by 50% compared to traditional methods. The tracking algorithm incorporates target intent recognition, predicting the target's next action (takeoff, parking, turning) by analyzing the target's historical trajectory and airport control instructions, achieving an intent recognition accuracy of ≥90%. For small targets such as drones, micro-Doppler feature analysis is enabled to extract the Doppler frequency shift of rotating parts from radar echoes, distinguishing them from birds with an accuracy of ≥95%. In air logistics target tracking, cargo electronic tag information is linked to achieve full-chain traceability of "vehicle-cargo-flight," with a tracking information update delay of ≤1s. It supports querying cargo details and destination by target ID.

[0029] In this invention, the dynamic path prediction module, combined with the airport digital twin system, integrates real-time sensor data with virtual scenes to simulate the target's movement trend in virtual space, and the prediction results are presented through three-dimensional visualization. The path planning algorithm introduces an eco-driving factor, adjusting the suggested speed for ground vehicles based on conflict risks to reduce energy consumption under safe conditions, achieving an energy saving rate of ≥5%. For flight taxiing paths, it optimizes acceleration / deceleration timing by combining engine thrust characteristics and taxiway gradient, shortening taxiing time by 5-10%. The prediction model has self-learning capabilities; when the deviation between the actual trajectory and the predicted trajectory exceeds a threshold for three consecutive times, it automatically triggers fine-tuning of model parameters, updating weights using an online gradient descent method to adapt to changes in airport operation modes.

[0030] In this invention, the conflict risk assessment module establishes a dynamic safety distance model. The safety distance is adaptively adjusted according to the target speed, weather conditions, and road surface conditions, as shown in the formula: ,in For dynamic safety distance, Here, v is the velocity coefficient (value 0.5-1.2), and v is the target velocity. Here, w represents the meteorological coefficient (0.8 for sunny days, 1.2 for rainy days, and 1.5 for foggy days), and w represents the wind speed. The pavement coefficient is 0.9 for dry conditions and 1.3 for wet conditions, and r is the pavement slope. This model makes the calculation of safe distance more consistent with the actual operating environment and improves the accuracy of early warning by 15-20% under complex conditions.

[0031] In this invention, the early warning information generation and push module has augmented reality (AR) visualization capabilities. When dispatchers wear AR glasses, the early warning information is superimposed on the actual scene as a virtual identifier, displaying the location, movement trajectory, and risk level of the conflict target. The spatial alignment error between the virtual identifier and the real target is ≤0.5 meters. For ground support vehicles, the vehicle terminal integrates voice interaction, supporting the query of early warning details or confirmation of avoidance operations via voice commands, with a voice recognition accuracy rate ≥95%. An early warning effect feedback mechanism is established, recording the target response behavior (deceleration, steering, stopping) after each early warning, calculating the early warning effectiveness rate (the proportion of response behavior that conforms to the recommended operation). When the effectiveness rate is <80%, the early warning method or content is automatically adjusted to improve human-machine collaboration efficiency.

[0032] In this invention, the system self-optimization module uses blockchain technology to ensure data credibility. Each sensor node acts as a blockchain node, and the collected data is encrypted and stored on the chain. Hash value verification verifies data integrity. The on-chain data includes sensor ID, acquisition time, raw data summary, and preprocessing results, supporting traceability and auditing. Equipment health monitoring incorporates remaining service life prediction. By analyzing historical fault data, operating temperature, and vibration intensity of the sensors, an LSTM neural network is used to predict the remaining service life (RUL), with a prediction error ≤10%. System performance evaluation incorporates a comprehensive efficiency index, calculated using the following formula: Where E is the comprehensive performance index (0-100), A is the recognition accuracy, T is the tracking continuity, W is the early warning accuracy, and F is the equipment fault-free operation time. When E < 80, deep optimization is initiated, including model retraining, sensor calibration, and parameter reset, to maintain the system's long-term optimal operating state.

