Emergency command method and system for low-altitude flight

Abstract

The invention provides an emergency command method and system for low-altitude flight. The method comprises the following steps: acquiring real-time flight monitoring data of a target area; determining an emergency event type in the target area according to an abnormal data subset in the real-time flight monitoring data; calling a flight emergency prediction model corresponding to the emergency event type; the flight emergency prediction models are obtained by training based on flight features and airspace features screened in historical emergency event data, and the types of input parameters and output parameters of each flight emergency prediction model are consistent in a training stage and an application stage; outputting a flight path adjustment strategy and an emergency resource allocation strategy according to the flight emergency prediction model, and generating an emergency command instruction; and executing the emergency command instruction, and updating parameters of the flight emergency prediction model according to the executed real-time feedback data. According to the invention, the problem of cooperative improvement of flight safety and emergency response efficiency in a complex airspace environment can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular, to an emergency command method and system for low-altitude flight. Background Art

[0002] In recent years, with the wide application of low-altitude vehicles such as unmanned aerial vehicles and general aviation aircraft, low-altitude flight emergency command technology has become the core link to ensure flight safety. Most of the existing technologies generate commands based on a preset emergency rule library or a single-dimensional flight emergency prediction model. For example, a fixed obstacle avoidance path is triggered by the positioning data of the aircraft, or an emergency response program is started according to the meteorological warning threshold. Although such methods can handle conventional scenarios, it is difficult to make quick and accurate decisions under complex working conditions such as sudden airspace conflicts and extreme meteorological disturbances.

[0003] There are significant technical bottlenecks in the current low-altitude emergency command system: First, the static rule library cannot dynamically adapt to the multi-factor coupling relationship among sudden changes in the aircraft state, temporary airspace control, and fluctuations in meteorological parameters, resulting in a mismatch between the generated flight path adjustment instructions and the real-time environment; Second, the single flight emergency prediction model lacks a cross-dimensional data collaboration mechanism, and there are spatio-temporal logic conflicts between flight control instructions and resource scheduling strategies due to the fragmentation of data sources; Third, the structural deviation between the model training data and the application scenario parameters causes decision feature drift, resulting in deviation in the position of emergency resource allocation or failure of obstacle avoidance path planning; Fourth, there is a lack of an optimization mechanism driven by the feedback of the instruction execution effect, and the system response ability continuously decays in the face of high-frequency emergencies.

[0004] Due to insufficient dynamic adaptability, lack of multi-source data collaboration, and weak closed-loop optimization ability, the existing technologies lead to systematic defects in low-altitude emergency command such as lag in instruction generation, frequent occurrence of strategy mutual exclusivity, and repeated resource scheduling. It is urgent to build an emergency command architecture that can fuse multi-dimensional state parameters in real time, dynamically correct decision-making logic, and has self-optimization ability to break through the bottleneck of the coordinated improvement of flight safety and emergency response efficiency in complex airspace environments. Summary of the Invention

[0005] In view of this, at least one embodiment of the present invention provides an emergency command method for low-altitude flight.

[0006] The technical solution of the embodiment of the present invention is implemented as follows: On the one hand, the embodiment of the present invention provides an emergency command method for low-altitude flight, and the method includes: obtaining real-time flight monitoring data of a target area; the real-time flight monitoring data includes aircraft state parameters, meteorological parameters, and airspace occupancy parameters; determining the type of emergency event in the target area according to the abnormal data subset in the real-time flight monitoring data; calling a flight emergency prediction model corresponding to the type of emergency event; the flight emergency prediction model is trained based on flight characteristics and airspace characteristics selected from historical emergency event data, and the types of input parameters and output parameters of each flight emergency prediction model are consistent in the training stage and the application stage; generating an emergency command instruction according to the flight path adjustment strategy and the emergency resource allocation strategy output by the flight emergency prediction model; executing the emergency command instruction, and updating the parameters of the flight emergency prediction model according to the real-time feedback data after execution.

[0007] On the other hand, the embodiment of the present invention provides a computer system, including a memory and a processor, the memory stores a computer program that can run on the processor, and the processor implements the steps in the above method when executing the program.

[0008] The emergency command method for low-altitude flight provided by the present invention constructs a multi-dimensional data perception system by fusing aircraft state parameters, meteorological parameters, and airspace occupancy parameters in real time, dynamically matches the type of emergency event with the calling conditions of the flight emergency prediction model, and generates a joint decision instruction with both a flight path adjustment strategy and an emergency resource allocation strategy. It can effectively utilize the correlation between real-time dynamic data and historical training data in the low-altitude flight scenario, break through the limitations of single-dimensional decision-making in traditional emergency responses, and significantly improve the real-time performance and reliability of emergency command in complex airspace environments; by forcibly restricting the consistency of the types of input and output parameters of the flight emergency prediction model in the training stage and the application stage, it ensures that the heading, altitude, and speed parameters in the flight path adjustment strategy and the path, position, and supply parameters in the resource allocation strategy form a coordinated mapping relationship in the spatial dimension and the time dimension, eliminating the risk of strategy conflicts caused by model feature drift; dynamically updating the model parameters based on the feedback data after the instruction execution solves the interference problem of uncertain factors such as sudden weather changes and dynamic changes in airspace occupancy in the low-altitude flight scenario on emergency decision-making, and finally realizes the synchronous optimization of aircraft obstacle avoidance path planning and precise emergency resource delivery, comprehensively improving the efficiency and safety of low-altitude emergency event handling. Description of the Drawings

[0009] Figure 1 It is a schematic flowchart of the implementation of an emergency command method for low-altitude flight provided by an embodiment of the present invention; Figure 2Schematic diagram of the hardware entity of a computer system provided by an embodiment of the present invention. Detailed implementation manners

[0010] An embodiment of the present invention provides an emergency command method for low-altitude flight, which can be executed by a processor of a computer system. Herein, the computer system may refer to devices with data processing capabilities such as servers, laptop computers, tablet computers, desktop computers, and mobile devices.

[0011] Figure 1 Schematic diagram of the implementation process of an emergency command method for low-altitude flight provided by an embodiment of the present invention. As Figure 1 shown, the method includes: Step S100: Obtain real-time flight monitoring data of the target area; wherein, the real-time flight monitoring data includes aircraft state parameters, meteorological parameters, and airspace occupancy parameters.

[0012] Optionally, the real-time flight monitoring data is, for example, a comprehensive data set collected in real time through a sensor network, a radar system, and aviation communication equipment deployed in the target area, and is used to comprehensively reflect the dynamic state of the current low-altitude flight environment. Among them, the aircraft state parameters specifically include core operation indicators such as the real-time flight speed, flight altitude, heading angle, attitude angle, acceleration, remaining fuel quantity, and navigation system state of the aircraft. These parameters are transmitted to the command center in real time through the data link between the aircraft-borne equipment and the ground control system; the meteorological parameters cover the atmospheric environment indicators in the target area, including wind speed, wind direction, temperature, humidity, air pressure, visibility, precipitation intensity, thunderstorm activity area, and turbulence intensity. Such data is collected through multi-source fusion by meteorological satellites, ground meteorological stations, and meteorological detection equipment carried by unmanned aerial vehicles; the airspace occupancy parameters are used to describe the airspace resource usage conditions such as the distribution density of aircraft in the current airspace, the boundary coordinates of the temporary control area, the geographical location and height limits of obstacles (such as buildings, high-voltage lines, mountains), and other flight plan conflict areas, and are obtained through the collaborative monitoring of the airspace management system and the air traffic control radar. Specifically, the collection frequency of the aircraft state parameters needs to meet at least one update per second to ensure the accurate capture of the aircraft dynamics; the collection of the meteorological parameters needs to combine the short-term meteorological forecast model and the real-time observation data to cover the comprehensive influence of the static environment and dynamic changes; the update of the airspace occupancy parameters needs to be synchronized with the air traffic control instructions to ensure the timeliness of the airspace dynamic information. Through the parallel collection and fusion of the above three types of parameters, the real-time flight monitoring data can construct a three-dimensional dynamic mapping model of the aircraft, meteorological environment, and airspace resources in the target area, providing a data basis for subsequent emergency event identification and decision-making.

[0013] Step S200: Determine the type of emergency event in the target area according to the abnormal data subset in the real-time flight monitoring data.

[0014] Exemplarily, the abnormal data subset is, for example, a set of parameters that exceed the preset safety range and are extracted from real-time flight monitoring data through a dynamic threshold screening mechanism, which is used to characterize the possible flight risks or sudden events in the target area. Specifically, the dynamic threshold screening mechanism uses a preset aircraft state threshold, a preset meteorological fluctuation threshold, and a preset airspace capacity threshold as screening criteria: for aircraft state parameters, when the allowable deviation range of the flight speed from the route planning speed, the flight altitude is lower than the minimum safety altitude limit, or the heading angle mutation exceeds the angle tolerance, an abnormal aircraft data mark is triggered; for meteorological parameters, if the mutation gradient of the wind speed within a unit time exceeds the safe flight condition, the visibility decrease rate reaches the critical value affecting the navigation accuracy, or the air pressure fluctuation intensity causes the risk of aircraft attitude instability, it is marked as abnormal meteorological data; for airspace occupancy parameters, when the aircraft density in the airspace exceeds the maximum capacity limit, the overlap degree between the temporary control area and the flight path reaches the conflict threshold, or the obstacle distribution dispersion causes the increase in the difficulty of obstacle avoidance, it is determined as abnormal airspace data. After the abnormal data subset is cross-validated by multi-dimensional parameters, it is input into the event type association model. Based on the mapping relationship between the same abnormal features and event types in the historical emergency event case library, this model calculates the similarity between the current abnormal data and the historical cases through a hierarchical matching mechanism, and outputs the basic classification label of the emergency event type and its priority level. For example, if the abnormal data subset includes the sudden drop in aircraft altitude and strong crosswind meteorological data, the model will match the composite event type of "aircraft out of control" and "meteorological mutation", and generate a priority ranking according to the historical processing time limit and resource requirement parameters. Finally, the event feature label will be associated with the subsequent model call conditions to ensure the pertinence of the emergency response strategy.

