An aircraft intelligent trajectory prediction method based on deep learning big data

Abstract

The application discloses an aircraft intelligent trajectory prediction method based on deep learning big data, and belongs to the technical field of trajectory prediction, and the method comprises the following steps: obtaining target aircraft multi-source observation and low-altitude scene data, fusing to form a historical trajectory sequence and generating a trajectory credibility label; extracting low-altitude scene constraint objects and determining their action attributes; generating a layered low-altitude reachable domain based on target state, motion ability and environmental constraints, and performing reachability counterexample filtering on the constraint objects; combining the credibility label and the effective objects after filtering to generate a candidate maneuver behavior chain; inputting the above features into a deep behavior chain prediction model to output the occurrence probability and arrival time distribution of each behavior chain; and generating main and backup prediction trajectories in combination with the reachable domain; the application effectively improves the accuracy and scene rationality of trajectory prediction in a complex low-altitude scene by combining physical reachability constraints with deep learning.

Description

Technical Field

[0001] This invention relates to the field of trajectory prediction technology, and in particular to an intelligent trajectory prediction method for aircraft based on deep learning big data. Background Technology

[0002] With the development of low-altitude economy, drone inspection, logistics delivery, plant protection operations, security monitoring, and low-altitude control applications, the number of aircraft in low-altitude airspace is constantly increasing. Accurate prediction of the future movement of aircraft has become a crucial technical aspect of low-altitude surveillance, risk warning, path conflict assessment, and control and management. This is especially true for low-altitude, slow, and small aircraft, whose flight areas are typically close to buildings, mountains, power towers, roads, rivers, transmission lines, no-fly zones, electronic fences, and other sensitive areas, making their future trajectories highly correlated with these environments.

[0003] Existing aircraft trajectory prediction methods typically extrapolate from historical trajectory sequences or input historical position, speed, heading, altitude, and other state data into a neural network model, which then directly outputs future trajectory points. These methods can achieve certain predictive results in scenarios where the target motion is relatively continuous, the observation data is relatively stable, and the operating environment is relatively open. Furthermore, the ability to fit general motion trends can be improved by increasing the amount of historical samples and the model complexity.

[0004] However, in complex low-altitude scenarios, an aircraft's future movement is not only affected by its own speed, turning, climb, or descent capabilities, but also by scene factors such as obstacles, no-fly zones, passageways, suspected take-off and landing points, return points, and observation attenuation areas. If the trajectory prediction process relies solely on historical trajectory changes, or only uses low-altitude map information as ordinary input features, the prediction results are prone to problems such as crossing obstacles, deviating from the actual reachable range, and failing to reflect maneuvering behaviors such as avoidance, flight along the route, approach, return, and hovering.

[0005] Therefore, how to improve the accuracy and scenario rationality of aircraft trajectory prediction in complex low-altitude scenarios, so that the prediction results can better match the actual reach of the aircraft and the continuous maneuvering behavior that may occur, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] One of the objectives of this invention is to provide an intelligent trajectory prediction method for aircraft based on deep learning big data, in order to solve the problem that in the existing technology, when predicting aircraft trajectories in complex low-altitude scenarios, the prediction process is difficult to fully combine low-altitude scenario constraints, the actual reachability of the target, and possible continuous maneuvering behaviors, resulting in insufficient accuracy of the predicted trajectory and scenario rationality.

[0007] This invention is achieved through the following technical solution:

[0008] On the one hand, a method for intelligent aircraft trajectory prediction based on deep learning big data is provided, including the following steps:

[0009] Acquire multi-source observation data and low-altitude scene data of the target aircraft, perform spatiotemporal alignment and target association on the multi-source observation data to form a historical trajectory sequence, and generate trajectory confidence markers for trajectory points in the historical trajectory sequence;

[0010] Extract low-altitude scene constraint objects from the low-altitude scene data, and determine the effect attributes based on how the low-altitude scene constraint objects affect the future motion of the target aircraft.

[0011] Based on the current state, maneuverability, and low-altitude environmental constraints of the target aircraft, a hierarchical low-altitude reachability domain is generated. The reachability of the low-altitude scenario constraint objects is then filtered by the hierarchical low-altitude reachability domain to obtain effective low-altitude scenario constraint objects for generating candidate maneuver behavior chains.

[0012] An effective observation window is formed based on the historical trajectory sequence and the corresponding trajectory confidence marker, and combined with the effective low-altitude scene constraint object, a candidate maneuver behavior chain that the target aircraft may experience in the future prediction time window is generated.

[0013] The effective observation window, the trajectory confidence marker, the effective low-altitude scene constraint object, the hierarchical low-altitude reachability domain, and the candidate maneuvering behavior chain are input into the deep behavior chain prediction model, and the occurrence probability and arrival time distribution of each candidate maneuvering behavior chain are output.

[0014] Based on the occurrence probability, the arrival time distribution, and the layered low-altitude reachability, the primary predicted trajectory and backup predicted trajectory of the target aircraft are generated.

[0015] Furthermore, the multi-source observation data includes at least one of radar detection data, photoelectric detection data, infrared detection data, radio spectrum detection data, acoustic detection data, Remote ID data, telemetry data, and historical trajectory record data;

[0016] Spatiotemporal alignment and target association are performed on the multi-source observation data, including converting observation points from different data sources to a unified spatial coordinate system and resampling according to a unified time reference;

[0017] When there are observation points from multiple sources at the same time, they are fused based on source confidence, position consistency, and motion continuity to form the target aircraft state at the corresponding time.

[0018] Further, generating trajectory confidence markers for trajectory points in the historical trajectory sequence includes:

[0019] The source confidence evaluation item, time continuity evaluation item, speed rationality evaluation item, altitude change rationality evaluation item, heading change rationality evaluation item, multi-source observation consistency evaluation item, and observation attenuation status evaluation item for the current trajectory point are calculated respectively. The above evaluation items are then weighted, summed, and normalized to obtain the trajectory confidence label for the trajectory point.

[0020] The speed rationality evaluation item is determined based on the deviation between the speed calculated from adjacent locations and the fused observation speed; the multi-source observation consistency evaluation item is determined based on the maximum or average deviation between the positioning locations of multiple data sources at the same time; and the observation attenuation status evaluation item is determined based on whether the target aircraft is in a radar blind zone, an electro-optical line-of-sight obstruction zone, or a weak communication coverage zone.

[0021] Furthermore, the functional attributes of the low-altitude scene constraint object include approach attribute, avoidance attribute, passage attribute, and observation attenuation attribute;

[0022] The constrained objects in low-altitude scenarios with proximity attributes include sensitive areas, inspection points, suspected take-off and landing points, historical return points, and drone nests or control points.

[0023] The constraints in low-altitude scenarios with avoidance attributes include buildings, mountains, power towers, viaducts, no-fly zones, or electronic fence boundaries;

[0024] Low-altitude scene constraints with passage attributes include roads, rivers, railways, pipelines, power transmission lines, walls, or low-altitude passages;

[0025] Constraints in low-altitude scenarios with observation attenuation properties include radar blind spots, electro-optical line-of-sight obstruction areas, or weak communication coverage areas.

[0026] Furthermore, the method also includes calculating an action competition quantity for each low-altitude scene constraint object, wherein the action competition quantity is used to characterize the degree of influence of the corresponding low-altitude scene constraint object on the future motion of the target aircraft within the current prediction period;

[0027] The action competition quantity is obtained by weighted calculation based on distance correlation, heading correlation, velocity trend correlation, altitude correlation, historical matching, and observation confidence correlation. The observation confidence correlation is determined based on the correlation between the trajectory confidence marker and trajectory state and the low-altitude scene constraint object, and is used to reduce the impact of low-confidence trajectory points on the calculation of the action competition quantity.

[0028] Further, generating the layered low-altitude reachability domain includes:

[0029] The future prediction time window is divided into multiple prediction steps; based on the set of reachable states in the previous layer, possible states are expanded under the control constraints of the target aircraft's maximum speed, maximum acceleration, maximum turning rate, maximum descent rate, and maximum climb rate.

[0030] The possible states obtained from the expansion are corrected based on the low-altitude wind field; and possible states that fall into obstacle areas, no-fly zones or impassable areas are eliminated, and layered low-altitude reachable domains are generated layer by layer to represent the actual three-dimensional spatial range that the target aircraft can reach at different predicted times.

[0031] Furthermore, the accessibility disproving filter for the low-altitude scene constraint object is performed using the layered low-altitude reachability domains, including: calculating the reachability correlation value between the low-altitude scene constraint object and each layer of low-altitude reachability domains respectively, and determining the reachability disproving value based on the reachability correlation value corresponding to each layer of low-altitude reachability domains.

[0032] When the low-altitude scene constraint object is a point object, the reachability association value is determined based on the minimum distance from the point object to the set of state points in the corresponding low-altitude reachable domain.

[0033] When the low-altitude scene constraint object is a linear object, the reachability association value is determined based on the overlap ratio or minimum distance between the linear object and the corresponding layer's low-altitude reachability domain projection area.

[0034] When the low-altitude scene constraint object is a planar object or a volumetric object, the reachability association value is determined based on the overlap area, overlap volume, or shortest boundary distance between the planar object or volumetric object and the corresponding low-altitude reachable domain.

[0035] If the connecting path from the current state to the low-altitude scene constraint object crosses the low-altitude scene constraint object with avoidance attribute, obstacle area, or no-fly boundary, then the reachable association value is reduced or set to zero.

[0036] When the reachability disproven value is greater than the preset disproven threshold, it is excluded from the valid low-altitude scene constraint objects.

[0037] Further, generating the candidate maneuver behavior chain includes:

[0038] Based on the valid low-altitude scene constraint objects retained after the accessibility counter-proof screening, candidate maneuver behavior chains consisting of multiple maneuver behavior stages connected sequentially are generated in chronological order.