[0033] The following two examples further illustrate the specific implementation of this system: Example 1: Implementation of Multi-Target Collaborative Surveillance and Conflict Early Warning at Large Hub Airports This embodiment targets a large hub airport with an annual passenger throughput exceeding 60 million and a cargo and mail throughput exceeding 1.5 million tons. This airport includes 4 runways, 8 main taxiways, 50 aircraft stands, and a 100,000 square meter cargo and logistics area. During daily operations, it simultaneously hosts various types of targets, including civil aircraft, general aviation aircraft, cargo trailers, refueling trucks, baggage transfer vehicles, and various personnel. During peak hours, the target density reaches 30 targets per square kilometer. The implementation must adapt to complex weather conditions and high-density operational scenarios. The specific implementation is as follows: 1. System module deployment and parameter configuration Multi-source sensor data acquisition module deployment: On the runway level, 8 millimeter-wave radars are symmetrically deployed 50 meters on either side of the centerline of each runway, totaling 32 units. The radar beam elevation angle is 3°, the detection range covers 50-800 meters, the angular resolution is 0.5°, and the refresh rate is 20Hz. On the taxiway level, one sensor node is set every 200 meters, totaling 40 nodes. Each node integrates one lidar and two high-definition cameras. The lidar point cloud density is 200 points / ㎡, the ranging accuracy is ±2cm, and the high-definition cameras are installed at a height of 6 meters, with a 30° elevation angle. The 360° panoramic lens covers a radius of 100 meters, the close-up lens has 20x optical zoom, a frame rate of 50fps, and infrared night vision capabilities. On the apron level, the aircraft stands are grouped into 10 groups of 4-6 units each, with each group equipped with one UWB positioning base station and three A... The DS-B receiver and UWB base station are deployed in the corners of the equipment positions to form a triangulation positioning system with a positioning error of ±10cm. The ADS-B ground station has a receiving radius of 10 kilometers. The mobile station is deployed with 20 support vehicles, with an update frequency of 2Hz. In the freight logistics area, 50 UWB tags and 30 fixed readers are deployed at the entrances and exits and along the conveyor belts. The reader identification distance is ≥5 meters. Ten meteorological sensors are deployed simultaneously, with a sampling frequency of 1Hz, to collect data on visibility, wind speed, and precipitation intensity. All sensors are synchronized using a TSN network with a synchronization error of ≤5ms. Sensors in critical locations use 1+1 hot backup, with a fault switching time of ≤100ms. The sensor operating temperature range is -40℃ to 70℃, and the humidity is 0-95% with no condensation. It has anti-corrosion and anti-electromagnetic interference capabilities.

[0034] Data preprocessing and spatiotemporal calibration module configuration: Radar data uses wavelet thresholding for noise reduction, with a soft threshold function threshold λ=0.8σ, where σ is the standard deviation of the original data. Clutter is subsequently removed using CFAR constant false alarm rate detection with a 16×16 window. Camera images undergo noise reduction using an improved BM3D algorithm, combined with a lighting compensation model to correct backlight and shadow areas. Under low light conditions, ISP wide dynamic range processing is automatically enabled, extending the dynamic range to 120dB. Time synchronization adopts a master-slave clock architecture, using the airport's UTC master clock as a reference. Slave devices synchronize via the PTPv2 protocol, with a synchronization message transmission interval of 10ms, jitter ≤1ms, and GPS second pulse synchronization. The step accuracy is ±1μs, and the local calibration cycle is 100ms. Spatial calibration is based on the airport's three-dimensional coordinate system, with the X-axis along the runway direction, the Y-axis perpendicular to the runway, and the Z-axis vertically upward. The sensor extrinsic parameter matrix is ​​established by calibrating targets, and Kalman filtering is used to correct coordinate deviations in real time. The static calibration error is ≤0.1m, and the dynamic tracking calibration error is ≤0.3m. A dynamic error map with a 10×10m grid is established to store the average calibration error of each area. In rainy or snowy weather, the lidar data is processed by a raindrop point cloud removal algorithm, and the effective data retention rate is ≥85%. When the lidar fails, the weight of the millimeter-wave radar is automatically increased to 0.7, and the weight of the camera is increased to 0.3.