[0015] As an implementation manner, in step S200, according to the abnormal data subset in the real-time flight monitoring data, determine the emergency event type in the target area, which may specifically include: step S210: Extract a first data subset corresponding to the aircraft state parameters, a second data subset corresponding to the meteorological parameters, and a third data subset corresponding to the airspace occupancy parameters from the real-time flight monitoring data.

[0016] As a collection of multi-source heterogeneous data, real-time flight monitoring data needs to be classified and extracted into three independent data subsets according to parameter types to support refined analysis. The first data subset is a dedicated collection of aircraft status parameters, specifically including core indicators of aircraft operation such as real-time speed, altitude, heading angle, attitude angle, acceleration, remaining fuel quantity, and navigation system warning status. Such parameters are transmitted in real time through the data link between aircraft on-board sensors and ground control systems, and are classified and stored according to the unique aircraft identifier with timestamps as indexes; The second data subset is a dedicated collection of meteorological parameters, covering wind speed, wind direction, temperature, humidity, air pressure, visibility, precipitation intensity, coordinates of thunderstorm activity areas, and turbulence intensity levels within the target area. Such data is collected through multi-node collaboration of meteorological satellites, ground meteorological radars, and micro-meteorological stations carried by unmanned aerial vehicles, and forms spatio-temporally continuous meteorological field data after being aligned according to grid airspace coordinates; The third data subset is a dedicated collection of airspace occupancy parameters, including the heat map of aircraft density distribution in the airspace, the geographical fence coordinates of the boundary of the temporary control area, the three-dimensional spatial coordinates and height limits of obstacles (such as high-rise buildings, high-voltage transmission towers, mountains), and the time-space occupancy status of other flight plan conflict areas. Such parameters are generated by fusing the dynamic topological map of the airspace management system and the scan data of the air traffic control radar. Specifically, during the extraction process, the original data needs to be classified and filtered according to parameter type tags. For example, aircraft status parameters are separated by parsing the protocol identifier in the data packet, the second data subset is extracted using the dedicated coding rules for meteorological data, and the third data subset is obtained based on the data interface protocol of the airspace management system to ensure the integrity and independence of the data in each subset and provide structured input for subsequent anomaly detection.

[0017] Step S220: Dynamically screen the first data subset, the second data subset, and the third data subset respectively to obtain abnormal aircraft data in the first data subset that exceeds the preset aircraft status threshold, abnormal meteorological data in the second data subset that exceeds the preset meteorological fluctuation threshold, and abnormal airspace data in the third data subset that exceeds the preset airspace capacity threshold.

[0018] Exemplarily, the dynamic threshold screening mechanism compares each data subset item by item through preset aircraft state safety thresholds, meteorological fluctuation tolerance thresholds, and airspace capacity limit thresholds to identify abnormal data deviating from the normal range. The preset aircraft state thresholds include the maximum allowable deviation value of the flight speed (e.g., ±15% of the route planning speed), the minimum safety limit of the flight altitude (such as a vertical buffer distance of 100 meters from ground obstacles), the sudden deviation tolerance of the heading angle (such as an angle change exceeding 10 degrees within 3 consecutive seconds), and the critical warning value of the fuel reserve (such as the remaining fuel supporting a flight time of less than 15 minutes); the preset meteorological fluctuation thresholds cover the instantaneous mutation gradient of the wind speed (such as an increase in wind speed of 5 m / s within 10 seconds), the sudden drop rate of visibility (such as a decrease of 500 meters per minute), the fluctuation intensity of air pressure (such as a change in air pressure exceeding 2 hPa within 1 minute), and the danger level of turbulence intensity (such as moderate or above turbulence lasting for more than 30 seconds); the preset airspace capacity thresholds include the maximum aircraft density within a unit airspace grid (such as no more than 5 aircraft per cubic kilometer), the upper limit of the overlap degree between the temporary control area and the flight path (such as a path overlap ratio exceeding 60%), and the safety factor of the dispersion of obstacle distribution (such as the obstacle spacing being less than 1.5 times the minimum avoidance distance). During the screening process, if the flight speed in the first data subset exceeds the maximum allowable deviation value for 3 consecutive sampling periods, it is marked as abnormal aircraft data; if the mutation gradient of the wind speed in the second data subset exceeds the preset threshold between two adjacent acquisition time points, it is determined as abnormal meteorological data; if the aircraft density heat map in the third data subset shows that the number of aircraft in a certain airspace grid exceeds the capacity limit, abnormal airspace data is generated. Through layer-by-layer threshold comparison and multi-period data sliding window verification, the accuracy of abnormal data determination and the anti-noise interference ability are ensured.

[0019] Step S230: Input the abnormal aircraft data, abnormal meteorological data, and abnormal airspace data into the event type association model to determine the priority and association parameters of the emergency event type.

[0020] Exemplarily, the event type association model is a multi-modal fusion classification model trained based on a historical emergency event case library. By analyzing the similarity between the combined features of abnormal data and historical event patterns, it outputs the emergency event type and its disposal priority. The model inputs include abnormal aircraft data (such as a sequence of sudden drops in flight altitude), abnormal meteorological data (such as the mutation gradient of strong crosswind speed), and abnormal airspace data (such as the excessive dispersion of obstacle distribution). First, standardization processing is performed on the three types of abnormal data: aircraft state parameters are mapped to the [0,1] interval by maximum-minimum normalization, meteorological parameters eliminate dimensional differences through Z-score standardization, and airspace parameters use piecewise linear scaling to adapt to different airspace scales. Subsequently, the model converts the standardized abnormal data into feature vectors respectively: the aircraft feature vector consists of the speed deviation, altitude offset, and heading angle offset rate; the meteorological feature vector includes the wind speed mutation gradient, visibility decrease rate, and air pressure fluctuation intensity; the airspace feature vector integrates the airspace occupancy density, control area overlap degree, and obstacle dispersion. After generating a comprehensive abnormal feature matrix through multi-dimensional splicing, the model uses a hierarchical matching mechanism to calculate the similarity with the feature matrix in the historical case library (such as cosine similarity or dynamic time warping algorithm), selects the historical case with the highest matching degree, and extracts its event type identifier (such as "aircraft out of control - meteorological mutation composite event"), processing priority weight (such as the resource emergency dispatch level), and resource association parameters (such as the type and quantity of required rescue equipment). Further, the model dynamically corrects the priority weight in combination with the dynamic change trend of real-time data (such as the accelerating continuous decrease rate of flight altitude), for example, upgrading an event originally designated as secondary priority to primary priority, and adding the fuel supply requirement in the associated parameters, so as to generate an emergency event classification result closely adapted to the current situation.

[0021] As an implementation, in step S230, the abnormal aircraft data, abnormal meteorological data, and abnormal airspace data are input into the event type association model to determine the priority and association parameters of the emergency event type, which may specifically include: step S231: Perform data standardization processing on the abnormal aircraft data, abnormal meteorological data, and abnormal airspace data respectively to generate a first standardized data set corresponding to the abnormal aircraft data, a second standardized data set corresponding to the abnormal meteorological data, and a third standardized data set corresponding to the airspace data.

[0022] Data standardization aims to eliminate the differences in dimension and numerical range of different parameters, enabling multi-source heterogeneous data to be fused and analyzed and input into models under a unified scale. For abnormal aircraft data, in the standardization process, first, according to the normal range of preset aircraft state parameters (for example, the standard range of flight speed is 200 - 300 knots, the safe range of flight altitude is 500 - 1500 meters, and the stable fluctuation range of heading angle is ±5 degrees), the deviation of each parameter (i.e., the difference between the actual value and the standard value) is calculated, and the deviation is mapped to the [0,1] interval through the max-min normalization method to generate the first standardized data set. For example, the flight speed deviation is the difference between the actual speed and the planned speed of the flight route. If the actual speed is 320 knots and the planned speed is 280 knots, the deviation is +40 knots, which is converted to 0.8 after normalization (assuming the maximum allowable deviation is 50 knots); for abnormal meteorological data, the Z-score standardization method is used. Based on the mean and standard deviation of historical meteorological data, the wind speed mutation gradient (such as the wind speed increases by 5 m / s within 10 seconds), the visibility decrease rate (such as decreases by 500 meters per minute), and the air pressure fluctuation intensity (such as changes by 2 hPa within 1 minute) are converted into standardized values that conform to the normal distribution to generate the second standardized data set; for abnormal airspace data, the airspace occupancy density (such as the number of aircraft per square kilometer is 8, exceeding the preset threshold of 5), the overlap degree of temporary control areas (such as the overlap ratio of the flight path and the control area reaches 70%), and the obstacle distribution dispersion (such as the distance between obstacles is less than 1.2 times the safe distance) are adjusted to a unified ratio according to the airspace scale and geographical coordinate range through the piecewise linear scaling method to generate the third standardized data set. The standardized data set has comparability and model input compatibility, providing a basis for subsequent feature vector construction.

[0023] Step S232: Generate an aircraft feature vector containing the flight speed deviation, altitude offset, and heading angle offset rate according to the types of aircraft state parameters in the first standardized data set; generate a meteorological feature vector containing the wind speed mutation gradient, visibility decrease rate, and air pressure fluctuation intensity according to the types of meteorological parameters in the second standardized data set; generate an airspace feature vector containing the airspace occupancy density, overlap degree of temporary control areas, and obstacle distribution dispersion according to the types of airspace parameters in the third standardized data set.