[0039] Each maneuver phase includes the maneuver state, the dominant low-altitude scene constraint object, and the corresponding phase time range.

[0040] The maneuvering behavior states include at least one of the following: straight-line crossing, flying along a low-altitude scene constraint object with a passage attribute, avoiding a low-altitude scene constraint object with a dodge attribute, hovering or low-speed observation, approaching a sensitive area, returning home, circling, descending or landing, entering an observation attenuation region, and reappearing from an observation attenuation region.

[0041] Furthermore, the deep behavior chain prediction model includes a temporal encoder, a behavior chain encoder, and a fusion scorer;

[0042] The timing encoder is used to extract historical motion codes containing the trajectory confidence markers;

[0043] The behavior chain encoder is used to extract the behavior chain code of the candidate maneuver behavior chain;

[0044] The fusion scorer is used to fuse the historical motion code, the behavior chain code, and the hierarchical low-altitude reachability features of the corresponding candidate maneuver behavior chains, and outputs the occurrence probability of each candidate maneuver behavior chain, the expected arrival time of the corresponding dominant low-altitude scene constraint object, and the arrival time uncertainty.

[0045] Furthermore, the training samples of the deep behavior chain prediction model include historical observation windows, trajectory confidence markers, low-altitude scene constraint objects, hierarchical low-altitude reachability domains, candidate maneuvering behavior chains, and real subsequent trajectories. The real subsequent trajectories are used to determine the candidate maneuvering behavior chain labels and arrival time labels.

[0046] Further, generating the primary prediction trajectory and the backup prediction trajectory includes:

[0047] The spatial extension of each candidate maneuvering behavior chain is restricted to the hierarchical low-altitude reachable domain of the corresponding time layer to generate the corresponding candidate trajectory.

[0048] Based on the occurrence probability of each candidate maneuvering behavior chain, combined with the reachability consistency evaluation used to characterize whether the candidate trajectory meets the reachability connectivity requirements of the corresponding time layer, the scene constraint consistency evaluation used to characterize whether the candidate trajectory matches the dominant low-altitude scene constraint object, and the trajectory credibility support evaluation based on trajectory credibility markers to characterize the degree of support for historical motion trends, the comprehensive score of each candidate trajectory is calculated by weighting.

[0049] The candidate trajectory with the highest comprehensive score is selected as the main predicted trajectory, and several candidate trajectories with high scores and different candidate maneuver behavior chain types are selected as backup predicted trajectories.

[0050] Furthermore, it also includes a discontinuous observation posterior update step: when the target aircraft is briefly lost, the current main predicted trajectory and backup predicted trajectory are maintained, and the corresponding candidate maneuver behavior chain is continued to be expanded;

[0051] When the target aircraft reappears and acquires a new observation status, the matching degree is determined based on the positional and heading deviations between the new observation status and the predicted status of each candidate maneuvering behavior chain at the current moment. The matching degree is then used to perform a posteriori update on the probability of occurrence of each candidate maneuvering behavior chain before the target was lost, so as to correct the main predicted trajectory and the backup predicted trajectory.

[0052] On the other hand, a deep learning-based intelligent trajectory prediction system for aircraft is provided, including:

[0053] The multi-source observation fusion and labeling module is used to acquire multi-source observation data and low-altitude scene data of the target aircraft, perform spatiotemporal alignment and target association on the multi-source observation data, form a historical trajectory sequence, and generate trajectory confidence labels for trajectory points in the historical trajectory sequence.

[0054] The constraint object extraction and attribute determination module is used to extract low-altitude scene constraint objects from the low-altitude scene data and determine the effect attributes based on the way the low-altitude scene constraint objects affect the future motion of the target aircraft.

[0055] The reachability domain generation and disproving module is used to generate a hierarchical low-altitude reachability domain based on the current state, maneuverability and low-altitude environmental constraints of the target aircraft, and to use the hierarchical low-altitude reachability domain to perform reachability disproving and filtering of the low-altitude scenario constraint objects to obtain effective low-altitude scenario constraint objects for generating candidate maneuvering behavior chains.

[0056] The candidate behavior chain generation module is used to form an effective observation window based on the historical trajectory sequence and the corresponding trajectory confidence marker, and in combination with the effective low-altitude scene constraint object, generate candidate maneuver behavior chains that the target aircraft may experience in the future prediction time window.

[0057] The depth prediction module is used to input the effective observation window, the trajectory confidence marker, the effective low-altitude scene constraint object, the hierarchical low-altitude reachability domain, and the candidate maneuver behavior chain into the depth behavior chain prediction model, and output the occurrence probability and arrival time distribution of each candidate maneuver behavior chain;

[0058] The primary and backup trajectory generation module is used to generate the primary predicted trajectory and backup predicted trajectory of the target aircraft based on the occurrence probability, the arrival time distribution, and the layered low-altitude reachability domain.

[0059] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0060] 1. This invention effectively overcomes trajectory jumps and noise interference caused by radar blind spots, photoelectric blockage, and signal discontinuity in complex low-altitude scenarios by performing spatiotemporal fusion of multi-source observation data and innovatively generating trajectory credibility markers for trajectory points. It also weakens the weight of abnormal trajectory points within the effective observation window, significantly improving the reliability and anti-interference capability of model input from the data source.

[0061] 2. This invention transforms conventional low-altitude map information into low-altitude scene constraint objects with functional attributes and dynamically calculates their functional competition quantity, enabling the prediction framework to intelligently filter out scene entities that truly affect the target's movement within the current period (such as avoiding buildings, flying along roads, or approaching landing points). This solves the technical defects of existing pure data-driven models that easily output trajectories that defy common sense, such as "passing through walls" or deviating from the actual flight path, and greatly improves the scene rationality of trajectory prediction.

[0062] 3. This invention deeply integrates physical laws and artificial intelligence. By combining aircraft maneuverability with low-altitude wind fields to construct a layered low-altitude reachability domain, and using this domain to perform reachability verification screening of scene constraint objects, the invention eliminates unreachable or abnormal states that cross obstacles using rigorous physical and kinematic boundaries before generating candidate maneuvers. This ensures that subsequent deep model predictions are strictly limited to a physically feasible safe space, significantly improving the accuracy and realism of the prediction results.

[0063] 4. This invention outputs the occurrence probability and arrival time distribution of candidate behavior chains with clear temporal logic through a deep behavior chain prediction model, thereby generating primary and backup prediction trajectories. This not only provides multi-level early warning references with both spatiotemporal dimensions for low-altitude control, but also introduces a discontinuous observation posterior update mechanism. When a target is lost for a short time and then reappears, the prediction probability can be dynamically corrected by using the matching degree between the re-observed state and the predicted state, which greatly improves the continuity and robustness of trajectory prediction under complex low-altitude discontinuous tracking conditions. Attached Figure Description

[0064] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0065] Figure 1 This is a schematic diagram of the flow steps of the prediction method provided in Embodiment 1 of the present invention.

[0066] Figure 2 This is a schematic diagram of the unified coordinate system and spatiotemporal alignment fusion of multi-source observation data in Embodiment 1 of the present invention.

[0067] Figure 3 This is a curve comparing the trajectory positioning errors before and after multi-source observation fusion in Embodiment 1 of the present invention;

[0068] Figure 4 This is a heatmap showing the changes in trajectory reliability evaluation items during different flight phases in Embodiment 1 of the present invention.

[0069] Figure 5 This is a dynamic evolution diagram of the multi-object interaction competition quantity on the flight trajectory of the target aircraft in Embodiment 1 of the present invention;

[0070] Figure 6 This is a corrected reachability state deviation diagram of the target aircraft affected by the low-altitude wind field in Embodiment 1 of the present invention;

[0071] Figure 7 This is a distribution map of weakened low-confidence abnormal trajectory points within the effective observation window in Embodiment 1 of the present invention;

[0072] Figure 8 This is the output diagram of the predicted arrival time distribution of different maneuvering behaviors by the deep behavior chain prediction model in Embodiment 1 of the present invention;

[0073] Figure 9 This is a comparison chart showing the improvement in training performance of the deep behavioral chain prediction model by incorporating low-altitude reachability physical constraints in Embodiment 1 of the present invention;

[0074] Figure 10 This is a posterior update curve of the probability of occurrence of candidate maneuver behavior chains under discontinuous target observation in Embodiment 1 of the present invention;

[0075] Figure 11 This is a system block diagram provided in Embodiment 2 of the present invention. Detailed Implementation

[0076] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0077] Example 1:

[0078] This embodiment discloses an intelligent trajectory prediction method for aircraft based on deep learning big data. This method is mainly applied to scenarios such as low-altitude airspace surveillance, UAV management and control, security of key areas, low-altitude route conflict early warning, trajectory simulation of patrol aircraft, monitoring of low, slow and small targets, and continuous trajectory prediction after short-term target loss.

[0079] The target aircraft can be drones, inspection drones, agricultural drones, logistics drones, model aircraft, racing drones, or other low-altitude flying targets.

[0080] Specifically, please refer to Figure 1 The method in this embodiment includes the following steps.

[0081] S100: Acquire multi-source observation data and low-altitude scene data of the target aircraft, perform spatiotemporal alignment and target association on the multi-source observation data, form a historical trajectory sequence, and generate trajectory confidence markers for trajectory points in the historical trajectory sequence.

[0082] Among them, multi-source observation data is used to describe the motion state of the target aircraft at the current time and historical time. It can include one or more of the following: radar detection data, photoelectric detection data, infrared detection data, radio spectrum detection data, acoustic detection data, Remote ID (unmanned aerial vehicle) data, telemetry data, and historical trajectory record data. Each observation point can include at least the observation time, target position, target altitude, target speed, target heading, vertical speed, data source, source confidence level, and target identification number.

[0083] In this embodiment, since different observation devices have different sampling frequencies, coordinate systems, and positioning errors, it is necessary to first convert the observation points from different data sources to a unified spatial coordinate system.

