Flight route dynamic optimization method based on air traffic control data

Abstract

The invention relates to the technical field of air traffic management, and discloses a flight route dynamic optimization method based on air traffic control data. The method comprises the following steps: firstly, acquiring real-time air traffic control data including a current aircraft state, an airspace environment parameter and dynamic airspace event information; initial flight path parameters are determined according to the two, hierarchical feature extraction is performed on the dynamic airspace event information based on a multi-level airspace topology network, and multi-level airspace event features are generated; updated flight situation parameters are generated through iterative feature fusion, and a similar flight situation parameter set is determined by combining the similarity with historical flight situation parameters; and generating a dynamic optimization prompt text, inputting the dynamic optimization prompt text into a trajectory prediction model to obtain predicted flight route parameters, and finally determining a target optimization flight route by combining a preset constraint condition. According to the method, the dynamic airspace change can be responded in real time, and the adaptability and reliability of flight route optimization are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of air traffic management, in particular to a flight route dynamic optimization method based on air traffic control data. BACKGROUND

[0002] With the rapid development of the aviation transportation industry, air traffic flow continues to grow, and the airspace environment is becoming increasingly complex, which puts higher requirements on flight route optimization. Traditional flight route planning methods are mostly based on static airspace information, which is difficult to respond to dynamic changes in the airspace environment in real time, resulting in reduced flight efficiency, increased potential conflict risk, and other problems.

[0003] In actual operation, the airspace environment is affected by various factors such as weather changes, temporary regulations, and sudden conditions of aircraft, etc. These dynamic airspace events will change the available state of the airspace in real time. Existing optimization methods often lack effective hierarchical analysis mechanisms when dealing with these dynamic events, and cannot fully extract the characteristic information of the events, making the judgment of the airspace state not accurate enough. At the same time, when combining historical flight data for route optimization, due to the lack of accurate identification and use of similar flight situations, it is difficult to learn from past experience to generate a reasonable optimization scheme, resulting in insufficient adaptability and reliability of the optimization results.

[0004] Traditional trajectory prediction models rely on labeled data in the training process, and the use of a large amount of unlabeled flight data is insufficient, limiting the generalization ability of the model. When faced with complex and variable airspace scenarios, the model is difficult to generate a predicted route that meets the actual constraint conditions, further affecting the effectiveness of flight route optimization. The existence of these problems makes the existing flight route optimization method difficult to meet the demand of modern air traffic management for high efficiency, safety, and dynamic adaptation, and a new technical solution is needed to solve it. SUMMARY

[0005] The purpose of the present application is to provide a flight route dynamic optimization method based on air traffic control data to solve the problems raised in the background.

[0006] To achieve the above purpose, the present application provides a flight route dynamic optimization method based on air traffic control data, which comprises: Obtaining real-time air traffic control data, the real-time air traffic control data including current aircraft state parameters, airspace environment parameters and dynamic airspace event information; Determining initial flight route parameters according to the current aircraft state parameters and the airspace environment parameters; Based on a multi-level airspace topology network, hierarchical feature extraction is performed on the dynamic airspace event information to generate multi-level airspace event features; According to the initial flight route parameters and the multi-level airspace event characteristics, iterative feature fusion processing is performed to generate updated flight situation parameters; According to the similarity between the updated flight situation parameters and the historical flight situation parameters corresponding to other aircraft, a similar flight situation parameter set is determined; According to the updated flight situation parameters and the similar flight situation parameter set, a dynamic optimization prompt text is generated; The dynamic optimization prompt text is input into a trajectory prediction model, and a predicted flight route parameter is output; According to the predicted flight route parameter and a preset flight route constraint condition, a target optimized flight route is determined.

[0007] Preferably, based on the multi-level airspace topology network, hierarchical feature extraction is performed on the dynamic airspace event information to generate multi-level airspace event characteristics, including: In the multi-level airspace topology network, a to-be-processed airspace event entity of the current level is determined; The to-be-processed airspace event entity is subjected to flight parameter feature conversion to generate an airspace event feature vector of the current level; In the multi-level airspace topology network, a to-be-processed airspace event entity of the next level of the current level is determined; According to the to-be-processed airspace event entity of the next level and the airspace event feature vector of the current level, feature fusion processing is performed to generate a target updated airspace event feature; The target updated airspace event feature is taken as an updated airspace event feature, the next level is taken as the current level, and the step of determining the to-be-processed airspace event entity of the next level is returned to be executed until the feature fusion processing of the highest level is completed; The updated airspace event feature corresponding to the highest level is taken as the multi-level airspace event characteristics.

[0008] Preferably, the to-be-processed airspace event entity is subjected to flight parameter feature conversion to generate an airspace event feature vector of the current level, including: The airspace event description text corresponding to the to-be-processed airspace event entity is acquired; The airspace event description text is processed by a pre-configured airspace semantic encoder to output an event semantic feature vector; The event semantic feature vector is subjected to flight parameter dimension conversion to generate the airspace event feature vector of the current level.

[0009] Preferably, the feature fusion processing is performed according to the to-be-processed airspace event entity of the next level and the airspace event feature vector of the current level to generate a target updated airspace event feature, including: performing mean value calculation on airspace event feature vectors corresponding to all to-be-processed airspace event entities of the next level to generate secondary airspace event features; performing weighted fusion processing on the secondary airspace event features and the updated airspace event features to generate the target updated airspace event feature.