[0024] Feature vector construction is a crucial step in converting the standardized parameter sequence into a structured feature representation. The aircraft feature vector is composed of the flight speed deviation (such as 0.8 after normalization), altitude offset (such as 0.6 after normalizing the difference between the current altitude and the safety altitude), and heading angle deviation rate (such as 0.7 after standardizing the change rate of the heading angle within 3 consecutive seconds) arranged in a fixed order, forming a vector with a dimension of 3; the meteorological feature vector integrates the wind speed mutation gradient (such as 1.2 after Z-score standardization), visibility decrease rate (after standardization of -0.5), and air pressure fluctuation intensity (after standardization of 0.9), forming a vector with a dimension of 3; the airspace feature vector includes the airspace occupancy density (such as 0.75 after scaling), temporary control area overlap degree (after scaling of 0.88), and obstacle distribution dispersion degree (after scaling of 0.63), also constituting a vector with a dimension of 3. Specifically, each element in the aircraft feature vector represents the abnormal degree of a specific aircraft state parameter, the meteorological feature vector reflects the impact intensity of meteorological mutations on flight safety, and the airspace feature vector quantifies the severity of airspace resource conflicts. For example, when the aircraft feature vector is [0.8, 0.6, 0.7], the meteorological feature vector is [1.2, -0.5, 0.9], and the airspace feature vector is [0.75, 0.88, 0.63], it can respectively represent the abnormal scenarios of the aircraft having too high a speed, too low an altitude, and unstable heading, accompanied by strong wind speed mutations and low visibility, and crowded airspace and high overlap of control areas.

[0025] Step S233: Multidimensionally splice the aircraft feature vector, meteorological feature vector, and airspace feature vector to generate a comprehensive abnormal feature matrix, and input the comprehensive abnormal feature matrix into the pre-trained event type association model; among them, the event type association model is trained based on the mapping relationship between the same type of feature vectors extracted from historical emergency event cases and the event type labels.

[0026] Multi-dimensional splicing constructs a comprehensive abnormal feature matrix with dimensions of 3×3 or 9×1 by stacking three feature vectors row by row or column by column. For example, if the aircraft feature vector is [0.8, 0.6, 0.7], the meteorological feature vector is [1.2, -0.5, 0.9], and the airspace feature vector is [0.75, 0.88, 0.63], after splicing row by row, a 3×3 matrix [[0.8, 0.6, 0.7], [1.2, -0.5, 0.9], [0.75, 0.88, 0.63]] is generated, or after splicing column by column, a 9×1 vector [0.8, 0.6, 0.7, 1.2, -0.5, 0.9, 0.75, 0.88, 0.63] is obtained. After this matrix is input into the pre-trained event type association model, the model, based on the mapping relationship between the same type of feature matrix and the event type label (such as "aircraft out of control - strong crosswind composite event") in the historical emergency event case library, performs pattern matching through a deep neural network or an ensemble learning algorithm. Each case in the historical case library contains a feature matrix, an event type identifier (such as the unique code E-1024), a processing priority weight (such as the first-level priority corresponding to emergency resource scheduling), and resource association parameters (such as the required number of fire-fighting drones being 2). The model establishes a mapping rule from the feature matrix to the event type and parameters through supervised learning to ensure that the current comprehensive abnormal feature matrix can trigger a classification logic similar to that of historical cases.

[0027] Step S234: Through the hierarchical matching mechanism in the event type association model, compare the similarity between the comprehensive abnormal feature matrix and the feature matrix in the historical emergency event cases, determine at least one historical case with the highest matching degree to the comprehensive abnormal feature matrix, and extract the event type identifier, processing priority weight, and resource association parameters recorded in the historical case.

[0028] The hierarchical matching mechanism adopts a multi-stage similarity calculation strategy: First, in the first layer, the top K (e.g., K = 10) candidate historical cases are quickly screened out through Euclidean distance or cosine similarity; in the second layer, the dynamic time warping (DTW) algorithm or Manhattan distance is used for refined comparison, considering the time series change trend of the feature matrix (such as the continuous increase in the flight speed deviation); finally, in the third layer, the context compatibility of the candidate cases is verified through an expert rule engine (such as whether the current airspace allows the scheduling of similar resources). For example, if the cosine similarity between the current comprehensive anomaly feature matrix and the feature matrix of historical case C-205 is 0.92 (the highest), and the similarity with C-198 is 0.85, then the event type identifier "Aircraft Fuel Warning - Airspace Congestion Composite Event", the processing priority weight "Level 1", and the resource association parameters "2 fuel supply drones, 1 set of temporary navigation beacons" of C-205 are preferentially extracted. During the matching process, the model simultaneously calculates the similarity confidence (e.g., 0.92 corresponds to a confidence of 95%), and only when the confidence exceeds the preset threshold (e.g., 80%), the match is determined to be valid.

[0029] Step S235: Determine the basic classification label of the current emergency event type according to the event type identifier of the historical case with the highest matching degree; according to the processing priority weight and resource association parameters, combined with the real-time change trend of the aircraft feature vector in the comprehensive anomaly feature matrix, dynamically adjust the priority level and the combination of association parameters corresponding to the basic classification label.

[0030] The basic classification label is inherited from the event type identifier of the matching historical case, such as "Aircraft Out of Control - Strong Crosswind Composite Event". The initial value of the priority level is the processing priority weight in the historical case (e.g., Level 2), but it needs to be dynamically adjusted according to the parameter change trend in the real-time aircraft feature vector: If the flight speed deviation continues to increase in the next 3 sampling periods (e.g., from 0.8 to 0.9), the priority is upgraded from Level 2 to Level 1; for the "number of fire drones" in the resource association parameters, if the historical case is 2 drones, but the current obstacle distribution has a higher dispersion (e.g., from 0.63 to 0.75), then 1 additional drone is added to cope with the complex airspace environment. The dynamic adjustment is achieved through a preset weight correction coefficient. For example, the priority upgrade coefficient is 0.2 (the coefficient accumulates for every 0.1 increase in the deviation per period), and the resource quantity correction formula is "base quantity × (1 + dispersion increment / 0.1)".

[0031] Step S236: Based on the adjusted priority level and the combination of association parameters, generate a structured output result including the emergency event type, priority ranking, and resource association rules, and dynamically bind the structured output result to the invocation conditions of the flight emergency prediction model.

[0032] The structured output result is encapsulated in JSON or XML format, including fields "Event Type: Aircraft Out-of-Control - Strong Crosswind Composite Event", "Priority: Level 1", "Resource Requirements: 3 fire-fighting drones, 1 set of temporary navigation beacons", "Processing Time Window: 5 minutes". The dynamic binding process verifies whether the invocation conditions of the flight emergency prediction model meet the resource type (such as whether the model supports "fire-fighting drone scheduling"), time window (such as whether the model can generate a strategy within 5 minutes), and airspace constraints (such as whether the model is compatible with the current airspace control rules). For example, if the invocation conditions of model M-1024 include "Resource Type = fire-fighting drone, Time Window ≤ 5 minutes", then bind this model to the structured output result; if model M-1025 only supports "Time Window ≥ 10 minutes", then exclude this model. The binding result generates a model invocation queue to ensure that only eligible prediction models are triggered in subsequent steps.

[0033] Step S240: Generate an event feature label containing the emergency event type according to the priority and associated parameters, and match the event feature label with the invocation conditions of the flight emergency prediction model.

[0034] The event feature label is a structured data object used to encapsulate the core attributes of an emergency event type and its disposal constraints. The priority field reflects the urgency of event handling. For example, a first-level priority indicates an out-of-control aircraft event that requires immediate response, and a second-level priority corresponds to a meteorological interference event that can be postponed for disposal. The associated parameter field includes resource type requirements (such as fire-fighting drones, temporary navigation beacons), processing time window (such as the need to complete path adjustment within 5 minutes), and airspace impact range (such as a controlled area with a radius of 2 kilometers). When generating the label, first parse the resource type requirement parameter in the associated parameters, refine the resource identifier into specific device models and deployment locations (such as "Model A fire-fighting drones need to be deployed to coordinates X, Y"), and sort the resource priorities according to the emergency resource scheduling rules (such as giving priority to scheduling the nearest available device). At the same time, based on the time window limit in the processing time parameter (such as generating a command instruction within 30 seconds), expand the time limit constraint identifier into a time sensitivity classification (such as high sensitivity requires real-time feedback) and a time superposition effect (such as cumulative delay control for parallel execution of multiple instructions). Further, combine the obstacle distribution density in the airspace impact range parameter to map the impact range identifier into a spatial weight distribution matrix (such as a higher avoidance weight needs to be assigned to areas with high-density obstacles) and a regional linkage relationship topology (such as the cooperative activation conditions for adjacent airspace control instructions). Finally, generate a composite event feature label containing resource allocation logic, time constraint logic, and spatial coverage logic through logical splicing. For example, "First-level priority - Aircraft fuel warning event: 2 fuel replenishment drones need to be dispatched to coordinates P, Q within 3 minutes, and other aircraft are prohibited from entering the airspace between altitudes H1 and H2". When matching this label with the invocation conditions of the flight emergency prediction model, it is necessary to verify the consistency between the types of model input parameters and the resource types, time windows, and spatial ranges in the label. For example, only invoke a prediction model that supports the "fuel replenishment" resource type and is compatible with the "3-minute response" time constraint to ensure the accurate adaptation of the model output strategy to the event disposal requirements.

[0035] As an implementation manner, in step S240, according to the priority and associated parameters, generate an event feature label including the emergency event type, which may specifically include: Step S241: Based on the grading result in the priority, parse the resource type requirement parameter, processing time parameter, and airspace impact range parameter in the associated parameters, and generate an initial label component corresponding to the emergency event type; the initial label component includes a resource type identifier, a time limit constraint identifier, and an impact range identifier.