[0084] Specifically, latitude and longitude data can be converted into a local northeast-sky coordinate system; radar data can be converted into three-dimensional rectangular coordinates based on distance, azimuth, and elevation angles; photoelectric angle measurement data can be converted into spatial coordinates based on the installation location of the photoelectric equipment, line-of-sight angle, and distance measurement information; and Remote ID data or telemetry data can be unified based on their own positioning coordinates and timestamps.

[0085] After unifying the spatial coordinates, the observation points from different data sources are resampled according to a unified time reference, and the targets are associated based on the target identification number, spatial distance threshold, velocity continuity, heading continuity, altitude continuity, and source confidence. When there are observation points from multiple sources at the same time, they can be fused based on source confidence, position consistency, and motion continuity to form the target aircraft status at the corresponding time.

[0086] It is understood that the target aircraft status in this embodiment includes information such as target position, altitude, speed, heading, vertical speed, and comprehensive source confidence. Multiple target aircraft statuses are arranged in chronological order to form a historical trajectory sequence.

[0087] To more intuitively illustrate the spatiotemporal alignment and fusion process and its effects, please refer to [link to relevant documentation]. Figure 2 , Figure 2This is a schematic diagram of the unified coordinate system and spatiotemporal alignment and fusion of multi-source observation data in Embodiment 1 of the present invention; it can be seen that the original radar and photoelectric observation points often exhibit certain discreteness and systematic offsets under the unified coordinate system, while after alignment and fusion, a coherent and smooth trajectory is formed. Further combined with... Figure 3 As shown in the trajectory positioning error comparison curve before and after multi-source observation fusion in Embodiment 1 of the present invention, compared with the error jump that is very easy to occur in a specific area by a single observation source, the state error after fusion of this scheme always remains at a very low and stable level.

[0088] After forming the aforementioned historical trajectory sequence, this embodiment can further generate trajectory confidence markers for the trajectory points in the historical trajectory sequence. This is because target aircraft in complex low-altitude scenarios are prone to issues such as weak radar echoes, electro-optical blockage, intermittent radio signals, inconsistent multi-source positioning, trajectory jumps, and short-term loss. If such observation points are directly used as input to the model as the actual trajectory, the model may mistakenly identify low-confidence observations as actual maneuvers.

[0089] Therefore, to avoid the above problems, this scheme can use trajectory reliability markers to characterize the reliability of the corresponding trajectory points in reflecting the true motion state of the target aircraft. The expression is as follows:

[0090] ;

[0091] In the formula, Indicates the first Trajectory confidence markers for each trajectory point; Represents the normalization function; The source confidence evaluation item can be determined based on the equipment accuracy, current operating conditions, and historical reliability of the observation data source. The evaluation item for time continuity can be determined based on whether the time interval between adjacent observation points meets the sampling continuity requirement; The velocity rationality evaluation item can be determined based on the deviation between the velocity calculated from adjacent locations and the fused observation velocity; The evaluation item for the rationality of altitude change can be determined based on the matching relationship between the altitude change of adjacent trajectory points and the climb or descent capability of the target aircraft; This indicates the evaluation item for the reasonableness of the course change, which is determined based on whether the course change between adjacent trajectory points exceeds the turning capability of the target aircraft; This represents a multi-source observation consistency evaluation item, determined based on the maximum or average deviation between the location positions of multiple data sources at the same time. The observation attenuation status evaluation item is determined based on whether the target aircraft is in a radar blind zone, an electro-optical line-of-sight obstruction zone, or a weak communication coverage zone. to These are the weights of the corresponding evaluation items.

[0092] It should be noted that the quantitative calculation of each evaluation item (such as the speed rationality evaluation item mentioned above) in this embodiment can be performed using a step threshold method, an exponential decay function, or an inverse proportional function for normalization mapping. For example, when the speed deviation is less than the safety threshold, the evaluation item takes a value of 1; when the deviation is greater than the safety threshold, an exponential function is used. (in For speed deviation, (where the attenuation coefficient is used) to smoothly decay the evaluation item to near 0. Furthermore, the weights of each item in the above formula ( to The weights can be calculated using objective weighting methods such as the Analytic Hierarchy Process (AHP) and entropy weighting, or determined by domain experts based on the application scenario (e.g., lowering the weight in water areas where radar is susceptible to interference). or The weighting of radar components is pre-calibrated based on prior experience.

[0093] Understandably, the purpose of using the aforementioned trajectory confidence labeling is not only to enable subsequent models to obtain the target aircraft's historical position, speed, and heading changes, but also to allow them to understand the reliability of each historical trajectory point. Therefore, in subsequent processes such as constructing effective observation windows, calculating the competitive forces of constraints in low-altitude scenarios, generating candidate maneuver chains, and comprehensive trajectory scoring, low-confidence trajectory points will not have the same impact as high-confidence trajectory points, thereby reducing the interference of jump points, occlusion, multi-source bias, and short-term weak observations on the prediction results. As... Figure 4 As shown, this is a heatmap illustrating the changes in trajectory reliability evaluation items during different flight stages in this embodiment of the invention. The heatmap demonstrates that the trajectory reliability labeling formula constructed in this embodiment can accurately quantify the dynamic changes of various evaluation sub-items during different stages, such as stable flight, entering blind zones, or experiencing trajectory jumps. This provides objective data support for the effective screening and mitigation of anomaly observations.

[0094] S200. Extract low-altitude scene constraint objects from low-altitude scene data, and determine the effect attributes based on how the low-altitude scene constraint objects affect the future motion of the target aircraft.

[0095] In the above steps, low-altitude scene data is used to describe objects in the low-altitude environment where the target aircraft is located that may affect its future trajectory. This may include 3D map data, low-altitude airspace management data, no-fly zone data, electronic fence data, road data, river data, railway data, pipeline data, power transmission line data, suspected take-off and landing point data, historical return point data, UAV nest data, control point data, radar blind spot data, electro-optical line-of-sight obstruction area data, weak communication coverage area data, and low-altitude meteorological data.

[0096] This embodiment extracts low-altitude scene constraint objects from low-altitude scene data; wherein, low-altitude scene constraint objects refer to low-altitude scene objects that may affect the target aircraft's direction of motion, speed, altitude, reachability, or observation reliability within the future prediction time window.

[0097] For example, based on geometric morphology, low-altitude scene constraint objects include point objects, line objects, area objects, and volume objects. Among them, point objects can be suspected take-off and landing points, historical return points, drone nests, control points, or inspection points; line objects can be roads, rivers, railways, pipelines, power transmission lines, walls, or low-altitude corridors; area objects can be sensitive areas, no-fly zones, electronic fence areas, or areas with weak communication coverage; and volume objects can be buildings, mountains, viaducts, power tower spatial envelopes, or photoelectric line-of-sight obstructions.

[0098] Furthermore, in this embodiment, based on the way the low-altitude scene constraint object affects the future motion of the target aircraft, the effect attribute of the low-altitude scene constraint object is determined as an approach attribute, avoidance attribute, passage attribute, or observation attenuation attribute.

[0099] For example, low-altitude scene constraints with proximity attributes include sensitive areas, inspection points, suspected take-off and landing points, historical return points, UAV nests, or control points; low-altitude scene constraints with avoidance attributes include buildings, mountains, power towers, viaducts, no-fly zones, or electronic fence boundaries; low-altitude scene constraints with passage attributes include roads, rivers, railways, pipelines, power transmission lines, walls, or low-altitude corridors; and low-altitude scene constraints with observation attenuation attributes include radar blind spots, electro-optical line-of-sight obstruction areas, or areas with weak communication coverage.

[0100] It is understandable that the same low-altitude scene object can have one or more functional attributes in different application scenarios. For example, a building can serve as a avoidance attribute object for flight paths, and can also form an observation attenuation attribute object for photoelectric observation; a power transmission line can serve as a passage attribute object in inspection tasks, and can also serve as a avoidance attribute object that requires maintaining a safe distance in non-inspection tasks; and the above-mentioned functional attributes can be pre-configured according to task type, target category, low-altitude management rules, and monitoring equipment deployment method.

[0101] When multiple low-altitude scene constraints exist around the target aircraft, it is necessary to determine which objects are more likely to affect the future movement of the target aircraft within the current prediction period.

[0102] Therefore, this embodiment can calculate the action contention quantity for each low-altitude scene constraint object. Specifically, the action contention quantity for a low-altitude scene constraint object can be expressed as follows:

[0103] ;

[0104] In the formula, Indicates the first The competitive force of the constraint objects in a low-altitude scene; Indicates distance-related items; Indicates the heading-related item; Indicates the correlation terms of speed trend; Indicates highly related items; Indicates historical matches; Indicates the correlation term for observation confidence; to For the corresponding weights.

[0105] In specific quantization, each related term can be implemented using the following physical or geometric mapping function:

[0106] For example, the distance correlation term can be calculated based on the reciprocal of the spatial distance between the target aircraft and the low-altitude scene constraint object or a Gaussian decay function; the heading correlation term... It can be based on the cosine value of the angle between the current velocity vector direction of the target and the direction of the line connecting the target to the constrained object. The mapping shows that the smaller the included angle, the closer the cosine value is to 1, indicating a more consistent heading. The calculation of other correlation terms can be achieved by setting corresponding geometric or physical matching functions; velocity trend correlation terms... The height correlation term can be determined by normalization based on the magnitude of the projection of the target's current velocity vector in the direction pointing towards the object; It can be obtained by decaying a negative exponential function based on the difference between the target's current height and the object's height range; historical matching items. The setting can be based on the frequency proportion of interactions with this type of object in historical trajectory data (such as past approach or avoidance behaviors); observation reliability correlation items This can be directly positively correlated with the trajectory confidence markers generated in the aforementioned steps. Through the above specific quantitative calculations, the multidimensional physical space state can be uniformly mapped to a dimensionless competitive score.