[0035] The target collaborative identification and tracking module is configured as follows: A recognition network is built based on a multimodal Transformer model. Inputs include radar point clouds, image features, and ADS-B identity information. Outputs five target categories: aircraft, cargo vehicles, refueling trucks, baggage carts, and personnel. The model is trained on 500,000 labeled data points, achieving a recognition confidence level ≥98% and a recognition speed ≤50ms. Federated distillation technology is used to optimize the model. A general feature extractor is generated by the central server, and each sensor node trains its classifier locally before distilling it into the central model. This reduces model parameters by 60% and improves inference speed by 50% compared to traditional methods. The tracking algorithm uses a framework combining improved particle filtering and the Hungarian algorithm, assigning a unique ID to each target with an ID retention rate ≥99%. A high-dynamic tracking mode is activated during aircraft takeoff and landing. The state estimator uses an extended Kalman filter to handle target motion at speeds ≥300km / h. For cargo trailers, load status is identified through lidar point cloud volume calculation, categorized as empty, heavily loaded, and overloaded, with an error ≤5%. A target motion feature database is established, storing data on the maximum speed of civil aircraft up to 300km / h and acceleration up to 2m / s². 2 The maximum speed of freight vehicles is 50 km / h and the acceleration is 1 m / s². 2 Parameters such as these are used for abnormal motion detection; for small targets such as drones, micro-Doppler feature analysis is enabled, with a differentiation accuracy of ≥95%; in air logistics target tracking, the electronic tag information of the cargo is linked, and the tracking information update delay is ≤1s.

[0036] The dynamic path prediction and conflict risk assessment module is configured as follows: It employs a spatiotemporal graph neural network (ST-GNN) model, taking as input trajectory points from the past 30 seconds at a sampling interval of 0.5 seconds, current speed, acceleration, and environmental constraints such as restricted areas and construction zones. The output is a path prediction for the next 10-60 seconds, with a short-term prediction deviation of ≤0.5 meters for 10 seconds and a long-term prediction deviation of ≤3 meters for 60 seconds. The prediction step size is 10 seconds during busy periods and 60 seconds during idle periods. The prediction frequency is 10Hz when the target speed is ≥50km / h and 5Hz at low speeds. For general aviation small aircraft, the prediction model incorporates vertical takeoff and landing constraints and adjusts the dwell time and path of freight vehicles based on the loading and unloading plans of the logistics scheduling system. The conflict risk assessment constructs a four-dimensional conflict determination space, including x, y, and y coordinates. The four dimensions are z and t. The minimum safe distance between aircraft is ≥100 meters on the runway, ≥50 meters on the taxiway, and ≥20 meters on the apron. The distance between aircraft and vehicles is ≥30 meters, and the distance between vehicles and personnel is ≥10 meters. The conflict time total (TTC) is calculated based on the relative motion vector: TTC = d / v_rel. Risk levels are divided into five levels: Level 1 TTC > 60s, Level 2 30s < TTC ≤ 60s, Level 3 15s < TTC ≤ 30s, Level 4 5s < TTC ≤ 15s, and Level 5 TTC ≤ 5s. A target priority coefficient is introduced: 1.0 for flights, 0.9 for emergency vehicles, 0.7 for cargo vehicles, 0.5 for general vehicles, and 0.8 for personnel. The comprehensive risk value is the weighted sum of distance and time risks. The conflict probability distribution is calculated through integration, using the formula: ,Pick =0s, =30s, d(t) is the real-time distance function between targets. =50 meters is the minimum safe distance between aircraft on the taxiway. =2 meters, calculated P=0.03, judged as low conflict probability; dynamic safety distance model formula is: The speed of a freight vehicle, v=30km / h, is equivalent to 8.33m / s. =0.8, rainy day =1.2, wind speed =5m / s, road surface is slippery =1.3, slope r=0.02, calculated as follows =0.8×8.33+1.2×5+1.3×0.02≈6.66+6+0.026≈12.686 meters, matching the safety distance requirement between vehicles and pedestrians; acceleration rate of change correction path prediction is introduced, the formula is... Δt=0.1s, acceleration at time t =0.5m / s 2 , time =2m / s 2 Calculated =(2-0.5) / 0.1=15m / s3 If it is determined to be a rapid acceleration, the prediction step size is shortened to 10 seconds.