[0036] The construction of the initial label component is indexed by the priority level. By parsing the resource type requirement parameters (such as "fire drones, fuel supply equipment"), processing time limit parameters (such as "the path adjustment needs to be completed within 5 minutes"), and airspace impact range parameters (such as "a control area with a radius of 2 kilometers") in the associated parameters, basic label elements are generated. The resource type identifier is a standardized encoded string. For example, "RES-FIRE-DRONE-001" represents a fire drone with model number 001; the time limit constraint identifier adopts a combination of a timestamp interval and a time sensitivity marker. For example, "T_WINDOW: [T0, T0 + 300 seconds], SENSITIVITY: HIGH" represents a 5-minute response window with high sensitivity; the impact range identifier defines the spatial range through geofence coordinates and vertical altitude limits. For example, "GEO_ZONE: POLYGON((x1,y1), (x2,y2),...), ALTITUDE: [100 meters, 500 meters]" represents an airspace area with a horizontal boundary of polygon coordinates and a vertical altitude limit of 100 meters to 500 meters. Specifically, during the parsing process, the device names and models in the resource type requirement parameters need to be mapped to a predefined resource coding library to ensure the uniqueness of the identifier; the processing time limit parameters need to be converted into an absolute timestamp sequence and synchronized with the system clock; the airspace impact range parameters are converted into a standardized spatial description format through a Geographic Information System (GIS) interface. The initial label component is stored in the form of a tuple or key-value pair. For example, "{resource: RES-FIRE-DRONE-001, time limit: T_WINDOW, space: GEO_ZONE}", providing a structured input for subsequent refinement and expansion.

[0037] Step S242: According to the emergency resource scheduling rules in the resource type requirement parameters, refine the resource type identifier in the initial label component to generate an updated resource type identifier that includes resource priority sorting and resource conflict avoidance strategies.

[0038] The refinement process expands the deployment priority, scheduling path, and conflict avoidance conditions of resource type identifiers according to the strategies in the emergency resource scheduling rule library. For example, the initial resource type identifier "RES-FIRE-DRONE-001" is refined into "RES-FIRE-DRONE-001-P1 (priority 1, deployed to coordinate P1, closest to the current airspace)" and "RES-FIRE-DRONE-001-P2 (priority 2, deployed to coordinate P2, backup node)" through the "distance priority" strategy in the rule library. At the same time, if the resource type requirement parameters contain conflicting devices (such as simultaneously requiring the use of a fire-fighting drone and a meteorological monitoring drone), the conflict avoidance strategy "ALT_AVOID: H_LAYER≠300 meters" is added to prohibit the two types of devices from operating at the same altitude layer. The refined resource type identifiers form executable resource scheduling instructions by appending priority tags, deployment coordinates, and conflict rules. For example, "the resource type identifier is updated to: {RES-FIRE-DRONE-001-P1@P1, RES-WEATHER-DRONE-005-P2@P2, AVOID_H_LAYER=300 meters}", ensuring the compatibility of the resource scheduling logic with airspace control rules.

[0039] Step S243: Based on the time window limit condition in the processing time parameter, dynamically expand the time constraint identifier in the initial label component to generate an updated time constraint identifier that includes time sensitivity grading and time effect superposition.

[0040] Dynamic expansion upgrades the simple time window of the initial aging constraint identifier to a multi-dimensional time logic description by introducing a time sensitivity grading mechanism and an aging superposition effect model. For example, the initial identifier "T_WINDOW: [T0, T0 + 300 seconds]" is expanded to "T_SENSITIVITY: HIGH (red alert, real-time monitoring), T_SUB_TASKS: {Path planning: [T0, T0 + 60 seconds], Resource scheduling: [T0 + 60 seconds, T0 + 200 seconds], Instruction issuance: [T0 + 200 seconds, T0 + 300 seconds]}, T_OVERLAP_TOLERANCE: ≤10 seconds (allowing sub-task superposition delay)". The time sensitivity grading is automatically associated according to the event priority. For example, a first-level priority event corresponds to "HIGH" sensitivity, requiring real-time feedback. The aging superposition effect ensures that the overall response time meets the window constraint by calculating the cumulative delay upper limit when multiple tasks are executed in parallel (e.g., the total delay does not exceed 10 seconds). The expanded aging constraint identifier is stored through a time tree structure or hierarchical key-value pairs, such as "{Main window: T0 - T0 + 300 seconds, Sub-tasks: [Planning, Scheduling, Issuance], Superposition tolerance: 10 seconds}", providing precise control logic for time-driven instruction execution.

[0041] Step S244: Combine the regional overlap degree data and obstacle distribution density in the airspace influence range parameters to perform multi-dimensional mapping on the influence range identifier in the initial label component, generating an updated influence range identifier that includes spatial weight allocation and regional linkage relationships.

[0042] Multi-dimensional mapping upgrades the static geographical description of the initial influence range identifier to a dynamic spatial strategy through a spatial weight allocation algorithm and a regional linkage rule engine. For example, the initial identifier "GEO_ZONE: POLYGON((x1,y1),...), ALTITUDE: [100 meters, 500 meters]" combined with the obstacle distribution density data (such as the obstacle density in area A is 0.8 per square kilometer) is mapped to "SPACE_WEIGHT: ZONE_A = 0.8 (high avoidance priority), ZONE_B = 0.3 (low priority)"; at the same time, if there is an airspace control linkage relationship between area A and adjacent area C (such as area C enters the monitoring state when area A is activated), then append the linkage rule "LINKED_ZONES: ZONE_A → ZONE_C (Monitoring mode: preparatory control)". The updated influence range identifier forms a dynamic avoidance strategy and a regional coordination mechanism through a spatial weight matrix and a linkage relationship table, such as "Spatial identifier updated to: {Weight: ZONE_A = 0.8, ZONE_B = 0.3, Linkage: ZONE_A → ZONE_C}", providing a decision-making basis for the adaptive allocation of airspace resources.

[0043] Step S245: logically concatenate the updated resource type identifier, the updated time constraint identifier, and the updated impact range identifier to form a composite event feature label including resource allocation logic, time constraint logic, and space coverage logic.

[0044] Logical concatenation integrates the detailed parameters of the three types of identifiers into a unified event feature label through logical operators and conditional expressions. For example, the resource type identifier "RES-FIRE-DRONE-001-P1@P1" and the conflict avoidance rule "AVOID_H_LAYER=300 meters", the time constraint identifier "T_SENSITIVITY: HIGH" and the space identifier "SPACE_WEIGHT: ZONE_A=0.8" are combined through the logical AND relationship to form a compound logical statement: "IF resource = RES-FIRE-DRONE-001-P1@P1 AND time sensitivity = HIGH AND space weight ≥ 0.8 THEN execute avoidance strategy: ALTITUDE ≠ 300 meters". Composite event feature tags are stored in structured text or XML / JSON format, for example, "{resource allocation logic: schedule RES-FIRE-DRONE-001-P1 to P1 and avoid the 300-meter altitude layer, time constraint logic: complete within T0+300 seconds and the subtask superposition delay is ≤10 seconds, spatial coverage logic: give priority to avoiding areas with weight ≥0.8 and link monitoring area C}", to ensure the independence and parsability of each logical unit.

[0045] Step S246: According to the compatibility verification results of each logical unit in the composite event feature label, the conflicting part of the internal parameters of the label is adjusted, and the unique identification code of the emergency event type is injected to generate an event feature label that matches the flight emergency prediction model calling conditions.

[0046] Compatibility verification detects conflicts in resource, time, and space logic through a rules engine. For example, if there is a coordinate overlap between the resource scheduling path "RES-FIRE-DRONE-001-P1@P1" and the spatial weight area "ZONE_A = 0.8" and aircraft are prohibited from entering this area, a conflict alert is triggered, and the resource deployment coordinates are automatically adjusted to "P1'" (outside ZONE_A); at the same time, if there is a conflict between the time window "T0 + 300 seconds" and the resource arrival time "T0 + 250 seconds" (such as the resource not arriving on time), the time window is dynamically extended to "T0 + 350 seconds". After conflict resolution, a unique identification code (such as "EVENT-ID-20231105001") is injected into the label to generate the final event feature label: "EVENT-ID-20231105001: Resource = RES-FIRE-DRONE-001-P1@P1', Time = T0 + 350 seconds, Space = Outside ZONE_A and weight ≥ 0.5, Compatibility status = Verified". The unique identification code is created through a hashing algorithm or a serial number generator to ensure global uniqueness and traceability.

[0047] Step S247: Establish a mapping relationship between the emergency event type and the input parameters of the target flight emergency prediction model through the resource allocation logic, time constraint logic, and spatial coverage logic in the event feature label to ensure parameter consistency during the model call phase.

[0048] The mapping relationship is implemented through a parameter matching table, which maps the logical units in the event feature label to the model input parameters one by one. For example, "RES-FIRE-DRONE-001-P1@P1'" in the resource allocation logic is mapped to the model input parameter "resource_type = fire_drone, deploy_coord = P1'"; "T0 + 350 seconds" in the time constraint logic is mapped to "time_limit = 350"; "Outside ZONE_A and weight ≥ 0.5" in the spatial coverage logic is converted to "avoid_zones = ZONE_A, priority_weight = 0.5". If the input conditions of the target flight emergency prediction model M-1024 require "resource_type = fire_drone, time_limit ≤ 400, avoid_zones ≠ NULL", the parameters of the event feature label match it exactly, triggering the model call. During the mapping process, the consistency of parameter types, value ranges, and logical constraints needs to be verified. For example, whether the time limit is an integer value and whether the coordinates of the avoidance area are in a valid geographical format to ensure the legality of the model input and the reliability of the prediction results.

[0049] Step S300: Invoke a flight emergency prediction model corresponding to the type of emergency event; the flight emergency prediction model is trained based on flight features and airspace features screened from historical emergency event data, and the types of input parameters and output parameters of each flight emergency prediction model are consistent in the training stage and the application stage.