[0010] Preferably, the determining of the similar flight situation parameter set according to the similarity between the updated flight situation parameter and historical flight situation parameters corresponding to other aircrafts comprises: calculating airspace state similarity between the updated flight situation parameter and each historical flight situation parameter; screening specific historical flight situation parameters whose airspace state similarity exceeds a preset similarity threshold; adding historical flight route parameters corresponding to the specific historical flight situation parameters to the similar flight situation parameter set.

[0011] Preferably, the generating of the dynamic optimization prompt text according to the updated flight situation parameter and the similar flight situation parameter set comprises: determining a current flight constraint condition according to the updated flight situation parameter; extracting similar flight route trajectory data from the similar flight situation parameter set; performing path description text synthesis on the current flight constraint condition and the similar flight route trajectory data to generate the dynamic optimization prompt text.

[0012] Preferably, before the acquiring of the real-time air traffic control data, the method further comprises: constructing a multi-level airspace topology network, comprising: acquiring historical airspace operation data, the historical airspace operation data containing aircraft trajectory instances, airspace regulation event instances and flight conflict instances; determining airspace topology relationships between the aircraft trajectory instances, the airspace regulation event instances and the flight conflict instances; generating the multi-level airspace topology network according to the airspace topology relationships.

[0013] Preferably, the trajectory prediction model is obtained by training through the following steps: acquiring a flight training sample set, the flight training sample set containing labeled flight route parameters; processing sample dynamic optimization prompt texts through the trajectory prediction model to output predicted flight route parameters; adjusting parameters of the trajectory prediction model according to a difference value between the predicted flight route parameters and the labeled flight route parameters.

[0014] Preferably, before processing the sample dynamic optimization prompt text through the trajectory prediction model, the method further comprises: Obtaining flight pre-training samples of unannotated flight situation parameters; Performing contrast learning feature coding on the flight pre-training samples to generate flight contrast feature vectors; Processing the flight contrast feature vectors through an initial trajectory prediction model to output predicted flight feature parameters; Determining a model pre-training loss value according to the predicted flight feature parameters and the flight contrast feature vectors; Optimizing parameters of the initial trajectory prediction model according to the model pre-training loss value to obtain a pre-trained trajectory prediction model.

[0015] Preferably, the method of performing contrast learning feature coding on the flight pre-training samples to generate flight contrast feature vectors comprises: Identifying airspace topology patterns in the flight pre-training samples; Matching core flight feature encodings corresponding to the airspace topology patterns; Generating contrast feature encodings of the core flight feature encodings; Performing feature coupling processing on the core flight feature encodings and the contrast feature encodings to generate the flight contrast feature vectors.

[0016] Compared with the prior art, the method has the following beneficial effects: By obtaining real-time air traffic control data, current aircraft state parameters, airspace environment parameters, and dynamic airspace event information are covered, providing comprehensive and real-time basic information for dynamic optimization of flight routes, which can timely capture various changes in the airspace.

[0017] When processing dynamic airspace event information, a multi-level airspace topology network is used for hierarchical feature extraction to generate multi-level airspace event features. This hierarchical processing method can deeply mine the features of dynamic airspace events from different levels, considering both the details of local airspace events and the associations between events at different levels, making the understanding of dynamic airspace events more comprehensive and in-depth, and providing a more accurate basis for subsequent route optimization.

[0018] Through iterative feature fusion processing of initial flight route parameters and multi-level airspace event features, updated flight situation parameters are generated, realizing effective combination of the initial route and dynamic event features, dynamically adjusting the judgment of the flight situation, and making the description of the flight situation more consistent with the real-time airspace situation.

[0019] In the process of determining the similar flight situation parameter set, according to the similarity between the updated flight situation parameter and the historical flight situation parameters of other aircrafts, the historical flight data that can be used for reference can be accurately screened out. These similar historical data contain the flight experience in the similar airspace situation in the past, which provides valuable reference for the current route optimization, and helps to generate a more adaptive optimization scheme.

[0020] According to the updated flight situation parameter and the similar flight situation parameter set, a dynamic optimization prompt text is generated, which is input into the trajectory prediction model to output a predicted flight route parameter, and a target optimized flight route is determined combined with the preset constraint condition, realizing the effective fusion of historical experience and real-time data. This way makes full use of the effective information in the historical flight data, and combines real-time dynamic adjustment, so that the generated target optimized flight route can better adapt to the complex and changeable airspace environment.

[0021] In the process of constructing the multi-level airspace topology network, based on the historical airspace operation data, the airspace topology relationship between various instances is determined, providing a reasonable network structure basis for hierarchical feature extraction, ensuring the effectiveness and systematicness of feature extraction.

[0022] In the training process of the trajectory prediction model, not only the labeled flight training samples are used, but also the unlabeled flight pre-training samples are processed through contrastive learning, making full use of the information in a large amount of unlabeled data and improving the generalization ability of the model. When facing different airspace scenes, the model can more flexibly generate a predicted route that meets the actual needs, enhancing the reliability of route optimization. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 The timing diagram of the flight route dynamic optimization method based on air traffic control data described in the present application; Figure 2 The flowchart for generating multi-level airspace event features; Figure 3 The flowchart for generating airspace event feature vectors; Figure 4 The flowchart for determining the similar flight situation parameter set; Figure 5 The flowchart for constructing the multi-level airspace topology network. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0025] Referring to Figure 1 The application provides a flight route dynamic optimization method based on air traffic control data, and the method comprises.