[0037] The configuration of the early warning information generation, push, and system self-optimization module is as follows: Level 1 risks are only recorded by the system; Level 2 risks are pushed to the dispatch center's large screen; Level 3 risks are sent as text prompts through the dispatch terminal; Level 4 risks trigger audible and visual alarms on vehicle terminals and display avoidance paths; Level 5 risks are linked to the target braking system. Early warning information includes the conflict target ID, latitude and longitude ±0.0001°, estimated conflict time, and suggested avoidance direction. It utilizes 5G slicing communication with a latency ≤100ms, airport-dedicated communication network, and BeiDou short message push. International flights support bilingual (Chinese and English) early warnings, with a voice TTS synthesis response latency ≤0.3s. When dispatchers wear AR glasses, the early warning information is superimposed on the actual scene as a virtual identifier, with a spatial alignment error ≤0.5 meters. Ground vehicles... The terminal integrates voice interaction with a voice recognition accuracy of ≥95%. The system's self-optimization module generates a performance report every 24 hours. It employs a federated learning framework, with participating nodes including the airport dispatch center, three airlines, and two ground support units. Local model training uses the Adam optimizer with an initial learning rate of 0.001. The weighted average weight of the federated aggregation is positively correlated with the amount of data per node. Sensor health indices are calculated based on signal-to-noise ratio, data integrity, and response latency, ranging from 0 to 100. A maintenance prompt is triggered when the index is <60. Digital twin verification is introduced, and a full system calibration is performed monthly. Blockchain technology is used to store sensor data, supporting traceability and auditing. The equipment's remaining lifespan prediction error is ≤10%, and the overall efficiency index is [not specified]. A represents the recognition accuracy, T represents the tracking continuity, W represents the early warning accuracy, and F represents the equipment's fault-free operating time.

[0038] 2. Effect Verification and Table Analysis Table 1 Key Operational Indicators of Large Hub Airport System (Example 1)

[0039] Table 1 shows data based on 30 days of continuous system operation monitoring, comprehensively reflecting the operational effectiveness of this invention in complex scenarios at large hub airports. The target recognition accuracy consistently remained above 97.8%, maintaining a high level even in low-visibility foggy weather and at night. This is attributed to the complementary fusion of multi-source sensor data and environmental adaptation technologies such as illumination compensation and raindrop cloud removal, effectively overcoming the performance degradation problem of single sensors. The tracking ID retention rate exceeded 98.5%, demonstrating that the improved multi-target tracking framework and target intent recognition technology can effectively handle complex situations such as target occlusion and high-speed movement, achieving continuous tracking throughout the entire process, especially suitable for the frequent vehicle traffic in freight areas. The early warning accuracy exceeded 96.9%, with a false alarm rate controlled below 2.1%, reflecting the scientific nature of the dynamic safety distance model and refined collision probability assessment, avoiding the risk of missed warnings while reducing the impact of excessive warnings on operational efficiency. The equipment's fault-free operation time reached a maximum of 7200 hours, verifying the effectiveness of the sensor redundancy design and health monitoring module, ensuring long-term stable system operation, and meeting the high-density, high-load operational needs of large hub airports.

[0040] Example 2: Implementation of Multi-Target Collaborative Surveillance and Conflict Early Warning at General Aviation Airports This embodiment targets a medium-sized airport primarily engaged in general aviation production services while also accommodating regional air logistics transshipment. The airport has one runway, three taxiways, 15 parking stands, and a 50,000㎡ cargo area. Daily operations involve small general aviation aircraft, helicopters, logistics transport vehicles, maintenance vehicles, and ground personnel. The general aviation aircraft are characterized by vertical takeoff and landing and low-speed taxiing. The cargo area primarily uses small and medium-sized transport vehicles, with a moderate operational pace. The airport needs to adapt to the scenario requiring flexible general aviation operations and efficient logistics transshipment. The specific implementation is as follows: 1. System module deployment and parameter configuration Multi-source sensor data acquisition module deployment: Eight millimeter-wave radars are symmetrically deployed 50 meters on each side of the runway centerline on the runway level. The radar beam elevation angle is 3°, the detection range is 50-800 meters, the angular resolution is 0.5°, and the refresh rate is 20Hz. On the taxiway level, one sensor node is set every 200 meters, for a total of 15 nodes. Each node integrates one lidar and two high-definition cameras. The lidar point cloud density is 200 points / ㎡, the ranging accuracy is ±2cm, and the high-definition cameras are installed at a height of 6 meters, with a 30° elevation angle. The panoramic lens covers a radius of 100 meters, the close-up lens has 20x optical zoom, a frame rate of 50fps, and infrared night vision capability for nighttime operations. On the apron level, four... - Six camera positions are grouped into three groups, each equipped with one UWB positioning base station and three ADS-B receivers. The UWB positioning error is ±10cm, and the ADS-B ground station reception radius is 10 kilometers. The mobile station is deployed with eight support vehicles. In the freight area, 20 UWB tags and 15 fixed readers are deployed. The reader recognition distance is ≥5 meters. Five meteorological sensors are deployed simultaneously with a sampling frequency of 1Hz. All sensors are synchronized through the TSN network with a synchronization error of ≤5ms. Key position sensors have 1+1 hot backup, and the fault switching time is ≤100ms. The working environment is adaptable to temperatures from -40℃ to 70℃ and humidity of 0-95% without condensation.