[0050] Exemplarily, the flight emergency prediction model is a machine learning model customized for a specific type of emergency event, and its training data is sourced from the flight feature set and airspace feature set after feature screening in historical emergency event cases. The flight feature set includes parameters directly related to the flight path such as the speed adjustment record, altitude correction trajectory, and heading angle recovery rate of the aircraft in historical events; the airspace feature set contains the dynamic change data of airspace capacity during the handling of historical events, the obstacle avoidance path planning record, and the temporary control area adjustment strategy. In the model training stage, redundant features (such as irrelevant environmental noise data) with a correlation degree lower than the preset threshold with the emergency handling result are removed through redundancy analysis, and the core flight features and core airspace features are retained as input variables. At the same time, the flight path adjustment strategy and emergency resource allocation strategy verified to be effective in historical events are used as output variables to ensure the structural consistency of the model input and output in the training and application stages. For example, for the event type of "avoidance of low-altitude thunderstorm weather", the model inputs are the time-series data of the wind speed in the thunderstorm area, the current altitude of the aircraft, and the available avoidance paths in the airspace, and the outputs are the climb altitude instruction, the coordinates of the bypass path, and the adjustment parameters of the weather radar scanning frequency. When invoking the model, the target model with the highest response weight and the smallest error range is screened from the candidate model pool according to the event feature label, and the real-time abnormal data is reorganized into the input data set in the standardized format of the training stage to trigger the model to perform path simulation and resource demand prediction.

[0051] As an implementation manner, step S300, invoking a flight emergency prediction model corresponding to the type of emergency event, may specifically include: step S310: Obtain multiple candidate flight emergency prediction models associated with the event feature label; each candidate flight emergency prediction model corresponds to a different emergency event handling stage, and the type of input parameters used by each candidate flight emergency prediction model in the training stage is consistent with the parameter type in the current real-time flight monitoring data.

[0052] The candidate flight emergency prediction model is screened and generated from a predefined model pool through the resource allocation logic, time constraint logic, and space coverage logic in the event feature tags. For example, if the event feature tags contain "resource type identifier = fire drone model A-001", "time constraint identifier = 5-minute response window", and "space coverage identifier = avoidance area ZONE_A", then all models in the model pool that support the dispatching of fire drones, the 5-minute time window, and the avoidance rules for the ZONE_A area are listed as candidates. Each candidate model corresponds to a specific emergency response stage. For example, model M-1024 is dedicated to path planning during the emergency avoidance stage of the flight path, and the input parameters include the aircraft speed, altitude, heading angle, and obstacle coordinates; model M-1025 is dedicated to demand prediction during the resource scheduling stage, and the input parameters cover the resource type, deployment coordinates, and time window. The types of input parameters of the candidate model must be strictly consistent with the parameters in the current real-time flight monitoring data. For example, if the real-time data contains the aircraft speed deviation, wind speed mutation gradient, and airspace occupancy density, then the training data of the candidate model must contain the same type of parameters to ensure the structural compatibility of input and output. By matching the event feature tags with the model metadata, such as comparing the "resource type identifier" in the tags with the "supported resource type" field in the model metadata, all eligible candidate models are screened out to form an initial candidate queue.

[0053] Step S320: Determine the response weight and error range of each candidate flight emergency prediction model according to the historical verification results of the candidate flight emergency prediction model.

[0054] The historical verification results are obtained by evaluating the performance of the model in historical emergency event cases, including indicators such as response speed, prediction accuracy, and resource scheduling success rate. The response weight reflects the priority of the model in similar events. For example, if model M-1024 successfully generated an effective path strategy 95 times in 100 historical verifications, then its response weight is 0.95; the error range is calculated based on the deviation between the prediction result and the actual disposal result. For example, if the average absolute error between the predicted resource demand quantity and the actual usage quantity of model M-1025 is 8%, then its error range is 8%. Specifically, the response weight is calculated using the weighted average method, considering that the performance weight of the model in the recent period is higher (such as the verification results in the last 30 days account for 70% of the weight); the error range is quantified by the root mean square error (RMSE) or the mean absolute percentage error (MAPE). For example, the historical response weight of model M-1024 is 0.92 (9 successes in the recent 10 verifications), and the error range is 5%; the response weight of model M-1025 is 0.85, and the error range is 12%. The weight and error data are stored in the model performance database and are retrieved through the real-time query interface and updated to the candidate model queue.

[0055] Step S330: Screen out at least one target flight emergency prediction model from the candidate flight emergency prediction models, where the response weight is higher than the preset weight threshold and the error range is lower than the preset error threshold.

[0056] The preset weight threshold and error threshold are dynamically adjusted according to the safety level of the emergency event type. For example, for a first-level priority event, the response weight is required to be ≥0.9 and the error range is ≤10%. For a second-level priority event, it is relaxed to a weight ≥0.8 and an error ≤15%. During the screening process, the candidate models are sorted in descending order of response weight, and the models with excessive errors are excluded. For example, the candidate queue contains models M-1024 (weight 0.92, error 5%), M-1025 (weight 0.85, error 12%), and M-1026 (weight 0.88, error 9%). If the preset threshold is a weight ≥0.85 and an error ≤10%, then M-1025 is excluded due to an error of 12%, and the final target models are M-1024 and M-1026. If multiple models meet the conditions, a single model or a combination of models is selected according to the scenario requirements. For example, a high-complexity event triggers multi-model collaborative prediction (such as M-1024 is responsible for path planning and M-1026 is responsible for resource conflict detection), while a simple event only invokes a single model. The screening results of the target models form the final call list and are injected into the event processing pipeline.

[0057] Step S340: Reorganize the abnormal aircraft data, abnormal meteorological data, and abnormal airspace data according to the input format of the target flight emergency prediction model during the training phase to generate a standardized input data set.

[0058] The data reorganization must strictly follow the training data format of the target model, including the parameter order, data dimension, and standardization rules. For example, the input format of model M-1024 requires a four-dimensional vector of [flight speed deviation, altitude offset, heading angle deviation rate, obstacle distribution dispersion]. If the flight speed deviation in the real-time abnormal data is 0.8, the altitude offset is 0.6, the heading angle deviation rate is 0.7, and the obstacle distribution dispersion is 0.75, then it is reorganized into the vector [0.8, 0.6, 0.7, 0.75]. If the model input contains time series data (such as the wind speed mutation gradient in the past 5 sampling periods), then the real-time abnormal meteorological data needs to be slid and aligned according to the time window to generate a time series matrix. For example, the wind speed mutation gradient sequence in the abnormal meteorological data is [0.5, 0.7, 0.9, 1.1, 1.3], and it is reorganized into [[0.5, 0.7, 0.9], [0.7, 0.9, 1.1], [0.9, 1.1, 1.3]] according to the 3-order time window required by the model. During the data reorganization process, the matching of parameter types and dimensions needs to be verified, such as whether the airspace occupancy density parameter is a floating-point number and whether the time series length meets the model requirements, to ensure the integrity and legality of the input data.

[0059] Step S350: Input the standardized input data set into the target flight emergency prediction model, trigger the target flight emergency prediction model to perform flight path simulation and resource demand prediction, and obtain a flight path adjustment strategy and an emergency resource allocation strategy.

[0060] The target model executes the prediction logic based on the standardized input data set. For example, model M-1024 generates a heading correction angle, a target altitude layer, and a speed regulation curve through neural network inference, forming a flight path adjustment strategy of "heading +15 degrees, climb to 600 meters, speed reduce to 250 knots"; model M-1026 predicts the required number of fire drones, deployment coordinates, and supply cycles through regression analysis, generating an emergency resource allocation strategy of "dispatch 2 drones to coordinates (X1, Y1), supply once every 30 minutes". During the prediction process, the built-in conflict detection module in the model checks the spatio-temporal compatibility of the path and resources. For example, it detects whether the heading correction angle conflicts with the drone deployment area. If there is a conflict, it triggers the iterative optimization of the strategy until a conflict-free solution is generated. The finally output strategy is encapsulated into a structured instruction set. For example, the flight path adjustment strategy includes a waypoint sequence [WPT1, WPT2, WPT3], an altitude instruction [HOLD 600 meters], and a speed instruction [MAINTAIN 250 knots]; the emergency resource allocation strategy includes a resource scheduling queue [DRONE-001 → (X1, Y1), DRONE-002 → (X2, Y2)] and a time trigger [T0 + 1800 seconds for supply]. The strategy data is encoded through a standardized protocol (such as JSON) and transmitted to the instruction execution module to complete the closed-loop from model prediction to actual control.

[0061] As an implementation, the training process of the flight emergency prediction model includes the following steps: Step S10: Obtain a historical emergency event data set; the historical emergency event data set includes multiple processed emergency event cases, and each emergency event case contains aircraft parameters, meteorological parameters, airspace parameters at the time of the emergency event, as well as path adjustment records and resource allocation records after the emergency event is processed.

[0062] The historical emergency event dataset is constructed by integrating the archived data of the air traffic control system, meteorological monitoring database, and airspace management platform. Each case covers the complete data of the emergency event handling cycle. The aircraft parameters include the real-time flight speed, altitude, heading angle, attitude angle, fuel reserve, and navigation warning status at the time of the event. For example, in a case of "aircraft fuel warning event", the aircraft parameters are recorded as "speed 280 knots, altitude 600 meters, heading angle deviation +12 degrees, fuel reserve 10%"; the meteorological parameters include the time series data of wind speed, wind direction, visibility, air pressure, and turbulence intensity in the event area. For example, "the wind speed mutation gradient is 5 m / s², and the visibility drops to 500 meters within 30 minutes"; the airspace parameters record the airspace occupancy density, obstacle distribution coordinates, and the boundary of the temporary control area. For example, "the obstacle distribution dispersion is 0.8 (the high-density area accounts for 80%), and the overlap degree between the control area and the flight path is 70%"; the path adjustment record details the actual executed heading correction angle, climb / descent command, and speed regulation curve. For example, "heading +15 degrees correction, climb to 800 meters, speed reduce to 250 knots"; the resource allocation record includes the types of dispatched equipment (such as fire drone model A-001), deployment coordinates, supply cycle, and execution effect score (such as the deviation of resource arrival time ≤ 30 seconds). The dataset is aligned by timestamp and airspace coordinates to ensure the spatio-temporal consistency of multi-source data and provide high-quality input for feature extraction and model training.