[0026] The application discloses a flight route dynamic optimization method based on air traffic control data, and the overall implementation scheme is as follows: Real-time air traffic control data is acquired, wherein current aircraft state parameters cover real-time three-dimensional coordinates, instantaneous ground speed, heading angle, climb rate, residual fuel quantity and the like quantitative indexes; airspace environment parameters include real-time wind speed vector, visibility level, cloud layer height, air pressure field distribution, terrain elevation data and the like environmental elements; dynamic airspace event information relates to time-sensitive information such as temporary restriction area division setting notification, air traffic flow saturation early warning, weather mutation alarm and aircraft abnormal state report.

[0027] Initial flight route parameters are determined according to the current aircraft state parameters and the airspace environment parameters. In combination with the coordinate reference of the aircraft takeoff airport and the destination airport, the initial route scheme is generated by referring to the standard instrument departure procedure and the standard instrument approach procedure, fusing the limitation of the cruising altitude layer to the real-time weather conditions, and containing the longitude and latitude of the navigation points, the expected magnetic heading of each flight segment, the planned altitude layer sequence and the expected airspeed of each stage, wherein the navigation point spacing is dynamically adjusted according to the airspace type and the aircraft performance characteristics.

[0028] Based on the multi-level airspace topology network, the hierarchical feature extraction is performed on the dynamic airspace event information to generate multi-level airspace event features. The multi-level airspace topology network is divided into five levels of national control area, regional control area, terminal control area, approach control area and tower control area according to the airspace management responsibility, each level contains corresponding airspace unit boundary, control sector division, communication frequency coverage range and the like basic attributes, and the event features are transmitted and aggregated layer by layer through the association mapping between the levels.

[0029] Iterative feature fusion processing is performed according to the initial flight route parameters and the multi-level airspace event features to generate updated flight situation parameters. The initial route parameter sequence is time-series encoded by using a gated recurrent unit, and the multi-level event features are spatially compressed by using a convolutional neural network, the weight proportion of different features is allocated through an attention mechanism in a feature fusion layer, and after three times of iterative updating, a comprehensive situation parameter set containing spatial position deviation, time delay risk and conflict probability index is formed.

[0030] According to the similarity between the updated flight situation parameters and the historical flight situation parameters corresponding to other aircraft, a similar flight situation parameter set is determined. A six-dimensional feature space containing position, height, speed, weather conditions, and airspace event type is constructed, the similarity value between the current situation and the historical situation is calculated by Mahalanobis distance, and a dynamic threshold is set to screen the historical records that meet the conditions to form a similar case library.

[0031] According to the updated flight situation parameters and the similar flight situation parameter set, a dynamic optimization prompt text is generated. Key constraint elements in the current situation are extracted, including minimum safe height, maximum turning slope, and no-fly zone boundary coordinates, and the route adjustment segment in the similar case is intercepted, and the structured data is converted into a text sequence containing spatial description, timing instruction, and constraint condition by a natural language generation model.

[0032] The dynamic optimization prompt text is input into the trajectory prediction model, and the predicted flight route parameters are output. The trajectory prediction model adopts an encoder-decoder architecture, the encoder performs semantic analysis on the text prompt, and the decoder outputs a flight path point sequence containing the next 120 minutes, each flight path point containing timestamp, latitude and longitude, height, and ground speed.

[0033] According to the predicted flight route parameters and the preset flight route constraint conditions, a target optimized flight route is determined. The preset constraint conditions cover aircraft performance limitations, air traffic rules, and airspace usage restrictions, and the predicted route is verified and corrected by a constraint satisfaction degree checking algorithm, and finally an executable optimized route is generated.

[0034] Embodiment 1: refer to Figure 2 When generating multi-level airspace event feature based on multi-level airspace topology network, first determine the to-be-processed airspace event entity in the current level in the multi-level airspace topology network. The multi-level airspace topology network is divided into levels according to the actual architecture of airspace management, and each level covers a geographical range and management responsibilities, and the airspace event types and processing priorities contained in different levels are different. The determination of the to-be-processed airspace event entity needs to combine the specific location, influence range and airspace management category of the event, and through spatial coordinate matching and attribute label recognition, the event is accurately corresponded to the corresponding level. Each level maintains an event queue, which is sorted according to the event receiving time and influence degree to ensure the subsequent processing in order.

[0035] The flight parameter feature conversion is performed on the to-be-processed airspace event entity to generate an airspace event feature vector of the current level. This process needs to extract various key attributes of the event, including the starting time of the event, the expected duration, the geographical boundary of the affected area, the type of aircraft involved, the associated weather condition changes, etc. These attributes cover multiple dimensions such as time, space, and physical restrictions, and need to be converted into a unified numerical form through standardization processing to eliminate the influence of dimensional differences. For example, the time attribute is converted into seconds relative to a certain reference time, the geographical boundary is converted into a sequence of latitude and longitude coordinates, and the impact degree is converted into a corresponding numerical value according to the preset level standard. After conversion, these attributes will be combined to form a fixed-length vector, with each element corresponding to a feature, fully reflecting the core information of the airspace event.