[0041] Data preprocessing and spatiotemporal calibration module configuration: Radar data is denoised using wavelet thresholding (λ=0.8σ, where σ is the standard deviation of the original data) and clutter is removed using 16×16 window CFAR detection; Camera images are denoised and illumination compensated using an improved BM3D algorithm, with ISP wide dynamic range processing enabled in low light conditions, achieving a dynamic range of 120dB; Time synchronization adopts a master-slave clock architecture, with a PTPv2 protocol synchronization message interval of 10ms, jitter ≤1ms, GPS second pulse synchronization accuracy ±1μs, and a local calibration cycle of 100ms; Spatial calibration is based on the airport's three-dimensional coordinate system, with the X-axis along the runway direction, the Y-axis perpendicular to the runway, and the Z-axis vertically upwards. Kalman filtering corrects coordinate deviations in real time, achieving a static calibration error ≤0.1m and a dynamic tracking calibration error ≤0.3m, establishing a 10×10m grid dynamic error map; Rain and snow weather LiDAR data is filtered by raindrop point clouds, achieving an effective data retention rate ≥85%. In case of LiDAR failure, the millimeter-wave radar weight is 0.7, and the camera weight is 0.3.

[0042] The collaborative target identification and tracking module is configured as follows: The multimodal Transformer identification model takes radar point cloud, image features, and ADS-B identity information as input, and outputs five target categories: general aviation small aircraft, helicopters, logistics transport vehicles, maintenance vehicles, and personnel. The identification confidence level is ≥98%, and the identification speed is ≤50ms. A federated distillation optimization model is used, reducing parameters by 60% and improving inference speed by 50%. The tracking algorithm combines improved particle filtering and the Hungarian algorithm, achieving a target unique ID retention rate of ≥99%. For general aviation small aircraft, a tracking mode adapted to vertical takeoff and landing characteristics is enabled, and the state estimator uses an extended Kalman filter. For logistics transport vehicles, load status is identified through lidar point cloud volume calculation, categorized into empty, heavily loaded, and overloaded types, with an error ≤5%. A target motion feature database is established, storing data for small aircraft with a maximum speed of 200km / h and an acceleration of 1.5m / s². 2 The maximum speed of the logistics vehicles is 40 km / h, and the acceleration is 0.8 m / s². 2 Parameters such as; drones and birds are distinguished by micro-Doppler feature analysis with an accuracy of ≥95%; electronic tags are associated with air logistics targets to achieve "vehicle-cargo-flight" traceability with a tracking information update delay of ≤1s.

[0043] Dynamic Path Prediction and Conflict Risk Assessment Module Configuration: The ST-GNN model takes as input the trajectory points of the past 30 seconds, current speed, acceleration, and environmental constraints, and outputs path predictions for the next 10-60 seconds. Short-term deviation is ≤0.5 meters in 10 seconds and long-term deviation is ≤3 meters in 60 seconds. The prediction step size is 10 seconds during busy periods and 60 seconds during idle periods. The prediction frequency is 10Hz when the target speed is ≥50km / h and 5Hz at low speeds. The prediction model for general aviation small aircraft incorporates vertical takeoff and landing constraints, and the path for logistics vehicles is corrected based on loading and unloading plans. The four-dimensional conflict determination space calculates the spatiotemporal intersection of targets, with minimum safe distances between aircraft (runway ≥100 meters, taxiway ≥50 meters, apron ≥20 meters), between aircraft and vehicles ≥30 meters, and between vehicles and personnel ≥10 meters. TTC calculation is combined with relative motion vectors, and risk levels are divided into five levels. Target priority coefficients are: Flight aircraft 1.0, Emergency vehicles 0.9, Logistics vehicles 0.7, Ordinary vehicles 0.5, and Personnel 0.8. Conflict probability integral formula. , =0s, =30s, =30 meters is the safe distance between an aircraft and a vehicle. =1.5 meters, calculated P=0.02; Dynamic safety distance formula The speed of a small aircraft, v = 80 km / h, is equivalent to 22.22 m / s. =1.0, Sunny =0.8, wind speed w=3m / s, road surface dry =0.9, slope r=0.01, calculated as follows =1.0×22.22+0.8×3+0.9×0.01≈22.22+2.4+0.009≈24.629 meters, which meets the safe distance requirement between the aircraft and the vehicle; Acceleration rate formula , =0.1s, =0.3m / s 2 , =1.5m / s 2 Calculated =(1.5-0.3) / 0.1=12m / s 3 It was determined to be a rapid acceleration, and the predicted path was corrected.