[0063] Step S20: Extract the flight feature set related to flight speed, altitude deviation, and heading angle offset from the aircraft parameters of each emergency event case, and extract the airspace feature set related to airspace capacity, obstacle distribution, and temporary control area from the airspace parameters.

[0064] The extraction of the flight feature set focuses on the key indicators of the aircraft's dynamic behavior: the flight speed feature is characterized by calculating the deviation amount between the actual speed at the time of the event and the planned route speed (such as +20 knots); the altitude deviation is the difference between the actual altitude and the safe altitude (such as -100 meters); the heading angle offset is quantified by statistically calculating the standard deviation or the maximum instantaneous offset rate of the heading angle during the event duration (such as an offset of +10 degrees within 5 seconds). For example, in a certain case, the flight feature set is "speed deviation +20 knots, altitude deviation -100 meters, heading angle offset rate 0.8"; the airspace feature set is extracted from the airspace parameters including airspace capacity (such as the number of aircraft in a unit grid is 6, exceeding the threshold of 5), obstacle distribution (such as the area where the obstacle spacing is less than 1.2 times the safe distance accounts for 60%), and the overlap degree between the temporary control area and the flight path (such as the coordinate overlap ratio is 75%). The feature extraction process removes noise data (such as sensor instantaneous outliers) through data cleaning rules and uses the sliding window statistical method to generate time series features (such as the average altitude deviation in the past 5 minutes) to ensure that the feature set comprehensively reflects the event situation.

[0065] Step S30: Perform redundancy analysis on the flight feature set and the airspace feature set, remove redundant features with a correlation degree lower than a preset correlation threshold with the emergency event processing result, and obtain an optimized target flight feature set and target airspace feature set.

[0066] The redundancy analysis calculates the statistical correlation degree between the feature and the emergency event processing result (such as the path adjustment effect score, resource scheduling success rate) to screen out the core features. The Pearson correlation coefficient is used to quantify the linear correlation. For example, the correlation coefficient between the flight speed deviation and the path adjustment effect is 0.85 (strong positive correlation), while the correlation coefficient between the fuel reserve and the resource scheduling success rate is 0.1 (weak correlation); the mutual information method evaluates the non-linear correlation. For example, the mutual information value between the obstacle distribution dispersion and the complexity of the avoidance path is 0.7 (high correlation). The preset correlation threshold is 0.3 (features below this threshold are regarded as redundant). For example, the correlation coefficient between a certain airspace feature "historical usage frequency of the temporary control area" and the processing result is 0.2, so it is removed. Further, through the analysis of the influence of feature combinations, the synergy effect of the remaining features is verified. For example, the combined contribution degree of the "altitude deviation + airspace capacity" combination to the path adjustment effect is 0.9, which is higher than the contribution degree of a single feature. Finally, the target flight feature set retains "speed deviation, altitude deviation, heading angle offset rate", and the target airspace feature set retains "airspace capacity, obstacle distribution dispersion, temporary control area overlap degree" to form a concise and highly interpretable feature set.

[0067] As an implementation manner, in step S30, performing redundancy analysis on the flight feature set and the airspace feature set, removing redundant features with a correlation degree lower than a preset correlation threshold with the emergency event processing result, and obtaining an optimized target flight feature set and target airspace feature set may specifically include: step S31: Calculate the first correlation coefficient between each flight feature in the flight feature set and the emergency event processing result, and the second correlation coefficient between each airspace feature in the airspace feature set and the emergency event processing result.

[0068] The first correlation coefficient and the second correlation coefficient quantify the influence degree of flight characteristics and airspace characteristics on the emergency event handling results through statistical analysis methods. The first correlation coefficient adopts the Pearson correlation coefficient or the Spearman rank correlation coefficient to measure the linear or non-linear association between flight characteristics (such as flight speed deviation amount, altitude deviation, heading angle deviation rate) and handling results (such as path adjustment effect score, resource scheduling success rate). For example, the Pearson correlation coefficient between the flight speed deviation amount and the path adjustment effect is calculated to be 0.85, indicating a strong positive correlation between the two; while the Spearman rank correlation coefficient between the fuel reserve and the resource scheduling success rate is 0.1, showing a weak correlation or no significant association. The second correlation coefficient is applied to the relationship between airspace characteristics (such as airspace capacity, obstacle distribution dispersion, temporary control area overlap degree) and handling results in the same way. For example, the correlation coefficient between the obstacle distribution dispersion and the avoidance path complexity is 0.7, indicating that a higher dispersion area requires a more complex avoidance strategy. The correlation coefficient calculation is based on the eigenvalue and actual handling result records in the historical emergency event dataset, and a complete association matrix is generated by traversing all feature-result pairs, providing a quantitative basis for subsequent feature screening.

[0069] Step S32: Remove the target flight characteristics with the first correlation coefficient lower than the first correlation coefficient threshold from the flight characteristics set, and remove the target airspace characteristics with the second correlation coefficient lower than the second correlation coefficient threshold from the airspace characteristics set.

[0070] The preset first correlation coefficient threshold and the second correlation coefficient threshold are set according to the safety tolerance requirements of the emergency event type. For example, the first correlation coefficient threshold is 0.3 (removing weak correlation features with the absolute value of the correlation coefficient < 0.3), and the second correlation coefficient threshold is 0.25. The removal process traverses the association matrix, marks the features below the threshold as redundant and performs the deletion operation. For example, "fuel reserve" in the flight characteristics set is removed because the first correlation coefficient is 0.1 (< 0.3); "historical usage frequency of the temporary control area" in the airspace characteristics set is removed because the second correlation coefficient is 0.2 (< 0.25). The removal operation needs to verify the feature independence to avoid misdeleting features with potential synergistic effects. For example, although the individual correlation coefficients of "heading angle deviation rate" and "altitude deviation" are relatively low, their combination may have a significant impact on the handling result. In this case, the deletion needs to be postponed and enter the combined analysis stage.

[0071] Step S33: Conduct a combined analysis on the remaining flight characteristics and the remaining airspace characteristics after removal to determine the influence weights of different feature combinations on the handling results.

[0072] Combined analysis uses a multiple regression model or a random forest algorithm to evaluate the combined contribution of feature combinations to the processing results. For example, after combining the "flight speed deviation" and "altitude deviation" in the remaining flight features, the regression coefficient of their impact on the path adjustment effect calculated by the multiple regression model is 0.65 (the individual features are 0.45 and 0.40 respectively), indicating that the combined features have a higher influence weight; after combining the "airspace capacity" and "obstacle distribution dispersion" in the remaining airspace features, the feature importance score output by the random forest algorithm is 0.8 (the individual features are 0.5 and 0.6 respectively), reflecting a significant synergistic effect of the combined features. Combined analysis generates a feature combination weight matrix. For example, the weight of "flight speed deviation + airspace capacity" is 0.75, and the weight of "altitude deviation + obstacle distribution dispersion" is 0.68, providing data support for screening core features.

[0073] Step S34: According to the influence weights, screen out the top N core flight features from the remaining flight features and the top M core airspace features from the remaining airspace features.

[0074] The weight ranking is based on the results of combined analysis, and the truncation thresholds for the top N and top M are set (e.g., N = 3, M = 2). For example, among the remaining flight features, the weight of "flight speed deviation" is 0.75, the weight of "altitude deviation" is 0.68, and the weight of "heading angle offset rate" is 0.62. The top three are retained; among the remaining airspace features, the weight of "airspace capacity" is 0.80 and the weight of "obstacle distribution dispersion" is 0.73. The top two are retained. During the screening process, the significance of the weight difference needs to be verified. For example, use a T-test to confirm whether there is a statistically significant difference (P < 0.05) between the weight of the "heading angle offset rate" (0.62), ranked third, and the "remaining fuel warning times" (0.55), ranked fourth. If the difference is not significant, expand the truncation range (e.g., N = 4). The finally screened core flight features and core airspace features need to cover more than 80% of the processing result variations in historical cases to ensure the interpretability of the feature set.

[0075] Step S35: Combine the core flight features and core airspace features into a target flight feature set and a target airspace feature set.

[0076] The merging operation is achieved by aligning the feature names with the data structure. For example, the core flight features "flight speed deviation, altitude deviation, heading angle offset rate" and the core airspace features "airspace capacity, obstacle distribution dispersion" are classified according to the parameter type, respectively forming the target flight feature set and the target airspace feature set. The merged feature set needs to meet the requirements of the model input format. For example, time series features need to have a unified sampling frequency (such as once per second), and spatial features need to be converted to a standard coordinate system (such as WGS-84). Verify the integrity of the merged feature set, such as checking for missing values or inconsistent dimensions (such as flight speed in knots and airspace capacity in aircraft per square kilometer), and eliminate the dimensional differences through standardization processing (such as Z-score). Finally, the target feature set is used for subsequent model training. For example, the target flight feature set is input into the path adjustment model, and the target airspace feature set is input into the resource scheduling model to ensure the parameter consistency between the model training data and the real-time emergency event data.

[0077] Step S40: Based on the target flight feature set and the target airspace feature set, construct multiple initial flight emergency prediction models, and use cross-validation to iteratively train each initial flight emergency prediction model.

[0078] The initial model construction is based on diverse algorithms to adapt to different prediction requirements: The first type of model uses supervised learning algorithms (such as random forest, gradient boosting decision tree), with the target flight feature set and the target airspace feature set as inputs and the path adjustment strategy as the output. For example, when the input is "speed deviation + 20 knots, airspace capacity 6 aircraft", the output is "heading correction + 10 degrees"; The second type of model applies time series analysis (such as LSTM, ARIMA), with time series flight features (such as the altitude deviation sequence in the past 10 minutes) and meteorological parameters as inputs to predict the resource demand curve. Cross-validation adopts the K-fold strategy (such as 5-fold), dividing the historical data set into a training subset and a validation subset. For example, the first 4 folds of data are used to train the model, and the 5th fold is used to verify the prediction accuracy. In each iteration, the model adjusts the parameters through backpropagation or gradient descent methods. For example, the random forest model optimizes the node splitting rule through the Gini coefficient, and the LSTM model updates the weight matrix through the error backpropagation of the time step. Record the mean squared error (MSE) or classification accuracy of each validation during the training process, and dynamically adjust the hyperparameters (such as learning rate, tree depth) to improve the generalization ability.