[0107] In addition, the weights of each item in the formula for the amount of competition ( to Similarly, it can be set based on the above-mentioned hierarchical analysis method or according to the control importance of low-altitude scene attributes (such as appropriately increasing the weight of related items of avoidance attributes).

[0108] It is also understandable that the distance correlation term is used to characterize the distance relationship between the target aircraft's current position and the low-altitude scene constraint object. The closer the distance, the more likely the object is to influence the target's future movement within the current prediction period. The heading correlation term is used to characterize the consistency between the target aircraft's current heading and the direction pointing to the low-altitude scene constraint object or the direction in which the low-altitude scene constraint object extends. The speed trend correlation term is used to characterize whether the target aircraft's speed change trend points towards or away from the low-altitude scene constraint object. The altitude correlation term is used to characterize the degree of matching between the target aircraft's current altitude or altitude change trend and the altitude range of the low-altitude scene constraint object. The historical matching term is used to characterize whether the target aircraft has ever flown along, approached, avoided, or entered the observation attenuation region of this type of object in the historical trajectory. The observation confidence correlation term is determined based on the trajectory confidence label and the degree of correlation between the trajectory state and the low-altitude scene constraint object, and is used to reduce the impact of low-confidence trajectory points on the calculation of the action competition quantity.

[0109] Therefore, by setting a competitive factor, low-altitude scene objects can be transformed from ordinary map information into constraint objects that can participate in trajectory prediction. This ensures that subsequent candidate maneuver behavior chains are not generated indiscriminately from all map objects, but are preferentially generated around low-altitude scene constraint objects that are more likely to affect the target's future movement within the current prediction period. This reduces the interference of irrelevant scene objects on the model input and improves the scene rationality of the prediction results. Figure 5 The diagram shown in this embodiment illustrates the dynamic evolution of the multi-object interaction competition on the target aircraft's flight trajectory. As the target flies, the interaction competition of different constraint objects exhibits a physical dynamic alternation, allowing the system to accurately pinpoint the dominant scene constraint object within the current time period.

[0110] S300: Generate a hierarchical low-altitude reachability domain based on the target aircraft's current state, maneuverability, and low-altitude environmental constraints. Then, use the hierarchical low-altitude reachability domain to perform reachability verification and screening of low-altitude scenario constraint objects to obtain effective low-altitude scenario constraint objects for generating candidate maneuver behavior chains.

[0111] The current state of the target aircraft can be determined by the latest high-confidence trajectory point in the historical trajectory sequence, or by a comprehensive estimate of multiple recent high-confidence trajectory points. The target aircraft's motion capabilities include maximum speed, maximum acceleration, maximum turn rate, maximum descent rate, maximum climb rate, minimum turning radius, or a speed range limited by the mission scenario. Furthermore, low-altitude environmental constraints include obstacle areas, no-fly zones, impassable areas, low-altitude passage boundaries, terrain altitude, building height, and low-altitude wind fields.

[0112] In one implementation, the future prediction time window is divided into multiple prediction steps, and under the condition of satisfying the motion capability and low-altitude environment constraints, the low-altitude reachability range of the target aircraft at different prediction times is generated layer by layer, and the low-altitude reachability ranges at each prediction time are combined in chronological order to form the layered low-altitude reachability domain.

[0113] Specifically, the expression for this layered low-altitude reachability domain is:

[0114] ;

[0115] in, Indicates the low-altitude reachability of the layered system; Indicates the target aircraft in the The predicted time or the first A set of three-dimensional spatial states that may actually be reached within a predicted time layer; This indicates the number of prediction steps obtained by dividing the future prediction time window. This number of prediction steps is the number of layers in the layered low-altitude reachable domain.

[0116] When generating each low-altitude reachable domain, the current state is used as the initial reachable state, and the possible states of the next layer are expanded based on the reachable state set of the previous layer. During the expansion process, the motion capability constraints of the target aircraft, such as maximum speed, maximum acceleration, maximum turning rate, maximum descent rate, and maximum climb rate, need to be met.

[0117] For example, when generating each low-altitude reachability layer, the reachable state points in the previous prediction time layer can be used as the starting point for expansion. Multiple candidate control variables are applied within a preset time step, including candidate acceleration, candidate heading angular velocity, and candidate vertical velocity. For each set of candidate control variables, it is first determined whether they satisfy the constraints of the target aircraft's maximum speed, maximum acceleration, maximum turning angular velocity, maximum rate of climb, and maximum rate of descent. Then, candidate state points for the next prediction time layer are obtained based on these candidate control variables. Subsequently, the position, altitude, or heading of the candidate state point is corrected based on the low-altitude wind field at its location, and it is determined whether the corrected candidate state point falls into an obstacle area, a no-fly zone, or an impassable area. If it falls into such an area, the candidate state point is removed; otherwise, it is added to the reachability set of the next prediction time layer.

[0118] If the corrected candidate state point is located in the observation attenuation region, the candidate state point can be retained and an observation attenuation marker can be added. The observation attenuation marker is used to characterize that the candidate state point is in a radar blind zone, an electro-optical line-of-sight obstruction zone, or a weak communication coverage zone, so that subsequent candidate maneuver behavior chain generation and deep behavior chain prediction models can identify whether the target aircraft may enter or leave the observation attenuation region. Figure 6The diagram shown in this embodiment illustrates the reachability state offset correction of the target aircraft affected by the low-altitude wind field. It can be seen that the original theoretical reachability domain boundary underwent a significant physical offset after the wind field vector was superimposed. This ensures that the generated layered reachability domain can truly reflect the actual motion boundary under meteorological conditions.

[0119] It should be understood that the purpose of using a layered low-altitude reachability domain in this embodiment is to determine the actual three-dimensional spatial range that the target aircraft can reach at different predicted times in the future before generating candidate behaviors and predicted trajectories. Therefore, unlike judging reachability solely based on horizontal distance, this embodiment considers the target aircraft's own motion capability, wind field correction, obstacles, no-fly zones, and altitude change capability simultaneously through a layered low-altitude reachability domain, which can provide a more realistic constraint basis for subsequent reachability verification screening.

[0120] After obtaining the hierarchical low-altitude reachability domain, the hierarchical low-altitude reachability domain is further used to perform reachability verification and screening of low-altitude scene constraint objects, so as to determine whether a certain low-altitude scene constraint object can be used as a valid basis for behavior in the current prediction period before the candidate maneuver behavior chain is generated.

[0121] In specific calculations, for the first... For each low-altitude scene constraint object, the reachability correlation value between the object and each low-altitude reachable domain is calculated, and the reachability counter-proof value is determined based on the reachability correlation value corresponding to each low-altitude reachable domain; the calculation process satisfies the following expression:

[0122] ;

[0123] In the formula, Indicates the first The reachability counter-proof value of a low-altitude scene constraint object; Indicates the first The first low-altitude scene constraint object and the first The reachability correlation values ​​between the lower-level reachable regions, and The value of can be normalized to between 0 and 1; This indicates the number of layers in the low-altitude reachable domain.

[0124] It should be noted that the above-mentioned reachable correlation values The determination can be based on the geometric type of the constraint object in the low-altitude scene. For example:

[0125] When the constraint object in a low-altitude scene is a point object, the reachability association value can be determined based on the minimum distance from the point object to the set of state points in the corresponding low-altitude reachable domain.

[0126] When the constraint object in a low-altitude scene is a linear object, the reachability association value can be determined based on the overlap ratio or minimum distance between the linear object and the corresponding low-altitude reachability domain projection area.

[0127] When the constraint object in a low-altitude scene is a planar object or a volumetric object, the reachability association value can be determined based on the overlap area, overlap volume, or shortest boundary distance between the planar object or volumetric object and the corresponding low-altitude reachable domain.

[0128] Furthermore, it should be noted that if the connection path from the current state to a certain low-altitude scene constraint object requires traversing a low-altitude scene constraint object with avoidance attributes, an obstacle area, or a no-fly zone, then the reachability association value corresponding to that low-altitude scene constraint object should be reduced or set to zero.

[0129] Furthermore, reachable state points in each prediction time layer can be used as state nodes, and state transition edges can be established between state nodes in adjacent prediction time layers that satisfy motion capability constraints. If the spatial line segment corresponding to the state transition edge intersects with an obstacle area, a no-fly zone, or an impassable area, then the state transition edge is deleted. This forms a hierarchical reachable state graph, and based on this hierarchical reachable state graph, it is determined whether there is a state transition path between the current state and the neighborhood of the low-altitude scene constraint object. If there is no state transition path, then the reachability association value of the low-altitude scene constraint object is reduced or set to zero.

[0130] Regarding the aforementioned reachability counter-proof value Understandably, if a constraint object in a low-altitude scene is difficult to effectively correlate with the hierarchical low-altitude reachability domain across all prediction time layers, then... Smaller A larger value indicates that the object is unreachable within the current prediction window or should not be considered a valid basis for action; if the object has a strong correlation with the hierarchical low-altitude reachability domain within at least one prediction time layer, then... A smaller value indicates that the object can continue to participate in the generation of candidate maneuver behavior chains.

[0131] Furthermore, when the reachability disproven value is greater than the preset disproven threshold, the competitive amount of the corresponding low-altitude scene constraint object can be reduced, or it can be directly excluded from the effective low-altitude scene constraint objects. Through this process, objects that are inaccessible, disconnected, or do not conform to low-altitude environment constraints within the current prediction time window can be excluded before the candidate maneuvering behavior chain is generated. This not only reduces the burden on the subsequent deep behavior chain prediction model to score unreasonable candidate behaviors, but also reduces the risk of the predicted trajectory crossing obstacles, no-fly zones, or short-term inaccessible areas.