[0036] In the multi-level airspace topology network, the next level of to-be-processed airspace event entity of the current level is determined. There is a clear inclusion or association relationship between levels, and the airspace range of the previous level usually covers multiple next level airspace units. According to the impact range and spatial position of the current level event, all next level airspace units associated with the current event are retrieved through the level mapping relationship. Among these next level airspace units, the airspace events that are occurring or about to occur are screened out to form the next level of to-be-processed event set. This process needs to consider the spatial overlap and event association to ensure that the next level events have actual impact or association with the current level events, and avoid the introduction of irrelevant events.

[0037] According to the next level of to-be-processed airspace event entity and the airspace event feature vector of the current level, the feature fusion processing is performed to generate the target updated airspace event feature. First, the same feature conversion is performed on each to-be-processed event of the next level to obtain their respective feature vectors. Then, according to the association closeness of these events with the current level event, a corresponding weight is assigned to the feature vector of each next level event. The association closeness can be calculated by factors such as spatial distance, impact range overlap, and event type similarity. Then, the weighted next level event feature vectors are aggregated to form a comprehensive secondary feature vector. The secondary feature vector is then fused with the airspace event feature vector of the current level, and the fusion process adopts a weighted summation method of vector elements, with the weights being set according to the importance of the level and the impact of the event. Finally, a new feature vector is generated, which is the target updated airspace event feature.

[0038] The target update airspace event feature is taken as the update airspace event feature, the next level is taken as the current level, and the step of determining the next level to be processed airspace event entity is returned for execution until the feature fusion processing of the highest level is completed. After each level progresses, the event retrieval, feature conversion and fusion process described above needs to be repeated. In this iteration process, the features of each level absorb relevant event information from the next level, so that the feature vector is constantly enriched and improved, and gradually transmitted upwards. With the promotion of the level, the airspace range covered by the feature vector is continuously expanded, and the granularity of the event information is also adjusted accordingly, from specific local events to more macro regional event features. In each iteration, the normalized feature vector after fusion needs to be processed to ensure that the numerical range of the vector remains stable and to avoid numerical overflow or feature distortion after multiple fusions.

[0039] The update airspace event feature corresponding to the highest level is taken as the multi-level airspace event feature. When the iteration process reaches the preset highest level, the update airspace event feature at this time has fused all relevant airspace event information from the bottom level to the highest level, and contains event features of different granularities and ranges. This feature vector can comprehensively reflect the overall situation of various events in the current airspace environment, including both local subtle event changes and large-scale event trends. After generating the final multi-level airspace event feature, the elements in the vector can be selected by a feature selection algorithm to retain more influential features for subsequent processing, remove redundant or secondary features, and further optimize the quality of the feature vector to make it more suitable for dynamic optimization analysis of flight routes.

[0040] Embodiment 2: see Figure 3 When the flight parameter feature conversion is performed on the to-be-processed airspace event entity to generate the airspace event feature vector of the current level, the airspace event description text corresponding to the to-be-processed airspace event entity is first obtained. These text information usually comes from real-time reports of the air traffic control system, warning information of the meteorological department and temporary notifications of airspace management, etc., covering the basic attributes of the event, such as the specific time of the event, the airspace range involved, the nature of the event (such as weather mutation, temporary flight ban, traffic control, etc.), the expected duration and the type of impact on flight, etc. The text content needs to be organized according to uniform format specifications to ensure the integrity and accuracy of the key information and to avoid ambiguous or ambiguous expressions, so that the semantic analysis can accurately capture the core content of the event.

[0041] The airspace event description text is processed by the pre-set airspace semantic encoder to output an event semantic feature vector. The encoder is constructed based on a deep neural network architecture and trained on a large amount of aviation field text data, and can understand the specific meaning of aviation terminology and the context information of event description. During processing, the text is first segmented into several basic semantic units (such as words or sub-phrases), and then each semantic unit is converted into a low-dimensional vector representation through an embedding layer to capture its basic semantic information. Then, through a multi-layer encoding structure, the vectors are processed to consider the order and relationship between semantic units, such as the causal relationship between "thunderstorm area" and "prohibited entry", the logical relationship between "traffic saturation" and "speed limit instruction", etc. After such processing, the text information is converted into an event semantic feature vector of a fixed dimension, and each element in the vector corresponds to different dimensional semantic information in the text, which can fully reflect the deep meaning of the event description.

[0042] The event semantic feature vector is converted into a flight parameter dimension to generate an airspace event feature vector of the current level. Since the event semantic feature vector mainly reflects the semantic information of the text, its dimension and numerical range do not completely match the feature space of flight parameters (such as position, speed, height, etc.), so dimension conversion is needed to make the generated feature vector directly fuse with the state parameters of the aircraft, airspace environmental parameters, etc. The conversion process is implemented through a pre-set mapping matrix, which is constructed according to the characteristics of the flight parameter feature space and the distribution of semantic features, and can adjust the dimension of the semantic feature vector to adapt to the dimension of the flight parameter feature space. In the conversion process, the key information in the semantic features, such as the size of the event's impact range and the degree of impact, should be preserved, and the converted feature vector should be consistent in numerical distribution with other flight parameters to facilitate subsequent feature fusion.