[0079] As an implementation, in step S40, according to the target flight feature set and the target airspace feature set, constructing multiple initial flight emergency prediction models may specifically include: Step S41: Divide the historical emergency event data set into a training subset and a validation subset; The training subset is used to generate the initial flight emergency prediction model, and the validation subset is used to evaluate the model performance.

[0080] The historical emergency event dataset is divided into a training subset and a validation subset according to a preset ratio (such as 8:2 or 7:3) to ensure the balance of the two datasets in terms of event types, time distribution, and airspace feature coverage. The training subset is used for fitting and optimizing the model parameters. For example, 80% of the historical case data (such as 80 out of 100 cases) is selected, covering various emergency event types such as "aircraft out of control", "meteorological mutation", and "airspace congestion". The validation subset retains the remaining 20% of the case data to evaluate the generalization ability of the model on unknown data. During the division process, the problem of time leakage needs to be avoided. For example, the case timestamps in the validation subset should be later than those in the training subset to prevent the model from overfitting due to time correlation. Specifically, if the historical dataset is arranged in chronological order (from the earliest case in terms of event occurrence time to the latest case), then the data in the first 80% of the time period is used as the training subset, and the last 20% is used as the validation subset. After data division, it is necessary to verify the distribution consistency of the target flight feature set and the target airspace feature set in the two subsets. For example, the mean value of the flight speed deviation in the training subset is +15 knots, and that in the validation subset is +16 knots, indicating that there is no obvious deviation in the data distribution and it meets the model training requirements.

[0081] Step S42: According to the target flight feature set and the target airspace feature set, extract the training data corresponding to the core flight features and the core airspace features from the training subset.

[0082] Based on the target flight feature set (such as flight speed deviation, altitude deviation, heading angle deviation rate) and the target airspace feature set (such as airspace capacity, obstacle distribution dispersion), the training data extraction filters the corresponding parameter columns from the training subset and performs data cleaning and formatting. For example, the "flight speed deviation" in the target flight feature set corresponds to the difference between the real-time speed and the planned speed of each case record in the training subset, which is extracted through a numeric field; the "altitude deviation" calculates the difference between the actual altitude and the safe altitude from the aircraft altitude parameter of the case. The "airspace capacity" in the airspace feature set is obtained by counting the number of aircraft in the airspace grid, and the "obstacle distribution dispersion" is calculated based on the ratio of the obstacle spacing to the safe distance. The extracted training data needs to be processed for missing values (such as filling with linear interpolation) and standardized (such as Z-score standardization) to ensure that the input data meets the model training requirements. For example, the flight speed deviation of a certain case is +20 knots, and after standardization, it is converted to 0.75 (assuming the mean is +15 knots and the standard deviation is 6.7 knots), and the airspace capacity is 6 aircraft per square kilometer, and after standardization, it is 1.2 (the mean is 5 and the standard deviation is 0.83).

[0083] Step S43: Use the supervised learning algorithm to perform multiple rounds of fitting on the training data to generate the first type of flight emergency prediction model; the first type of flight emergency prediction model is used to predict the flight path adjustment strategy.

[0084] Supervised learning algorithms train a model to generate path planning instructions by establishing a mapping relationship between input features (target flight feature set and target airspace feature set) and output labels (path adjustment strategies in historical cases). For example, the random forest algorithm takes the flight speed deviation, altitude deviation, and airspace capacity as inputs and outputs the heading correction angle, climb altitude, and speed adjustment amount. During the training process, the model optimizes the node splitting rules and feature weights through multiple rounds of iteration. For example, in the first round of iteration, the model may preferentially split nodes based on the flight speed deviation to generate a preliminary path strategy; in subsequent iterations, the airspace capacity feature is introduced to optimize the complexity of the avoidance path. After each round of fitting, the model evaluates the prediction accuracy (such as the mean absolute error between the predicted heading angle and the actual value) through a validation subset and adjusts the hyperparameters (such as tree depth, learning rate) according to the error gradient. The finally generated model can output path adjustment instructions highly consistent with the historical handling strategies based on real-time flight features and airspace features. For example, when the input is a flight speed deviation of +20 knots and an airspace capacity of 6 aircraft, the model outputs "heading correction +12 degrees, climb to 800 meters".

[0085] As an implementation, in step S43, a supervised learning algorithm is used to perform multiple rounds of fitting on the training data to generate a first type of flight emergency prediction model, which may specifically include: Step S431: Determine the labeled result set related to flight path adjustment in the training data; the labeled result set includes the path change instructions actually executed in historical cases and the corresponding execution effect scores.

[0086] The labeled result set is extracted from the path adjustment records of historical cases, covering specific instructions such as the heading correction angle, target altitude, and speed adjustment amount, as well as the execution effect scores (such as the conflict avoidance success rate and time deviation amount after the instruction is executed). For example, the labeled result of a certain case is "heading correction +15 degrees, climb to 800 meters, speed reduced to 250 knots", and the execution effect score is 0.92 (a comprehensive score based on path deviation and time efficiency). During the labeling process, the instruction format needs to be unified. For example, the heading angle is in degrees, the altitude is in meters, and the speed is in knots to ensure the consistency of the model output. The execution effect score is calculated through quantitative indicators. For example, the time deviation amount is the difference between the actual execution time and the planned time (such as -30 seconds), which is converted into a standardized score in the 0-1 interval (such as 1 - |deviation amount| / maximum allowable deviation).

[0087] Step S432: Use the core flight features and core airspace features as input variables, and use the path change instructions and execution effect scores as output variables to construct a regression analysis model.

[0088] The regression analysis model uses the target flight characteristics (flight speed deviation, altitude deviation, heading angle deviation rate) and airspace characteristics (airspace capacity, obstacle distribution dispersion) as independent variables, and the path change instruction (such as heading correction angle) and execution effect score as dependent variables to establish a multi-output regression model. For example, the Gradient Boosting Regression Tree (GBRT) model generates a comprehensive path strategy by jointly optimizing the prediction errors of the heading angle, altitude, and speed. During model training, the loss function is designed as the weighted mean squared error (MSE), where the weight of the heading angle error is 0.6, the altitude error is 0.3, and the speed error is 0.1, to reflect the priority differences of different instructions. After the input data is standardized (such as the flight speed deviation is standardized to 0.75), the model updates the decision tree structure and leaf node weights through multiple rounds of iteration, gradually approaching the optimal fitting state.

[0089] Step S433: During each round of fitting, according to the deviation value between the predicted path instruction and the actual path instruction of the regression analysis model, adjust the weight coefficients in the model.

[0090] After each round of fitting, the model calculates the deviation value between the predicted instruction and the actual instruction (such as the heading angle deviation is +3 degrees, the altitude deviation is -20 meters), and updates the weight coefficients through the backpropagation algorithm. For example, the gradient descent method adjusts the gain threshold of feature splitting in the regression tree according to the deviation gradient, and preferentially optimizes the high-weight error terms (such as the heading angle deviation). At the same time, a regularization term (such as L2 regularization) is introduced into the loss function to prevent the model from overfitting due to excessive complexity. The adjusted weight coefficients need to re-evaluate the performance of the validation subset. If the validation error continues to decrease, continue the iteration; if an upward trend appears, trigger the early stopping mechanism to prevent performance degradation.

[0091] Step S434: When the change amount of the deviation value for continuous K rounds of fitting is less than the preset change amount threshold, it is determined that the regression analysis model has reached the convergence state, and the model parameters at this time are saved as the first type of flight emergency prediction model.

[0092] The convergence determination condition is set that the change amount of the deviation value for continuous K = 5 rounds of training (such as the fluctuation range of the heading angle error) is less than the threshold Δ = 0.01. For example, in the iteration of a certain model, the heading angle error sequence is 0.15 → 0.12 → 0.11 → 0.10 → 0.09, and its maximum change amount is 0.03 (0.15 - 0.12), which does not meet the threshold requirement; after continued training, the error sequence is 0.08 → 0.07 → 0.06 → 0.055 → 0.053, and the change amount is 0.027 (0.08 - 0.053), still higher than the threshold; finally, when the sequence is 0.05 → 0.049 → 0.048 → 0.047 → 0.046, the change amount is 0.004, lower than the threshold, and it is determined to converge. The model parameters (such as the decision tree structure, leaf node weights, feature importance) are serialized and stored for real-time prediction calls.

[0093] Step S44: Use a time series analysis algorithm to perform trend fitting on the training data to generate a second - type flight emergency prediction model; the second - type flight emergency prediction model is used to predict the emergency resource allocation strategy.

[0094] The time series analysis algorithm predicts future resource requirements and replenishment cycles by capturing the temporal patterns of historical resource scheduling data. For example, the autoregressive integrated moving average (ARIMA) model takes the temporal data of resource scheduling duration, resource consumption rate, and replenishment interval as input to construct a resource demand prediction curve. During the training process, the model first performs differencing on non - stationary temporal data (such as first - order differencing to eliminate the trend term), and then determines the optimal lag order through the autocorrelation function (ACF) and partial autocorrelation function (PACF) (such as the AR term order p = 2 and the MA term order q = 1). After model fitting, the prediction results are evaluated against the actual resource usage records through a validation subset (such as the root mean square error RMSE), and the parameters are adjusted to minimize the error. For example, in a certain case, the temporal data of the resource consumption rate is fitted by the ARIMA(2,1,1) model to generate a prediction curve of the consumption rate for the next 30 minutes, and the RMSE with the actual record curve is 0.08, indicating that the model has high - precision prediction ability. The final model can output a dynamic resource scheduling strategy, such as "Deploy 2 fire - fighting drones to coordinates (X,Y) every 20 minutes".