[0132] S400: Based on the historical trajectory sequence and the corresponding trajectory confidence markers, an effective observation window is formed, and combined with the effective low-altitude scene constraint objects, a candidate maneuver behavior chain that the target aircraft may experience in the future prediction time window is generated.

[0133] Specifically, in one implementation, trajectory points within a recent period are selected from the historical trajectory sequence, and a valid observation window is formed by combining the corresponding trajectory confidence markers. During this process, the system can divide and process the trajectory points based on the trajectory confidence markers generated in the aforementioned steps.

[0134] For example, when the trajectory confidence marker corresponding to a trajectory point is less than a preset first confidence threshold, it is identified as a trajectory point with low confidence, and its weight can be reduced in the effective observation window; when the trajectory confidence marker is further less than a preset second confidence threshold (the second confidence threshold is lower than the first confidence threshold), or when the time continuity evaluation item or speed rationality evaluation item corresponding to the trajectory point is lower than a preset safety baseline, it is identified as an abnormal trajectory point that clearly does not meet the motion continuity constraint, and can be weakened or removed from the effective observation window; thus, the effective observation window can more accurately reflect the recent real motion trend of the target aircraft, while avoiding excessive interference from low-confidence observations on the generation of subsequent candidate behavior chains; such as Figure 7 The distribution map of low-confidence abnormal trajectory points within the effective observation window shows that when the confidence score of a trajectory point drops sharply and falls below the threshold, its weight in the model input window is drastically reduced, physically isolating it from the interference of dirty data at the source.

[0135] The low-altitude scene constraint objects retained after accessibility verification are used as effective low-altitude scene constraint objects. Based on the effective observation window, effective low-altitude scene constraint objects, corresponding action attributes, action competition quantities, and corresponding hierarchical low-altitude reachability domains, candidate maneuver behavior chains that the target aircraft may experience in the future prediction time window are generated.

[0136] In one implementation, when generating candidate maneuver behavior chains, the effective low-altitude scene constraint objects are first sorted according to the action competition quantity, reachability counter-evidence value, and historical motion matching degree, and the top-ranked objects are selected as dominant candidate objects. Then, preset behavior templates are matched according to the action attributes of the dominant candidate objects, where the passage attribute objects correspond to the line-of-sight flight template, the avoidance attribute objects correspond to the bypass template, the approach attribute objects correspond to the approach, return, or descent template, and the observation attenuation attribute objects correspond to the entry into the observation attenuation region and reappearance template. Subsequently, the start and end time layers of the maneuver behavior stage are determined according to the predicted time layer of the first arrival in the neighborhood of the corresponding object in the hierarchical low-altitude reachability domain, and multiple maneuver behavior stages that are adjacent in time sequence and satisfy the connectivity path constraint are connected to form a candidate maneuver behavior chain; if there is no connectivity path between adjacent stages that satisfies the motion capability and low-altitude environment constraints, the corresponding candidate maneuver behavior chain is deleted.

[0137] Furthermore, candidate maneuver behavior chains are used to describe the continuous behavior phases that a target aircraft may undergo within a future prediction time window. Each candidate maneuver behavior chain includes multiple maneuver behavior phases connected in chronological order. Each maneuver behavior phase includes a maneuver behavior state, a dominant low-altitude scenario constraint object, and a corresponding phase time range.

[0138] Maneuver behavior states include at least one of the following: straight-line crossing, flying along low-altitude scene constraints with passage attributes, avoiding low-altitude scene constraints with avoidance attributes, hovering or low-speed observation, approaching sensitive areas, returning home, circling, descending or landing, entering the observation attenuation region, and reappearing from the observation attenuation region.

[0139] For example, when the target aircraft's current heading is consistent with the extension direction of a road, river, power line, pipeline, or low-altitude passage, and the accessibility attribute object still belongs to the valid low-altitude scenario constraint object after reachability verification, a maneuvering behavior stage of "flying along the accessibility attribute object" can be generated; if there are also reachable sensitive areas, inspection points, or suspected take-off and landing points near the extension direction of the accessibility attribute object, a candidate maneuvering behavior chain of "flying along the accessibility attribute object - approaching the constraint object" can be further generated.

[0140] When there are buildings, mountains, power towers, overpasses, no-fly zones, or electronic fence boundaries in front of the target aircraft, and there is accessible space on its left and right sides or up and down, a candidate maneuvering behavior chain of "straight flight - avoidance attribute object" can be generated.

[0141] When the target aircraft is at low speed and located in a sensitive area or near an inspection point, a candidate maneuvering behavior phase of "hovering or low-speed observation" can be generated. If there is an reachable historical return point, UAV nest, or control point after this phase, a candidate maneuvering behavior chain of "hovering or low-speed observation - return" can be generated.

[0142] When the target aircraft's current direction of motion is pointing towards the radar blind zone, the electro-optical line-of-sight obstruction zone, or the weak communication coverage zone, a candidate maneuvering behavior chain of "entering the observation attenuation zone - reappearing from the observation attenuation zone" can be generated.

[0143] In addition, the maximum number of stages, maximum duration, maximum number of turns, or maximum number of candidates for candidate maneuvering behavior chains can be set to control the computational scale of the deep behavior chain prediction model while covering the main possible trajectory types.

[0144] Therefore, through the candidate maneuver behavior chain generation step, this step transforms the prediction object from a single future coordinate into a continuous maneuver behavior structure with behavioral meaning, enabling the subsequent deep learning model to predict at two levels: "what behaviors may occur" and "when to reach the corresponding behavior stage", rather than just fitting future discrete location points.

[0145] The S500 inputs the effective observation window, trajectory confidence marker, effective low-altitude scene constraint object, hierarchical low-altitude reachability domain, and candidate maneuvering behavior chains into the deep behavior chain prediction model, and outputs the occurrence probability and arrival time distribution of each candidate maneuvering behavior chain.

[0146] It should be noted that the deep behavior chain prediction model in this embodiment does not directly predict future trajectory coordinates, but is used to make probability judgments and predict arrival time distributions for candidate maneuver behavior chains after they have been screened by accessibility proof; and the deep behavior chain prediction model includes a time sequence encoder, a behavior chain encoder and a fusion scorer.

[0147] The timing encoder is used to extract historical motion codes containing trajectory confidence markers. Its input includes the position, altitude, velocity, heading, vertical velocity, and trajectory confidence markers of each trajectory point within the effective observation window. The timing encoder can be an LSTM, GRU, TCN, or Transformer encoder. Because the input includes trajectory confidence markers, the timing encoder can distinguish between high-confidence and low-confidence observations when extracting historical motion trends, thus avoiding learning abnormal trajectory changes caused by occlusion, weak echoes, inconsistencies in multi-source positioning, or short-term signal interruptions as real maneuvering behavior.

[0148] The behavior chain encoder is used to extract the behavior chain code of candidate maneuver behavior chains. Since each maneuver behavior stage in the candidate maneuver behavior chain can be expressed by the behavior state type, the type of dominant low-altitude scene constraint object, the action attribute, the stage time range, the action competition quantity, and the corresponding reachable time layer, the behavior chain encoder can use sequence coding network, graph neural network, attention coding network, or embedding coding network to obtain the behavior chain code of the candidate maneuver behavior chain.

[0149] The fusion scorer integrates historical motion encoding, behavior chain encoding, and hierarchical low-altitude reachability features of corresponding candidate maneuver behavior chains, outputting the occurrence probability of each candidate maneuver behavior chain, the estimated arrival time of the corresponding dominant low-altitude scene constraint object, and the arrival time uncertainty. For example... Figure 8 As shown, the deep behavior chain prediction model can output the predicted arrival time distribution of different maneuvering behaviors, thus providing a time uncertainty reference for low-altitude early warning and control.

[0150] When performing feature fusion, the reachability distribution of candidate maneuvering behavior chains in different prediction time layers, their spatial relationship with the dominant low-altitude scene constraint objects, reachability domain boundary constraints, reachability domain connectivity, and observation attenuation markers can be extracted as hierarchical low-altitude reachability domain features. The effective observation window features, behavior chain features, low-altitude scene constraint object features, and the hierarchical low-altitude reachability domain features are aligned according to the dominant low-altitude scene constraint objects and prediction time layers corresponding to the candidate maneuvering behavior chains, and then fused features are formed through feature splicing, attention fusion, or gating fusion.

[0151] It should also be noted that during the model training phase, the training samples include historical observation windows, trajectory confidence markers, low-altitude scene constraint objects, hierarchical low-altitude reachability domains, candidate maneuvering behavior chains, and real subsequent trajectories. The real subsequent trajectories are used to determine the candidate maneuvering behavior chain labels and arrival time labels.

[0152] During the construction of training samples, the actual subsequent trajectory can be matched with each candidate maneuvering behavior chain to determine the corresponding candidate maneuvering behavior chain label. When the actual subsequent trajectory moves continuously along the extension direction of traffic attribute objects such as roads, rivers, power transmission lines, pipelines, or low-altitude channels, a flight along the route label is determined; when the actual subsequent trajectory avoids buildings, no-fly zones, or other avoidance attribute objects and remains outside the safe distance, an avoidance label is determined; when the actual subsequent trajectory enters the neighborhood range of approach attribute objects such as sensitive areas, inspection points, or suspected take-off and landing points, an approach label is determined; when the actual subsequent trajectory moves towards historical return points, UAV nests, or control points and enters their neighborhood range, a return label is determined; when the actual subsequent trajectory enters the observation decay region and reappears at a later time, an observation decay reappearance label is determined.

[0153] The arrival time label can be determined based on the time when the actual subsequent trajectory first enters the neighborhood of the corresponding dominant low-altitude scene constraint object; if the same actual subsequent trajectory satisfies multiple behavior labels at the same time, the main label can be determined based on the behavior duration, spatial matching degree and the competitive amount of the corresponding low-altitude scene constraint object, or it can be used as a multi-label training sample.