[0043] When generating the target updated airspace event feature according to the airspace event entity to be processed at the next level and the airspace event feature vector at the current level, first, the airspace event feature vectors corresponding to all airspace event entities to be processed at the next level are subjected to mean value calculation to generate a secondary airspace event feature. When there are multiple events to be processed at the next level, each event corresponds to an airspace event feature vector, and these vectors reflect the feature information of different events. In the mean value calculation process, these vectors need to be averaged according to the corresponding dimensions, that is, the mean value of each dimension is equal to the sum of all vector elements at the dimension divided by the number of vectors. For some special cases, such as an event whose influence range is much larger than that of other events or whose emergency level is obviously higher, a higher weight can be assigned to it in the mean value calculation to highlight its influence in the secondary feature. Through such mean value calculation, the feature information of multiple events at the next level is aggregated into a secondary airspace event feature vector, which can reflect the overall situation of events at the next level.

[0044] The secondary airspace event feature and the airspace event feature vector at the current level are subjected to weighted fusion processing to generate the target updated airspace event feature. In the fusion process, different weights need to be assigned to the two feature vectors according to the importance of the level and the influence of the event. Generally, the event feature at the current level reflects more macroscopic or more direct influence, and the weight is relatively high, while the secondary feature at the next level serves as supplementary information, and the weight is relatively low. The specific values of the weights are determined according to the actual needs of airspace management and the analysis results of historical data, and can be dynamically adjusted according to actual conditions. The calculation formula for weighted fusion is: wherein: represents the target updated airspace event feature, represents the airspace event feature vector at the current level, represents the secondary airspace event feature vector, represents the weight coefficient of the feature vector at the current level, represents the weight coefficient of the secondary feature vector, and Through such weighted summation, the features at the current level and the aggregated features at the next level are organically fused together to generate a target updated airspace event feature that not only retains the main information of the event at the current level but also integrates the related influence of the event at the next level, and can more comprehensively reflect the comprehensive situation of the airspace event. After the fusion is completed, the target updated airspace event feature needs to be subjected to standardization processing to adjust its numerical range to a preset interval, so as to avoid affecting the calculation accuracy of the subsequent calculation due to too large numerical difference.

[0045] Embodiment 3: see Figure 4According to the similarity between the updated flight situation parameter and the historical flight situation parameters corresponding to other aircrafts, when determining the similar flight situation parameter set, first, the airspace state similarity between the updated flight situation parameter and each historical flight situation parameter is calculated. The updated flight situation parameter covers the real-time position, flight height, cruising speed, heading angle, remaining fuel of the current aircraft, and the weather conditions, air traffic flow density, surrounding airspace restrictions and other information of the airspace where the aircraft is located. The historical flight situation parameter is the state record of the past aircraft in a similar airspace environment, which contains various attributes corresponding to the updated flight situation parameter. When calculating the similarity, each parameter needs to be compared one by one. The position parameter is converted into a latitude and longitude coordinate difference, the speed parameter is converted into a vector difference, the weather parameter is converted into a grade difference value, and the traffic flow parameter is converted into a density deviation rate. Through these conversions, different types of parameters are unified into quantifiable difference indicators, and the overall airspace state similarity is obtained through comprehensive calculation.

[0046] The specific historical flight situation parameters with airspace state similarity exceeding the preset similarity threshold are screened. The preset similarity threshold is dynamically set according to the airspace complexity. In high-altitude open airspace, the threshold can be appropriately reduced to include more reference cases. In complex airspace such as airport terminal area and air corridor, the threshold is correspondingly increased to ensure that the selected historical cases are highly matched with the current situation. During the screening process, the similarity value of each historical flight situation parameter needs to be judged. Only the historical parameters exceeding the threshold are considered as specific historical flight situation parameters. At the same time, in order to avoid the screening result being too single, it is also necessary to ensure that the specific historical flight situation parameters cover different flight stages and different weather condition combinations.

[0047] The historical flight route parameters corresponding to the specific historical flight situation parameters are added to the similar flight situation parameter set. The historical flight route parameters include the complete track data of the past aircraft in the corresponding situation, such as the coordinates of the navigation points passed, the flight time of each segment, the height layer adjustment record, the speed change curve, the number of heading corrections, etc. These parameters need to be structured and sorted according to time sequence and spatial coordinates, and abnormal data and repeated records need to be removed. When added to the set, each historical flight route parameter needs to be labeled with the corresponding historical flight situation characteristics, such as the control instructions at that time, the emergency weather response measures, etc.

[0048] When generating the dynamic optimization prompt text according to the updated flight situation parameter and the similar flight situation parameter set, first, the current flight constraint condition is determined according to the updated flight situation parameter. The current flight constraint condition includes hard constraints and flexible constraints. The hard constraints are, for example, the geographical boundary of the no-fly zone (a polygon area defined by latitude and longitude coordinates), the minimum safety height (determined according to the terrain elevation and obstacle height), and the maximum flight speed limit (set according to the air traffic flow and aircraft performance). The flexible constraints are, for example, the deviation allowed range of the recommended route, the avoidance suggestion of the traffic-intensive area, and the crossing time of the weather-affected area. These constraint conditions need to be dynamically generated according to the real-time data in the updated flight situation parameter. For example, when a new thunderstorm area is detected, it needs to be immediately included in the hard constraint of the no-fly zone.