[0044] Early warning information generation and push, and system self-optimization module configuration: Five-level risk corresponds to a graded response strategy; early warning information is pushed via 5G slicing communication, airport dedicated communication network, and BeiDou short message service; bilingual (Chinese and English) early warnings for international flights; voice response latency ≤0.3s; AR glasses virtual sign overlay with spatial alignment error ≤0.5 meters; vehicle terminal voice interaction recognition accuracy ≥95%; the system generates a performance report every 24 hours; federated learning nodes include the airport dispatch center, two general aviation companies, and one logistics company; the local model Adam optimizer has an initial learning rate of 0.001; sensor health index ranges from 0 to 100; maintenance prompts are triggered when the index <60; monthly full system calibration; blockchain-stored data supports traceability; equipment remaining life prediction error ≤10%; comprehensive efficiency index based on... calculate.

[0045] 2. Effect Verification and Table Analysis Table 2 Key Operational Indicators of General Aviation Airport System in Example 2

[0046] Table 2 shows data based on 30 consecutive days of system operation monitoring, demonstrating the adaptability and operational effectiveness of this invention in general aviation airport scenarios. The target recognition accuracy remained stable above 97.9%, with accuracy exceeding 98.3% during general aviation operation periods and helicopter take-off and landing periods. This proves that the multimodal recognition model can accurately capture the characteristics of distinctive targets such as small aircraft and helicopters, adapting to their significantly different appearance and movement characteristics compared to ordinary civil aircraft. The tracking ID retention rate exceeded 98.7%, maintaining stable tracking even in scenarios with frequent interaction between maintenance vehicles and personnel and significant target occlusion. This verifies the tracking algorithm's adaptability to low-to-medium speed, multi-interaction targets, meeting the full-process traceability requirements of logistics transfer vehicles. The early warning accuracy exceeded 97.1%, with a false alarm rate controlled below 1.9%, indicating that the dynamic safety distance model and conflict probability assessment effectively adapt to the flexible take-off and landing and low-speed movement characteristics of general aviation targets, avoiding early warning deviations caused by fixed parameters. The equipment has a fault-free operating time of up to 7,000 hours, which fully verifies the reliability and stability of the system under medium-load operation scenarios in general aviation airports. Its lightweight deployment and accurate identification characteristics can effectively support the coordinated and efficient operation of general aviation production services and regional logistics transfer.

[0047] Reference Figure 2 This figure visually demonstrates the target recognition capabilities of this invention in diverse scenarios. The recognition accuracy remains consistently above 97.8% in all scenarios, maintaining a high level even in complex environments such as low visibility fog. This result is attributed to the complementary fusion technology of multi-source sensor data. Data from millimeter-wave radar, lidar, and high-definition cameras complement each other, coupled with environmental adaptation algorithms such as illumination compensation and raindrop cloud removal, effectively overcoming the performance degradation of a single sensor under harsh conditions. The high accuracy in cargo areas and maintenance operation periods proves that the system can accurately identify different types of targets such as freight vehicles and maintenance equipment, adapting to the complex needs of air logistics and ground support, fully demonstrating the generalization ability and environmental adaptability of the multimodal recognition model.

[0048] Reference Figure 3 The figure illustrates the accuracy of the tiered early warning mechanism of this invention. As the risk level increases, the accuracy of the early warning continues to rise, while the false alarm rate decreases significantly. The early warning accuracy for Level 5 emergency risk reaches 99%, with a false alarm rate of only 0.5%. This result stems from the synergistic effect of the four-dimensional conflict determination space and the dynamic safety distance model. By comprehensively considering target distance, speed, priority, and environmental factors, a refined assessment of conflict risk is achieved. While the false alarm rate is relatively high for low-risk levels, it does not affect operations. High-risk levels provide accurate early warnings with an extremely low false alarm rate. This approach avoids safety hazards caused by missed warnings and reduces the interference of excessive early warnings on operational efficiency, perfectly balancing the safety and efficiency of airport surface operations, reflecting the scientific design of the conflict assessment module.