[0154] The training objectives of the deep behavior chain prediction model include behavior chain classification, arrival time regression, and reachability domain consistency constraint. Specifically, the behavior chain classification objective improves the accuracy of predicting the probability of candidate maneuvering behavior chains; the arrival time regression objective improves the accuracy of predicting the arrival time of dominant low-altitude scene constraint objects; and the reachability domain consistency constraint objective reduces the likelihood of the model outputting candidate behavior chains with high probability but not conforming to hierarchical low-altitude reachability domain constraints.

[0155] Preferably, the total loss function of the deep behavioral chain prediction model can be expressed as:

[0156] ;

[0157] in, Cross-entropy loss is used for behavior chain classification to optimize the prediction of occurrence probability; This is the mean squared error or smoothed L1 loss used for regression of the expected arrival time; This refers to the penalty loss generated by the reachability domain consistency constraint objective. When the high-probability behavior chain predicted by the model traverses the obstacle avoidance boundary in the hierarchical low-altitude reachability domain, this penalty loss is activated and significantly increased. The penalty coefficient is used; through the training of the above joint loss function, the deep network is forced to learn and obey the physical constraints in three-dimensional space while fitting the observed data.

[0158] Therefore, through the above model processing, the deep behavior chain prediction model can predict in the effective candidate space after low-altitude scene constraints and hierarchical low-altitude reachability filtering. This ensures that the output not only includes both the probability of behavior occurrence and temporal distribution information, but also eliminates the generation of anti-physics prediction results, thus providing a solid and accurate data basis for the generation of the main prediction trajectory and backup prediction trajectories in subsequent steps. Figure 9 As shown, this is a comparison chart illustrating the improvement in training performance of the depth prediction model by incorporating low-altitude reachability physical constraints in this embodiment; the convergence curve of the loss function clearly shows that after introducing physical consistency constraints, the model... The values ​​converge faster and oscillate less, proving that the introduction of physical rules overcomes the shortcomings of blind search in traditional pure data-driven models.

[0159] S600 generates the primary and backup predicted trajectories of the target aircraft based on the probability of occurrence, arrival time distribution, and layered low-altitude reachability.

[0160] Understandably, after obtaining the occurrence probability and arrival time distribution of each candidate maneuver chain, this scheme can generate corresponding candidate trajectories based on each candidate maneuver chain. When generating candidate trajectories, within the stage time range corresponding to each maneuver stage of the candidate maneuver chain, reachable state points that simultaneously satisfy behavioral state constraints, the spatial relationship of the dominant low-altitude scenario constraint object, and the target aircraft's motion capability constraints can be selected from the hierarchical low-altitude reachable domain as trajectory control points. Then, continuous trajectory segments are generated based on the state transition paths between adjacent trajectory control points, or spline interpolation, Bézier curves, or the shortest reachable path are used for connection. If the connected trajectory segment does not satisfy the constraints of speed, acceleration, turning angular velocity, rate of climb, or rate of descent, or crosses obstacle areas, no-fly zones, or impassable areas, the comprehensive score of the corresponding candidate trajectory is reduced or the candidate trajectory is deleted.

[0161] Specifically, for the flight phase along the accessibility attribute object, a trajectory can be generated along the extension direction of roads, rivers, power transmission lines, pipelines, or low-altitude channels, and the trajectory space extension is restricted to the layered low-altitude reachable domain of the corresponding time layer.

[0162] For the avoidance of attribute objects, an avoidance trajectory can be generated along the reachable area outside the boundary of the avoidance object, and the maximum turning angular velocity, maximum climb rate, maximum descent rate and no-fly boundary constraints can be met.

[0163] For the approach phase, trajectories can be generated toward sensitive areas, inspection points, suspected take-off and landing points, historical return points, or drone nests, and the trajectories are constrained to be within the reachable domain of the corresponding time layer.

[0164] For hovering or low-speed observation phases, low-speed trajectories can be generated within a finite radius near the current location or the target area. For entering the observation decay region and reappearing phases, candidate trajectories can be generated by combining the boundary of the observation decay region and the connectivity of the reachable domain.

[0165] After generating candidate trajectories, the comprehensive score of each candidate trajectory can be calculated by weighting the probability of occurrence of each candidate maneuver behavior chain, the consistency evaluation of reachability domain, the consistency evaluation of scenario constraints, and the trajectory credibility support evaluation based on trajectory credibility labeling.

[0166] For example, the first The combined score of the candidate trajectories can be expressed as:

[0167] ;

[0168] in, Indicates the first The overall score of the candidate trajectories; Indicates the first The probability of occurrence of a candidate maneuvering behavior chain; The representation indicates the reachability consistency evaluation, which is used to characterize whether the candidate trajectory is located within the hierarchical low-altitude reachability domain of the corresponding time layer, and whether it meets the reachability connectivity requirements; The scene constraint consistency evaluation is used to characterize whether the candidate trajectory matches the dominant low-altitude scene constraint object in the candidate maneuvering behavior chain. For example, whether the avoidance trajectory remains outside the avoidance object, whether the flight trajectory along the line is consistent with the direction of the passing object, and whether the approach trajectory is towards the approaching attribute object. The trajectory credibility support evaluation is used to characterize whether historical high-confidence trajectory points support the motion trend corresponding to the candidate trajectory. to The corresponding weights can be set according to task preferences and safety baselines; for example, in scenarios with extremely high security or obstacle avoidance requirements, it can significantly improve the consistency evaluation of the reachability domain. Consistency evaluation of scenario constraints The weighting of the model is adjusted to suppress unreasonable high-probability predictions generated by the model.

[0169] Understandably, by calculating the comprehensive score of each candidate trajectory, the risk of generating unreachable trajectories or trajectories with unreasonable scenarios by relying solely on model probabilities can be reduced; the main predicted trajectory is not determined solely by the highest probability of occurrence output by the deep behavioral chain prediction model, but also by considering whether the candidate trajectory conforms to the hierarchical low-altitude reachability domain, whether it conforms to the meaning of low-altitude scenario constraints, and whether it is supported by historical high-reliability trajectories.

[0170] Therefore, the candidate trajectory with the highest comprehensive score is selected as the main predicted trajectory, and several candidate trajectories with high scores and different behavior chain types are selected as backup predicted trajectories. Among them, backup predicted trajectories include avoidance backup trajectories, return-to-base backup trajectories, flight-along-line backup trajectories, and reappearance-after-entry backup trajectories. The backup predicted trajectories are used to express the uncertainty of the target aircraft's future behavior, thereby avoiding the system outputting only a single predicted trajectory, which would lead to subsequent early warning, interception, control, or scheduling decisions relying excessively on a single result.

[0171] Furthermore, as a preferred embodiment, after generating the main predicted trajectory and backup predicted trajectory of the target aircraft, a discontinuous observation posterior update step is also included.

[0172] In other words, when the target aircraft is briefly lost, the trajectory prediction is not terminated immediately, but the current main predicted trajectory and backup predicted trajectory are maintained, and the corresponding candidate maneuver behavior chain is continued to be expanded.

[0173] Therefore, when the target aircraft reappears and acquires a new observation state, the matching degree is determined based on the positional and heading deviations between the new observation state and the predicted states of each candidate maneuvering behavior chain at the current moment. The higher the matching degree, the more the new observation state matches the prediction results of the corresponding candidate maneuvering behavior chain.

[0174] The probability of occurrence of each candidate maneuver behavior chain before the target is lost is updated posteriorly based on the matching degree. The updated posterior value is then used to determine the probability of occurrence of each maneuver behavior chain before the target is lost. The probability of a candidate maneuver behavior chain occurring can be expressed as:

[0175] ;

[0176] in, Indicates the posterior updated th The probability of a candidate maneuver behavior chain occurring; Indicates the number of times the target was lost. The probability of occurrence of a candidate maneuvering behavior chain; Indicates the re-observation state and the first The degree of matching between the predicted states of the candidate maneuver behavior chains at the current moment; Indicates the number of candidate maneuver behavior chains; This represents the smallest positive number that prevents the denominator from being zero.

[0177] Furthermore, the matching degree can be determined based on the positional deviation between the re-observed position and the predicted position, and the heading deviation between the re-observed heading and the predicted heading. In one specific implementation, considering the dimensional difference between position (distance) and heading (angle), the two can be dimensionlessly fused using a double exponential decay function, the calculation process of which satisfies:

[0178] ;

[0179] in, Indicates the re-observation state and the first The degree of matching between the predicted states of the candidate maneuver behavior chains at the current moment; Indicates the true location of the re-observed state; Indicates the first The predicted position of each candidate maneuvering behavior chain at the corresponding time; Indicates positional deviation (Euclidean distance); The preset position deviation tolerance normalization coefficient; Indicates the true heading at the re-observed state; Indicates the corresponding predicted heading; Indicates the absolute deviation of the heading; This is the preset normalization coefficient for heading deviation tolerance.

[0180] This formula smoothly maps physical distance and angular deviations to a matching probability between 0 and 1; the smaller the deviation, the higher the probability of a match. The closer to 1, the higher the matching degree of the candidate maneuver behavior chain for avoidance. If the target aircraft reappears on the other side of a certain avoidance attribute object, the matching degree of the candidate maneuver behavior chain for avoidance increases, and the probability of its occurrence increases accordingly. If the target aircraft reappears at the historical return point or in the direction of the UAV nest, the matching degree of the candidate maneuver behavior chain for return increases. If the target aircraft reappears at the outer edge of the observation attenuation area, the matching degree of the candidate maneuver behavior chain that reappears after entering the observation attenuation area increases.

[0181] Furthermore, through the discontinuous observation post-hoc update step, this scheme can use the re-observation state to correct the probability of occurrence of the original candidate maneuvering behavior chain when the target aircraft reappears after a short period of loss, and correct the main predicted trajectory and backup predicted trajectory accordingly. This can improve the continuity and stability of the predicted trajectory under discontinuous observation conditions, and avoid directly interrupting the prediction or starting the prediction from scratch due to the short-term loss of the target.