[0049] The similar flight route trajectory data is extracted from the similar flight situation parameter set. The extraction process needs to focus on the trajectory segment that matches the current flight phase. For example, when the current aircraft is in the cruise phase, the key extraction data of the cruise phase in the historical case is highlighted. The trajectory data includes the coordinates of the key decision point, the environmental parameters at the time of decision, and the effect of the flight path change after the decision. At the same time, the deviation degree of these similar trajectories from the current initial flight route needs to be calculated, and the trajectory data with a deviation within a reasonable range is retained.

[0050] The path description text is synthesized for the current flight constraint condition and the similar flight route trajectory data, and the dynamic optimization prompt text is generated. The synthesis process uses a structured language template. First, the current constraint conditions are clearly listed in the format of “prohibited entry area: XXX latitude and longitude to XXX latitude and longitude”, “minimum safety height: XXX meters”, etc. Then, based on the similar flight route trajectory data, the flight route adjustment suggestions are extracted, such as “turn at north latitude XXX degrees, east longitude XXX degrees to avoid the traffic-intensive area 50 kilometers ahead, and reference the historical case, the turn can reduce the flight time by XXX minutes”, “it is recommended to maintain level flight at an altitude of XXX meters, and then continue to climb after passing through the cloud layer, which is consistent with the optimal trajectory under similar weather conditions in history”, etc. The text synthesis needs to ensure that the language is concise and accurate, consistent with the aviation communication specifications, and avoid ambiguous expressions. The comprehensive dynamic optimization prompt text is generated by weighted fusion, and the formula is as follows: wherein: represents the comprehensive feature value of the dynamic optimization prompt text, represents the feature value of the current flight constraint condition, represents the feature value of the similar flight route trajectory data, represents the weight coefficient of the current flight constraint condition, represents the weight coefficient of the similar flight route trajectory data, and Through this formula, the prompt text strictly follows the current constraints and fully absorbs historical experience, forming an optimized guide with normative and practical nature.

[0051] Embodiment 4: Refer to Figure 5 I will combine specific examples to describe the implementation of embodiment 4 from two aspects: building a multi-level airspace topology network and training a trajectory prediction model, which will include a related data table.

[0052] Before obtaining real-time air traffic control data, building a multi-level airspace topology network requires several steps. Obtain historical airspace operation data, which includes aircraft trajectory instances, airspace control event instances, and flight conflict instances. Aircraft trajectory instances can be the detailed trajectories of all flights of a certain airline on the Beijing-Shanghai route in the past six months. The trajectory data of each flight is recorded in chronological order, including latitude, longitude, altitude, ground speed, and other information every minute. Airspace control event instances may include temporary control notifications for the airspace along the route during this period, such as restrictions on travel in a certain airspace due to military activities during a specific time period. The notification will specify the time range of the control, the airspace coordinates involved, etc. Flight conflict instances may include records of two aircrafts coming too close to each other near the route, including the time of the conflict, the positions of the two aircrafts, and the speed information.

[0053] Through analysis, it can be found that the trajectories of some flights will bypass the controlled airspace during a certain time period, thereby forming a spatial avoidance relationship between trajectories and control events. For aircraft trajectory instances and flight conflict instances, the trajectory of one aircraft may cross the trajectory of another aircraft due to untimely speed adjustment, resulting in a conflict, which constitutes a causal relationship between trajectories and conflict events. Through such analysis, the association between various instances can be sorted out.

[0054] Generate a multi-level airspace topology network based on these airspace topology relationships. Different levels of the network correspond to different airspace management ranges, such as the national airspace level covering a wide range, including multiple regional airspace units; regional airspace includes several terminal control areas, etc. Each level of the node represents the corresponding airspace unit or event, and the edge between the nodes represents their association, such as inclusion relationship, time sequence relationship, etc.

[0055] The training of the trajectory prediction model also needs to be done in steps. First, obtain the flight training sample set, which contains labeled flight route parameters. These samples can be flight data on the Beijing-Shanghai route under different seasons and weather conditions, and the labeled flight route parameters are determined by experienced controllers based on actual conditions, including optimal waypoints, altitude layers, etc.

[0056] The trajectory prediction model processes the sample dynamic optimization prompt text and outputs predicted flight path parameters. The sample dynamic optimization prompt text may include weather conditions during the sample flight, such as visibility of 10 kilometers and wind speed of 5 meters per second, as well as airspace restriction information, such as airspace control from 8 a.m. to 10 a.m. After the model processes this text, it generates corresponding predicted route parameters.

[0057] The model parameters are adjusted according to the difference between the predicted flight path parameters and the labeled flight path parameters. Through continuous adjustment, the prediction results of the model become more accurate.

[0058] Table 1: Partial flight vehicle trajectory instance profile in historical airspace operation data.

[0059] When constructing a multi-level airspace topology network, the trajectory data in the above table is used in combination with airspace control events and flight conflict instances to sort out the relationships between various instances and generate a network. When training the trajectory prediction model, these trajectory data become part of the training samples, helping the model better learn and predict flight path parameters. In this way, whether it is the construction of a multi-level airspace topology network or the training of a trajectory prediction model, both can be based on actual flight data to adapt to complex airspace environments and flight requirements.