[0049] Reference Figure 4This figure visually demonstrates the self-optimization capability of this invention. The system's initial overall performance index was 85. As the operating time increased, the index continued to rise, reaching 93 after 120 days, and the upward trend gradually stabilized. This is attributed to the synergistic optimization of the federated learning framework and digital twin verification. The system generates a performance report every 24 hours, automatically fine-tuning the identification model and tracking algorithm parameters after accumulating data; simultaneously, the equipment health monitoring module predicts faults in advance, ensuring stable hardware operation. The continuous improvement in the performance index proves that the system can adapt to the dynamic changes in airport operating modes, achieve self-evolution through data accumulation, maintain optimal operating status in the long term, solve the pain point of performance degradation in traditional systems, and provide technical support for the long-term stable operation of airports.

[0050] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-source sensing collaborative monitoring and conflict early warning system for moving targets on the airport surface, characterized in that, Includes the following modules: The multi-source sensor data acquisition module consists of a heterogeneous sensor network composed of millimeter-wave radar, lidar, high-definition camera, ADS-B receiver, UWB positioning module and meteorological sensor; all sensors are synchronized through a time-sensitive network. The data preprocessing and spatiotemporal calibration module uses a cascaded filtering architecture to process the raw data. Radar data is denoised and clutter is removed by detection, and camera images are denoised and illumination compensation is used to correct abnormal areas. Multi-source data are synchronized in time through GPS timing and local crystal oscillator calibration. Spatial calibration is completed based on the airport's three-dimensional coordinate system, an external parameter matrix is ​​established, and coordinate deviations are corrected in real time. The activity target collaborative identification and tracking module, based on the multimodal Transformer model, achieves target classification, takes data as input, and outputs target categories; A multi-target tracking framework combining improved particle filtering and the Hungarian algorithm is adopted, and a unique ID is assigned to each target. During takeoff and landing, a high dynamic tracking mode is activated to handle target motion at speeds ≥300 km / h; The dynamic path prediction module integrates historical target trajectories, airport control rules, and flight schedules to generate predicted paths. It adopts a spatiotemporal graph neural network model, taking parameters from the past 30 seconds as input and outputting path predictions for the next 10-60 seconds. It also introduces a scene adaptive factor to dynamically adjust the prediction step size and frequency. The conflict risk assessment module constructs a four-dimensional conflict determination space and sets the minimum safe distance for different scenarios. Conflict time is calculated based on relative motion vectors, five risk levels are divided, and a weighted risk value is obtained by combining distance and time risks. The early warning information generation and push module generates graded response strategies based on risk levels, and gradually upgrades the processing from system recording to the linkage braking system. The system's self-optimization module constructs a closed-loop iterative mechanism, generates a performance report every 24 hours, updates the model using a federated learning framework, establishes a sensor health assessment model, and triggers maintenance prompts; it also introduces digital twin verification and correction parameters, and conducts monthly system-wide calibration to ensure stable operation.

2. The multi-source sensing airport surface moving target collaborative monitoring and conflict early warning system according to claim 1, characterized in that, It also includes a submodule for refined conflict risk assessment, which calculates the conflict probability distribution of the target over the predicted time period through integration, using the formula: Where P is the conflict probability. To predict the start time, To predict the termination time, Let be the function representing the real-time distance between targets as a function of time. For minimum safe distance, This represents the standard deviation of the distance fluctuation.

3. The multi-source sensing airport surface moving target collaborative monitoring and conflict early warning system according to claim 1, characterized in that, It also includes a dynamic path prediction optimization submodule, which introduces a path prediction model that corrects the rate of change of target motion acceleration. The formula for the rate of change of acceleration is: ,in For the rate of change of acceleration, Let be the target acceleration at time t. for The target acceleration at all times, The sampling time interval is set to 0.1s.