[0182] In summary, the prediction method in this embodiment does not directly input historical trajectory sequences into a deep learning model to regress future coordinates. Instead, it constructs a prediction framework that integrates physical rule constraints and data-driven approaches. Specifically, it first labels the multi-source observation data with credibility, then transforms low-altitude scene objects into constraint objects with functional attributes. Subsequently, it combines the target aircraft's maneuverability with the low-altitude environment to generate a hierarchical low-altitude reachability domain, and uses this domain to perform reachability verification and filtering of scene constraint objects. Based on this, candidate maneuver behavior chains are generated, and finally, a deep behavior chain prediction model performs probability judgment and arrival time prediction, thereby generating the main predicted trajectory and backup predicted trajectories.

[0183] Thus, through the above systematic processing, the prediction process of the deep learning model in this scheme no longer occurs in an unconstrained future coordinate space, but is strictly limited to an effective candidate space after observation confidence correction, low-altitude physical constraint screening, and explicit construction of behavioral structure. This eliminates the generation of anti-physics prediction results and significantly improves the accuracy, robustness, and interpretability of aircraft trajectory prediction results. Figure 10 It can be intuitively seen that during the short-term loss of the target, the probabilities of each candidate chain naturally drift due to the increased uncertainty. When the target is observed again at a certain time point, the model immediately calculates the matching degree according to the above formula and triggers the posterior update, so that the probability of the candidate chain (such as the avoidance chain) that matches the actual observation is instantly and significantly corrected. This improves the stability and continuity of the predicted trajectory under intermittent tracking conditions and avoids the defects of prediction interruption or restart from zero.

[0184] Example 2:

[0185] This embodiment discloses an intelligent aircraft trajectory prediction system based on deep learning big data. It is understood that this system can be deployed on low-altitude air traffic control platforms, drone monitoring platforms, key area security platforms, airport perimeter monitoring platforms, edge computing nodes, cloud servers, or distributed computing systems composed of multiple edge devices and cloud servers; and is used to execute the intelligent aircraft trajectory prediction method based on deep learning big data described in Embodiment 1 above.

[0186] Specifically, please refer to Figure 11 Its system consists of a multi-source observation fusion and labeling module, a constraint object extraction and attribute determination module, an reachability domain generation and counter-evidence module, a candidate behavior chain generation module, a depth prediction module, and a primary and backup trajectory generation module.

[0187] The multi-source observation fusion and labeling module is used to acquire multi-source observation data and low-altitude scene data of the target aircraft, perform spatiotemporal alignment and target association on the multi-source observation data, form a historical trajectory sequence, and generate trajectory credibility labels for trajectory points in the historical trajectory sequence. This module can receive one or more of the following: radar detection data, electro-optical detection data, infrared detection data, radio spectrum detection data, acoustic detection data, Remote ID data, telemetry data, and historical trajectory record data.

[0188] This module can also perform coordinate transformation, unified time base resampling, target association, source confidence fusion, and trajectory confidence label calculation.

[0189] In one implementation, the multi-source observation fusion and labeling module is used to perform the following processing:

[0190] The module transforms observation points from different data sources to a unified spatial coordinate system; establishes continuous sampling times according to a unified time reference; when multiple source observation points exist at the same time, they are fused based on source confidence, location consistency, and motion continuity; and calculates trajectory confidence labels based on source confidence evaluation items, time continuity evaluation items, velocity reasonableness evaluation items, altitude change reasonableness evaluation items, heading change reasonableness evaluation items, multi-source observation consistency evaluation items, and observation attenuation state evaluation items. The historical trajectory sequence and trajectory confidence labels output by this module serve as the basic inputs for subsequent candidate behavior chain generation and depth prediction.

[0191] The constraint object extraction and attribute determination module is used to extract low-altitude scene constraint objects from low-altitude scene data and determine the effect attributes based on how the low-altitude scene constraint objects affect the future motion of the target aircraft.

[0192] It can extract point objects, line objects, area objects, and volume objects from 3D maps, low-altitude control data, no-fly zone databases, electronic fence databases, geographic information systems, communication coverage data, and monitoring equipment visibility data, and identify them as approach attributes, avoidance attributes, passage attributes, or observation attenuation attributes. It can also calculate the competitive forces of constraint objects in low-altitude scenarios based on distance correlation terms, heading correlation terms, speed trend correlation terms, altitude correlation terms, historical matching terms, and observation confidence correlation terms.

[0193] The reachability generation and disproving module is used to generate a hierarchical low-altitude reachability domain based on the target aircraft’s current state, maneuverability and low-altitude environmental constraints, and to use the hierarchical low-altitude reachability domain to perform reachability disproving and filtering of low-altitude scene constraint objects.

[0194] It can divide the future prediction time window into multiple prediction steps, expand the next layer of possible states based on the set of reachable states in the previous layer, and apply constraints such as maximum speed, maximum acceleration, maximum turning rate, maximum descent rate, and maximum climb rate during the expansion process. This module can also correct possible states based on low-altitude wind fields and eliminate possible states that fall into obstacle areas, no-fly zones, or impassable areas.

[0195] During the accessibility disproving and filtering process, the accessibility domain generation and disproving module can calculate the accessibility association value between the low-altitude scene constraint object and the low-altitude accessibility domain of each layer, and determine the accessibility disproving value based on the accessibility association value.

[0196] For point objects, line objects, area objects, or volume objects, this module can determine the reachability correlation value by means of minimum distance, projection overlap ratio, overlap area, overlap volume, or shortest boundary distance, respectively. When the reachability disproving value is greater than the preset disproving threshold, this module reduces the competitive amount of the corresponding low-altitude scene constraint object, or excludes it from the effective low-altitude scene constraint objects.

[0197] The candidate behavior chain generation module is used to form an effective observation window based on historical trajectory sequences and corresponding trajectory confidence markers, and, in conjunction with effective low-altitude scene constraints, generate candidate maneuver behavior chains that the target aircraft may experience within the future prediction time window. It generates candidate maneuver behavior stages based on the effect attributes of the effective low-altitude scene constraints, the corrected effect competition quantity, the reachable time layer, spatial positional relationships, and historical motion trends, and connects multiple maneuver behavior stages in chronological order to form a candidate maneuver behavior chain.

[0198] The depth prediction module is used to input the effective observation window, trajectory confidence marker, effective low-altitude scene constraint object, hierarchical low-altitude reachability domain, and candidate maneuver behavior chains into the deep behavior chain prediction model, and output the occurrence probability and arrival time distribution of each candidate maneuver behavior chain.

[0199] The deep behavior chain prediction model includes a temporal encoder, a behavior chain encoder, and a fusion scorer. The temporal encoder is used to extract historical motion codes containing trajectory confidence markers; the behavior chain encoder is used to extract behavior chain codes for candidate maneuver behavior chains; and the fusion scorer is used to fuse historical motion codes, behavior chain codes, and hierarchical low-altitude reachability features of corresponding candidate maneuver behavior chains, and outputs the probability of occurrence, estimated arrival time, and arrival time uncertainty.

[0200] The primary and backup trajectory generation module generates the primary and backup predicted trajectories of the target aircraft based on the occurrence probability, arrival time distribution, and layered low-altitude reachability. This module can restrict the spatial extension of each candidate maneuvering behavior chain to the layered low-altitude reachability of the corresponding time layer, generate the corresponding candidate trajectories, and calculate the comprehensive score of the candidate trajectories based on the occurrence probability of the candidate maneuvering behavior chain, reachability consistency evaluation, scenario constraint consistency evaluation, and trajectory credibility support evaluation. The candidate trajectory with the highest comprehensive score is selected as the primary predicted trajectory, while several candidate trajectories with high scores and different behavior chain types are selected as backup predicted trajectories.

[0201] Furthermore, the primary and backup trajectory generation module can also perform intermittent observation posterior updates. When the target aircraft is briefly lost, the module maintains the current primary and backup predicted trajectories and continues to expand the corresponding candidate maneuvering behavior chains. When the target aircraft reappears, the module determines the matching degree based on the position and heading deviations between the re-observed state and the predicted state of each candidate maneuvering behavior chain at the current moment, and uses the matching degree to perform posterior updates on the occurrence probability of each candidate maneuvering behavior chain before the target was lost, so as to correct the primary and backup predicted trajectories.

[0202] Each module in this embodiment can be implemented by software programs, hardware circuits, or a combination of software and hardware. Data can be transferred between modules through memory, database, message queue, network interface, or data bus. It implements the method of Embodiment 1 in a modular way, so that aircraft trajectory prediction in low-altitude complex scenarios can be completed under the joint constraints of multi-source observation credibility, low-altitude scenario constraints, hierarchical low-altitude reachability domain, and candidate maneuver behavior chain.

[0203] Finally, it should be noted that the preset time parameters and preset thresholds involved in the embodiments of the present invention are not fixed absolute constants. Those skilled in the art can make adaptive configurations based on the actual flight characteristics of the target aircraft, the control requirements of low-altitude airspace, and the accuracy of the monitoring equipment.

[0204] For example, the future prediction time window and prediction step duration can be dynamically matched based on the average flight speed of the target aircraft and the spatial scale of the low-altitude control grid; the first confidence threshold and the second confidence threshold can be scientifically calibrated based on the historical signal-to-noise ratio and background noise statistical distribution of multi-source observation equipment under the current operating conditions; the "preset counter-evidence threshold" in the reachability counter-evidence screening can be flexibly set based on the physical collision tolerance and safety buffer margin of the specific task. This scheme does not limit the specific absolute values ​​of the above parameters. Any conventional parameter adaptive scaling and threshold matching based on the core logic of this invention are conventional technical means for those skilled in the art and do not depart from the protection scope of this invention.