[0060] Example 5: Before processing the sample dynamic optimization prompt text through the trajectory prediction model, flight pre-training samples without labeled flight situation parameters need to be obtained. These samples cover flight data in different regions and time periods, including flight trajectories of aircraft in various weather conditions, detailed information of airspace structure, and corresponding environmental parameters. For example, it includes flight paths of aircraft in different weather conditions such as sunny, rainy, and foggy days, trajectory records crossing different terrains such as mountains, plains, and oceans, and flight data in busy and empty airspace. These samples do not contain artificially labeled flight situation parameters, but only retain the original flight data information, such as the position coordinates of the aircraft recorded at regular intervals, flight altitude, flight speed, heading angle, etc., accompanied by meteorological data such as wind speed, wind direction, visibility, and the distribution of other aircraft in the airspace at the time.

[0061] When generating flight contrast feature vectors by contrastive learning feature encoding on pre-training flight samples, first, pre-process the sample data to remove outliers such as position jumps and speed anomalies caused by equipment failure, ensuring the reliability of the samples. Then, use data augmentation techniques to diversify the samples, such as slightly stretching or compressing the same flight trajectory in the time dimension to simulate minor changes in flight speed; slightly shifting the spatial position of the trajectory to simulate the flight path differences of different aircraft in the same airspace; or adding minor random disturbances to the trajectory to simulate minor heading adjustments that may occur during actual flight. Through these processes, multiple similar but not identical variant samples are generated for each original sample.

[0062] Process the flight contrast feature vectors through the initial trajectory prediction model to output predicted flight feature parameters. The network structure of the initial trajectory prediction model includes multiple hidden layers. The input layer receives the flight contrast feature vectors, and the output layer generates predicted flight feature parameters after calculation and processing by multiple layers of neural networks. These predicted parameters include information such as the possible position, altitude, and speed of the aircraft within a certain period of time. During processing, the model converts and extracts input features through connection weights between neurons in each layer, gradually learning the underlying patterns in the samples, such as typical flight trajectory characteristics of aircraft under different weather conditions and influence patterns of airspace structure on flight paths.

[0063] Determine the model pre-training loss value based on the predicted flight feature parameters and the flight contrast feature vectors. Compare the predicted flight feature parameters output by the model with the actual flight data contained in the flight contrast feature vectors and calculate the difference between them. The difference calculation involves multiple dimensions, such as the deviation of position coordinates, the difference in altitude, and the difference in speed, etc. The total loss value is obtained by combining the differences in these dimensions. This loss value reflects the degree of deviation of the model's prediction results from the actual situation. The smaller the loss value, the better the prediction effect of the model.

[0064] Optimize the parameters of the initial trajectory prediction model based on the model pre-training loss value to obtain the pre-trained trajectory prediction model. During optimization, adjust the connection weights and bias parameters of each layer of the model based on the calculated loss value. Through the backpropagation algorithm, the loss value is propagated from the output layer to the input layer in reverse, adjusting the parameters layer by layer to reduce the difference between the prediction results and the actual data. This process is repeated iteratively. Each iteration uses a portion of the pre-training samples to train the model, calculates the loss value and adjusts the parameters until the loss value stabilizes at a relatively low level. At this point, the model can better capture the flight patterns in the samples, forming the pre-trained trajectory prediction model.

[0065] When generating flight contrast feature vectors by contrastive learning feature encoding on pre-training samples, first, the spatial topology patterns in the samples are identified. By analyzing a large amount of flight trajectory data, different spatial usage patterns are summarized, such as the flight flow distribution characteristics of certain airspace in a specific time period, the common path patterns of aircraft when crossing airspace boundaries, and the differences in flight trajectories of different types of aircraft in the same airspace. These patterns are automatically classified by clustering algorithms, and flight samples with similar characteristics are classified into the same pattern category.

[0066] The core flight feature encoding corresponding to the matched spatial topology pattern is matched. A unique core feature encoding is assigned to each identified spatial topology pattern, which contains key feature information of flight trajectories in this pattern, such as typical height variation range, common speed interval, and main heading change node. The core feature encoding is obtained by statistical analysis and extraction of all samples in this pattern, and can represent the essential features of this pattern.

[0067] The contrast feature encoding of the core flight feature encoding is generated. The contrast feature encoding is generated based on the core feature encoding, and is obtained by controlled adjustment of part of the features in the core encoding. For example, small fluctuations are made within the height range reflected in the core encoding, slight adjustments are made within the typical speed interval, or slight shifts are made to the positions of the heading change nodes. These adjustments ensure that the contrast encoding has some differences from the core encoding, while maintaining a high degree of similarity, and can reflect the subtle changes in the same spatial topology pattern.

[0068] The core flight feature encoding and the contrast feature encoding are coupled to generate a flight contrast feature vector. The core encoding and the contrast encoding are combined according to certain rules, such as concatenating the feature elements of the two in order, or calculating the correlation values of the two in different dimensions, to form a comprehensive feature vector. This process makes the feature vector contain not only the core features of the spatial topology pattern, but also the change information within the pattern, which can provide more rich feature input for subsequent model pre-training, and help the model better learn the flight rules in different spatial topology patterns.

[0069] It should be noted that in this paper, relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device.