4. The multi-source sensing airport surface moving target collaborative monitoring and conflict early warning system according to claim 1, characterized in that, The multi-source sensor data acquisition module adopts a layered deployment and redundancy design; eight millimeter-wave radars are symmetrically deployed 50 meters on both sides of the runway centerline at the runway level to form 360° coverage without blind spots, and the radar beam elevation angle is set to 3° to simultaneously detect low-altitude aircraft and ground targets; one sensor node is set every 200 meters at the taxiway level; the apron level is grouped according to parking positions, with each group of 4-6 parking positions configured with one UWB positioning base station and three ADS-B receivers, and the base stations are deployed at the corners of the parking positions to form triangulation positioning; at the cargo logistics level, UWB tags and fixed readers are deployed at the entrances and exits of the cargo station and along the conveyor belt; the key position sensors adopt a 1+1 hot backup.

5. The multi-source sensing airport surface moving target collaborative monitoring and conflict early warning system according to claim 1, characterized in that, The data preprocessing and spatiotemporal calibration module introduces deep learning-assisted calibration, constructs a CNN-based sensor deviation prediction model, takes historical sensor deviation data, meteorological parameters, and equipment runtime as input, and outputs real-time deviation compensation values; time synchronization adopts a master-slave clock architecture, with slave devices synchronizing via the PTPv2 protocol; Spatial calibration establishes a dynamic error map, dividing the airport into a 10×10 meter grid, with each grid storing the average calibration error for that area.

6. The multi-source sensing airport surface moving target collaborative monitoring and conflict early warning system according to claim 1, characterized in that, The collaborative target identification and tracking module employs federated distillation technology to optimize the identification model. A general feature extractor is generated by the central server, and each sensor node trains a category classifier locally. The classifier knowledge is then distilled into the central model. The tracking algorithm incorporates target intent recognition, predicting the target's next action by analyzing the target's historical trajectory and airport control instructions. For small targets such as drones, micro-Doppler feature analysis is used to extract the Doppler frequency shift of rotating parts from radar echoes. In air logistics target tracking, cargo electronic tag information is linked.

7. The multi-source sensing airport surface moving target collaborative monitoring and conflict early warning system according to claim 1, characterized in that, The dynamic path prediction module, combined with the airport's digital twin system, integrates real-time sensor data with virtual scenes to simulate the target's movement trend in virtual space, and the prediction results are presented through three-dimensional visualization; the path planning algorithm incorporates ecological driving factors and adjusts the suggested speed for ground vehicles based on conflict risks. For flight taxiing paths, the timing of acceleration / deceleration is optimized by combining engine thrust characteristics and taxiway gradient. The prediction model has self-learning capabilities. When the deviation between the actual trajectory and the predicted trajectory exceeds the threshold for three consecutive times, the model parameters are automatically fine-tuned. The online gradient descent method is used to update the weights to adapt to changes in airport operation modes.

8. The multi-source sensing airport surface moving target collaborative monitoring and conflict early warning system according to claim 1, characterized in that, The conflict risk assessment module establishes a dynamic safety distance model, in which the safety distance adaptively adjusts according to target speed, weather conditions, and road surface conditions. The formula is as follows: ,in For dynamic safety distance, Here, v is the velocity coefficient, and v is the target velocity. Here, w is the meteorological coefficient, and w is the wind speed. denoted as pavement coefficient, and r as pavement slope.

9. The multi-source sensing airport surface moving target collaborative monitoring and conflict early warning system according to claim 1, characterized in that, The early warning information generation and push module has augmented reality visualization capabilities. When dispatchers wear AR glasses, early warning information is overlaid on the actual scene through virtual icons. For ground support vehicles, the vehicle terminal integrates voice interaction, supporting the query of early warning details or confirmation of avoidance operations through voice commands. An early warning effect feedback mechanism is established to record the target response behavior after each early warning, calculate the early warning effectiveness rate, and automatically adjust the early warning method or content when the effectiveness rate is less than 80%.

10. The multi-source sensing airport surface moving target collaborative monitoring and conflict early warning system according to claim 1, characterized in that, The system's self-optimization module uses blockchain technology to ensure data credibility. Each sensor node acts as a blockchain node, encrypting collected data before storing it on the chain, and verifying data integrity through hash value checks. Equipment health monitoring incorporates remaining life prediction, analyzing historical sensor fault data, operating temperature, and vibration intensity to predict remaining lifespan using an LSTM neural network. System performance evaluation incorporates a comprehensive efficiency index, calculated using the following formula: Where E is the comprehensive performance index, A is the identification accuracy, T is the tracking continuity, W is the early warning accuracy, and F is the equipment fault-free operation time. Deep optimization is initiated when E < 80.

Images