[0205] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent aircraft trajectory prediction based on deep learning big data, characterized in that, Includes the following steps: Acquire multi-source observation data and low-altitude scene data of the target aircraft, perform spatiotemporal alignment and target association on the multi-source observation data to form a historical trajectory sequence, and generate trajectory confidence markers for trajectory points in the historical trajectory sequence; Extract low-altitude scene constraint objects from the low-altitude scene data, and determine the effect attributes based on how the low-altitude scene constraint objects affect the future motion of the target aircraft. Based on the current state, maneuverability, and low-altitude environmental constraints of the target aircraft, a hierarchical low-altitude reachability domain is generated. The reachability of the low-altitude scenario constraint objects is then filtered by the hierarchical low-altitude reachability domain to obtain effective low-altitude scenario constraint objects for generating candidate maneuver behavior chains. An effective observation window is formed based on the historical trajectory sequence and the corresponding trajectory confidence marker, and combined with the effective low-altitude scene constraint object, a candidate maneuver behavior chain that the target aircraft may experience in the future prediction time window is generated. The effective observation window, the trajectory confidence marker, the effective low-altitude scene constraint object, the hierarchical low-altitude reachability domain, and the candidate maneuvering behavior chain are input into the deep behavior chain prediction model, and the occurrence probability and arrival time distribution of each candidate maneuvering behavior chain are output. Based on the occurrence probability, the arrival time distribution, and the layered low-altitude reachability, the primary predicted trajectory and backup predicted trajectory of the target aircraft are generated.

2. The intelligent aircraft trajectory prediction method based on deep learning big data according to claim 1, characterized in that, Generate trajectory confidence markers for trajectory points in the historical trajectory sequence, including: The source confidence evaluation item, time continuity evaluation item, speed rationality evaluation item, altitude change rationality evaluation item, heading change rationality evaluation item, multi-source observation consistency evaluation item, and observation attenuation status evaluation item for the current trajectory point are calculated respectively. The above evaluation items are then weighted, summed, and normalized to obtain the trajectory confidence label for the trajectory point. The speed rationality evaluation item is determined based on the deviation between the speed calculated from adjacent locations and the fused observation speed; the multi-source observation consistency evaluation item is determined based on the maximum or average deviation between the positioning locations of multiple data sources at the same time; and the observation attenuation status evaluation item is determined based on whether the target aircraft is in a radar blind zone, an electro-optical line-of-sight obstruction zone, or a weak communication coverage zone.

3. The intelligent aircraft trajectory prediction method based on deep learning big data according to claim 2, characterized in that, The method further includes calculating the action competition quantity for each low-altitude scenario constraint object, wherein the action competition quantity is used to characterize the degree of influence of the corresponding low-altitude scenario constraint object on the future motion of the target aircraft within the current prediction period; The competitive factor is calculated by weighting the distance correlation term, heading correlation term, speed trend correlation term, altitude correlation term, historical matching term, and observation confidence correlation term. The observation confidence correlation term is determined based on the correlation between the trajectory confidence marker and the trajectory state and the low-altitude scene constraint object, and is used to reduce the impact of low-confidence trajectory points on the calculation of the action competition quantity.

4. The intelligent aircraft trajectory prediction method based on deep learning big data according to claim 2, characterized in that, Generating the hierarchical low-altitude reachability domain includes: The future prediction time window is divided into multiple prediction steps; based on the set of reachable states in the previous layer, possible states are expanded under the control constraints of the target aircraft's maximum speed, maximum acceleration, maximum turning rate, maximum descent rate, and maximum climb rate. The possible states obtained from the expansion are corrected based on the low-altitude wind field; and possible states that fall into obstacle areas, no-fly zones or impassable areas are eliminated, and layered low-altitude reachable domains are generated layer by layer to represent the actual three-dimensional spatial range that the target aircraft can reach at different predicted times.

5. The intelligent aircraft trajectory prediction method based on deep learning big data according to claim 4, characterized in that, The accessibility verification of the low-altitude scene constraint objects is performed using the hierarchical low-altitude reachability domain, including: Calculate the reachability correlation value between the low-altitude scene constraint object and each layer of low-altitude reachable domain, and determine the reachability counter-proof value based on the reachability correlation value corresponding to each layer of low-altitude reachable domain; When the low-altitude scene constraint object is a point object, the reachability association value is determined based on the minimum distance from the point object to the set of state points in the corresponding low-altitude reachable domain. When the low-altitude scene constraint object is a linear object, the reachability association value is determined based on the overlap ratio or minimum distance between the linear object and the corresponding layer's low-altitude reachability domain projection area. When the low-altitude scene constraint object is a planar object or a volumetric object, the reachability association value is determined based on the overlap area, overlap volume, or shortest boundary distance between the planar object or volumetric object and the corresponding low-altitude reachable domain. If the connecting path from the current state to the low-altitude scene constraint object crosses the low-altitude scene constraint object with avoidance attribute, obstacle area, or no-fly boundary, then the reachable association value is reduced or set to zero. When the reachability disproven value is greater than the preset disproven threshold, it is excluded from the valid low-altitude scene constraint objects.

6. The intelligent aircraft trajectory prediction method based on deep learning big data according to claim 5, characterized in that, Generating the candidate maneuver behavior chain includes: Based on the valid low-altitude scene constraint objects retained after the accessibility counter-proof screening, candidate maneuver behavior chains consisting of multiple maneuver behavior stages connected sequentially are generated in chronological order. Each maneuver phase includes the maneuver state, the dominant low-altitude scene constraint object, and the corresponding phase time range. The maneuvering behavior states include at least one of the following: straight-line crossing, flying along a low-altitude scene constraint object with a passage attribute, avoiding a low-altitude scene constraint object with a dodge attribute, hovering or low-speed observation, approaching a sensitive area, returning home, circling, descending or landing, entering an observation attenuation region, and reappearing from an observation attenuation region.

7. The intelligent aircraft trajectory prediction method based on deep learning big data according to claim 6, characterized in that, The deep behavior chain prediction model includes a temporal encoder, a behavior chain encoder, and a fusion scorer. The timing encoder is used to extract historical motion codes containing the trajectory confidence markers; The behavior chain encoder is used to extract the behavior chain code of the candidate maneuver behavior chain; The fusion scorer is used to fuse the historical motion code, the behavior chain code, and the hierarchical low-altitude reachability features of the corresponding candidate maneuver behavior chains, and outputs the occurrence probability of each candidate maneuver behavior chain, the expected arrival time of the corresponding dominant low-altitude scene constraint object, and the arrival time uncertainty.

8. The intelligent aircraft trajectory prediction method based on deep learning big data according to claim 1, characterized in that, Generating the primary and backup predicted trajectories includes: The spatial extension of each candidate maneuvering behavior chain is restricted to the hierarchical low-altitude reachable domain of the corresponding time layer to generate the corresponding candidate trajectory. Based on the occurrence probability of each candidate maneuvering behavior chain, combined with the reachability consistency evaluation used to characterize whether the candidate trajectory meets the reachability connectivity requirements of the corresponding time layer, the scene constraint consistency evaluation used to characterize whether the candidate trajectory matches the dominant low-altitude scene constraint object, and the trajectory credibility support evaluation based on trajectory credibility markers to characterize the degree of support for historical motion trends, the comprehensive score of each candidate trajectory is calculated by weighting. The candidate trajectory with the highest comprehensive score is selected as the main predicted trajectory, and several candidate trajectories with high scores and different candidate maneuver behavior chain types are selected as backup predicted trajectories.

9. The intelligent aircraft trajectory prediction method based on deep learning big data according to claim 8, characterized in that, Generating the primary and backup prediction trajectories further includes: Discontinuous observation posterior update steps: When the target aircraft is briefly lost, maintain the current primary and backup predicted trajectories and continue to expand the corresponding candidate maneuver behavior chains; When the target aircraft reappears and acquires a new observation status, the matching degree is determined based on the positional and heading deviations between the new observation status and the predicted status of each candidate maneuvering behavior chain at the current moment. The matching degree is then used to perform a posteriori update on the probability of occurrence of each candidate maneuvering behavior chain before the target was lost, so as to correct the main predicted trajectory and the backup predicted trajectory.

10. An intelligent trajectory prediction system for aircraft based on deep learning big data, characterized in that, include: The multi-source observation fusion and labeling module is used to acquire multi-source observation data and low-altitude scene data of the target aircraft, perform spatiotemporal alignment and target association on the multi-source observation data, form a historical trajectory sequence, and generate trajectory confidence labels for trajectory points in the historical trajectory sequence. The constraint object extraction and attribute determination module is used to extract low-altitude scene constraint objects from the low-altitude scene data and determine the effect attributes based on the way the low-altitude scene constraint objects affect the future motion of the target aircraft. The reachability domain generation and disproving module is used to generate a hierarchical low-altitude reachability domain based on the current state, maneuverability and low-altitude environmental constraints of the target aircraft, and to use the hierarchical low-altitude reachability domain to perform reachability disproving and filtering of the low-altitude scenario constraint objects to obtain effective low-altitude scenario constraint objects for generating candidate maneuvering behavior chains. The candidate behavior chain generation module is used to form an effective observation window based on the historical trajectory sequence and the corresponding trajectory confidence marker, and in combination with the effective low-altitude scene constraint object, generate candidate maneuver behavior chains that the target aircraft may experience in the future prediction time window. The depth prediction module is used to input the effective observation window, the trajectory confidence marker, the effective low-altitude scene constraint object, the hierarchical low-altitude reachability domain, and the candidate maneuver behavior chain into the depth behavior chain prediction model, and output the occurrence probability and arrival time distribution of each candidate maneuver behavior chain; The primary and backup trajectory generation module is used to generate the primary predicted trajectory and backup predicted trajectory of the target aircraft based on the occurrence probability, the arrival time distribution, and the layered low-altitude reachability domain.

Images