[0070] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A flight route dynamic optimization method based on air traffic control data, characterized in that: include: Acquiring real-time air traffic control data, including current aircraft status parameters, airspace environment parameters, and dynamic airspace event information; Determining initial flight path parameters according to the current aircraft state parameters and the airspace environment parameters; Based on a multi-level airspace topology network, hierarchical feature extraction is performed on the dynamic airspace event information to generate multi-level airspace event features; Performing iterative feature fusion processing based on the initial flight path parameters and the multi-level airspace event features to generate updated flight situation parameters; Determining a similar flight situation parameter set based on similarities between the updated flight situation parameters and historical flight situation parameters corresponding to other aircraft; Generate a dynamic optimization prompt text according to the updated flight situation parameter and the similar flight situation parameter set; Inputting the dynamic optimization prompt text into a trajectory prediction model and outputting predicted flight path parameters; A target optimized flight route is determined based on the predicted flight route parameters and preset flight route constraints.

2. The flight route dynamic optimization method based on air traffic control data according to claim 1, characterized in that: The step of extracting hierarchical features from the dynamic airspace event information based on the multi-level airspace topology network to generate multi-level airspace event features includes: Determining an airspace event entity to be processed at a current level in the multi-level airspace topology network; Performing flight parameter feature conversion on the airspace event entity to be processed to generate an airspace event feature vector at the current level; Determining, in the multi-level airspace topology network, an airspace event entity to be processed at a level next to the current level; Performing feature fusion processing based on the airspace event entity to be processed at the next level and the airspace event feature vector at the current level to generate a target updated airspace event feature; Taking the target updated airspace event feature as the updated airspace event feature, taking the next level as the current level, and returning to the step of determining the airspace event entity to be processed at the next level until the feature fusion processing at the highest level is completed; The updated airspace event features corresponding to the highest level are used as the multi-level airspace event features.

3. The flight route dynamic optimization method based on air traffic control data according to claim 2, characterized in that: The step of performing flight parameter feature conversion on the airspace event entity to be processed to generate an airspace event feature vector at the current level includes: Obtain the airspace event description text corresponding to the airspace event entity to be processed; Processing the spatial domain event description text by a preset spatial domain semantic encoder to output an event semantic feature vector; The event semantic feature vector is converted into a flight parameter dimension to generate the airspace event feature vector of the current level.

4. The flight route dynamic optimization method based on air traffic control data according to claim 2, characterized in that: The step of performing feature fusion processing based on the next-level airspace event entity to be processed and the airspace event feature vector of the current level to generate a target updated airspace event feature includes: Calculating the mean of the airspace event feature vectors corresponding to all airspace event entities to be processed at the next level to generate secondary airspace event features; The secondary airspace event feature and the updated airspace event feature are weightedly fused to generate the target updated airspace event feature.

5. The flight route dynamic optimization method based on air traffic control data according to claim 1, characterized in that: Determining a set of similar flight situation parameters based on the similarity between the updated flight situation parameters and historical flight situation parameters corresponding to other aircraft includes: Calculating the airspace state similarity between the updated flight situation parameter and each historical flight situation parameter; Filtering specific historical flight situation parameters whose airspace state similarity exceeds a preset similarity threshold; The historical flight route parameters corresponding to the specific historical flight situation parameters are added to the similar flight situation parameter set.

6. The flight route dynamic optimization method based on air traffic control data according to claim 1, characterized in that: Generating a dynamic optimization prompt text according to the updated flight situation parameter and the similar flight situation parameter set includes: Determining current flight constraints based on the updated flight situation parameters; Extracting similar flight route trajectory data from the similar flight situation parameter set; The current flight constraint condition and the similar flight route trajectory data are synthesized into a path description text to generate the dynamic optimization prompt text.

7. The flight route dynamic optimization method based on air traffic control data according to claim 1, characterized in that: Before acquiring the real-time air traffic control data, the method further includes: Construct a multi-level airspace topology network, including: Acquiring historical airspace operation data, the historical airspace operation data including aircraft trajectory instances, airspace control event instances, and flight conflict instances; Determining an airspace topological relationship among the aircraft trajectory instance, the airspace control event instance, and the flight conflict instance; The multi-level spatial topology network is generated according to the spatial topology relationship.

8. The flight route dynamic optimization method based on air traffic control data according to claim 1, characterized in that: The trajectory prediction model is trained by the following steps: Acquire a flight training sample set, wherein the flight training sample set includes marked flight route parameters; Processing the sample dynamic optimization prompt text through the trajectory prediction model and outputting predicted flight path parameters; Parameters of the trajectory prediction model are adjusted according to the difference between the predicted flight path parameters and the annotated flight path parameters.

9. The flight route dynamic optimization method based on air traffic control data according to claim 8, characterized in that: Before processing the sample dynamic optimization prompt text by the trajectory prediction model, the method further includes: Obtain flight pre-training samples without labeled flight status parameters; Performing comparative learning feature encoding on the flight pre-training samples to generate a flight comparative feature vector; Processing the flight comparison feature vector through an initial trajectory prediction model to output predicted flight feature parameters; Determining a model pre-training loss value based on the predicted flight characteristic parameters and the flight comparison feature vector; Parameters of the initial trajectory prediction model are optimized according to the model pre-training loss value to obtain a pre-trained trajectory prediction model.

10. The flight route dynamic optimization method based on air traffic control data according to claim 9, characterized in that: The performing comparative learning feature encoding on the flight pre-training sample to generate a flight comparative feature vector includes: Identifying airspace topology patterns in the flight pre-training samples; Match the core flight feature code corresponding to the airspace topology pattern; generating a control feature code for the core flight feature code; Feature coupling processing is performed on the core flight feature code and the comparison feature code to generate the flight comparison feature vector.

Images