[0095] As an implementation, in step S44, using a time series analysis algorithm to perform trend fitting on the training data to generate a second - type flight emergency prediction model may specifically include: Step S441: Extract the time - series parameters related to resource allocation in the training data; the time - series parameters include resource scheduling duration, resource consumption rate, and resource replenishment interval.

[0096] The time - series parameters are extracted from historical resource allocation records according to a time window. For example, the resource scheduling duration sequence records the time from instruction issuance to resource arrival for each scheduling task (such as 30 minutes, 28 minutes, 32 minutes); the resource consumption rate sequence counts the resource usage per unit time (such as the fire - fighting foam consumption rate is 100 liters / minute); the resource replenishment interval sequence records the time difference between two consecutive replenishment operations (such as 120 minutes, 115 minutes, 125 minutes). During the extraction process, the timestamps need to be aligned to ensure the continuity of the sequence. For example, the resource consumption rate is sampled at 5 - minute intervals to generate equally - spaced temporal data.

[0097] Step S442: Perform a stationarity test on the time - series parameters, and perform differencing on non - stationary sequences until a stationary time series is obtained.

[0098] The stationarity test uses the Augmented Dickey-Fuller (ADF) test. If the ADF statistic is less than the critical value (e.g., p-value < 0.05), the sequence is determined to be stationary. For example, the ADF statistic of the resource scheduling duration sequence is -2.5 (p = 0.03), indicating that the sequence is stationary; the ADF statistic of the resource consumption rate sequence is -1.8 (p = 0.15), and first-order differencing is required. After differencing, the sequence is re-tested. If it is still not stationary, second-order differencing is performed until the ADF statistic meets the stationarity condition. For example, after first-order differencing, the ADF statistic of the original consumption rate sequence is -3.1 (p = 0.02), which is determined to be stationary.

[0099] Step S443: Based on the stationary time series, construct an autoregressive integrated moving average model and determine the optimal order and lag parameters of the model.

[0100] The optimal order of the autoregressive integrated moving average (ARIMA) model is determined through the analysis of the autocorrelation function (ACF) and the partial autocorrelation function (PACF). For example, if the ACF of the resource consumption rate sequence after stationarity truncates after lag 2 and the PACF truncates after lag 1, the model order is set to ARIMA(1,1,2). Parameter optimization uses the grid search method, traversing the combinations of p (0 - 3), d (1 - 2), and q (0 - 3), and selecting the parameter combination with the minimum AIC (Akaike Information Criterion) or BIC (Bayesian Information Criterion). For example, the AIC value of ARIMA(1,1,1) is 250, and that of ARIMA(2,1,1) is 245, and the latter is finally selected.

[0101] Step S444: Using the optimal order and lag parameters, perform backward fitting on the historical resource allocation data to generate a resource demand prediction curve.

[0102] After model fitting, the prediction accuracy is verified backward through historical data. For example, when the ARIMA(2,1,1) model inputs the resource consumption rate sequence for the historical 120 minutes to generate a prediction curve for the next 60 minutes, the root mean square error (RMSE) with the actual recorded curve is 0.05, indicating that the prediction result is reliable. The model also outputs a confidence interval (e.g., 95% confidence level) to provide a risk boundary reference for resource scheduling.

[0103] Step S445: According to the matching degree between the resource demand prediction curve and the actual resource usage curve, optimize the parameters of the autoregressive integrated moving average model, and mark the optimized model as the second type of flight emergency prediction model.

[0104] The optimization process improves the matching degree by adjusting the order of the ARIMA model or introducing seasonal parameters (SARIMA). For example, if there are systematic deviations in the predicted curve during periodic periods (such as daily peak hours), a seasonal order is added (such as SARIMA(2,1,1)(1,1,1,24)). After the optimized model is verified by the subset evaluation, if the RMSE is reduced to 0.03, it is marked as the second type of flight emergency prediction model, and the stored parameters are called for real-time prediction. For example, the model outputs "3 fire drones need to be dispatched within the next 1 hour, with a supply interval of 90 minutes", and the deviation from the actual demand is ±1 drone, meeting the emergency response accuracy requirements.

[0105] In an alternative derivative implementation, the method further includes the step of jointly verifying the first type of flight emergency prediction model and the second type of flight emergency prediction model, which may specifically include: Step S45: Obtain the heading change parameter, altitude correction parameter, and speed regulation parameter in the flight path adjustment strategy output by the first type of flight emergency prediction model, and extract the resource scheduling path parameter, resource delivery position parameter, and supply time interval parameter in the emergency resource allocation strategy output by the second type of flight emergency prediction model.

[0106] The heading change parameter is the correction amount indicating the heading angle of the flight path adjustment strategy. For example, the heading is offset by +15 degrees to bypass the obstacle area; the altitude correction parameter is used to adjust the vertical flight layer of the aircraft. For example, climb to 800 meters to avoid low-altitude turbulence; the speed regulation parameter defines the adjustment curve of the aircraft speed. For example, reduce the speed from 300 knots to 250 knots within 120 seconds to match the resource scheduling rhythm. The resource scheduling path parameter describes the sequence of trajectory coordinates of the movement of emergency resources (such as fire drones). For example, a straight-line path from the base coordinates (X1,Y1) to the target area coordinates (X2,Y2); the resource delivery position parameter specifies the specific geographical coordinates and altitude layer of the resource deployment. For example, deliver a navigation beacon at the coordinates (X3,Y3) and an altitude of 500 meters; the supply time interval parameter specifies the time difference between two consecutive resource supply operations. For example, perform a fuel supply every 180 minutes. The above parameters are extracted from the output results of the two types of models through a standardized data interface and converted into a unified airspace coordinate system (such as WGS-84) and timestamp format to ensure the spatio-temporal consistency of subsequent conflict detection.

[0107] Step S46: Determine the trajectory overlapping area of the two in the airspace coordinate system according to the spatial coverage range of the heading change parameter and the resource scheduling path parameter, and detect whether there is a conflict section in the trajectory overlapping area where the heading offset direction is opposite to the resource scheduling direction.

[0108] The spatial coverage is determined by calculating the intersection area between the path trajectory after adjusting the flight direction of the aircraft (such as the coordinate sequence of the aircraft in the next 10 minutes corresponding to a flight direction of +15 degrees) and the resource scheduling path (such as the movement trajectory of the UAV from X1, Y1 to X2, Y2) in the airspace grid. For example, if the aircraft trajectory passes through grid G7 within the time window from T0 + 300 seconds to T0 + 420 seconds, and the UAV passes through the same grid G7 within the time window from T0 + 360 seconds to T0 + 480 seconds, it is determined that there is a spatio-temporal overlap between the two trajectories in the G7 area. Further analyze the movement directions within the overlapping section: if the aircraft's flight direction offset is northeast (+15 degrees), while the UAV's movement direction is southwest (from X1, Y1 to X2, Y2), their directions are opposite, and there is a risk of head-on conflict. The conflict detection algorithm marks the conflict section through vector angle calculation (such as the angle between the aircraft's flight direction vector and the UAV's movement vector exceeds 150 degrees), and records the conflict time window (such as from T0 + 360 seconds to T0 + 420 seconds) and the spatial coordinate range (such as the coordinates within grid G7 from X = 123.45, Y = 45.67 to X = 123.50, Y = 45.70).

[0109] Step S47: Based on the aircraft climbing or descending instructions corresponding to the altitude correction parameter, analyze the vertical altitude limit condition in the resource delivery position parameter, and identify the compatibility status between the aircraft climbing or descending instruction and the resource delivery altitude threshold.

[0110] The aircraft climbing instruction (such as "climb to 800 meters") or descending instruction (such as "descend to 500 meters") corresponding to the altitude correction parameter needs to be verified for compatibility with the vertical altitude limit condition in the resource delivery position parameter. For example, the resource delivery position parameter stipulates that the navigation beacon needs to be deployed between 600 meters and 700 meters in altitude. If the target altitude of the aircraft climbing instruction is 800 meters, the aircraft will cross this altitude layer and there is no conflict with the resource delivery altitude threshold; however, if the target altitude of the aircraft descending instruction is 550 meters, and the resource delivery altitude threshold is 500 meters to 600 meters, then the aircraft altitude of 550 meters is within the resource delivery layer, and it is necessary to further detect whether there is spatial overlap. The compatibility status is divided into three categories: fully compatible (no intersection between the aircraft altitude and the resource altitude), partially compatible (there is a short overlap but the time windows are misaligned), and incompatible (the altitude layers overlap and the time windows conflict). For example, if the aircraft maintains an altitude of 550 meters within the time window from T0 + 300 seconds to T0 + 360 seconds, and the resource delivery time window is from T0 + 330 seconds to T0 + 390 seconds, it is determined to be in an incompatible state.

[0111] Step S48: Compare the expected flight speed change curve in the speed regulation parameter with the resource arrival time window in the supply time interval parameter, and verify whether the peak period of the speed change curve exceeds the tolerance boundary of the resource arrival time window.

[0112] The expected flight speed change curve in the speed control parameter describes the change trend of the aircraft speed over time, for example, it linearly decreases from 300 knots to 250 knots from T0+0 seconds to T0+120 seconds, and then maintains stability; the resource arrival time window in the replenishment time interval parameter stipulates that resources must be in place within a specific time range, for example, the first drone must arrive at coordinates X2, Y2 before T0+600 seconds. During verification, the peak period of the speed change curve is extracted (for example, the maximum speed gradient period from T0+0 seconds to T0+120 seconds in the deceleration stage is from T0+60 seconds to T0+90 seconds), and it is detected whether this period has a time conflict with the resource arrival window. For example, if the resource scheduling path is estimated to take 540 seconds, the estimated arrival time of the drone is T0+540 seconds, which does not overlap with the peak period of speed control (T0+60 seconds to T0+90 seconds), and is determined to be compatible; but if the resource scheduling time is extended to 660 seconds due to path adjustment (arrival time T0+660 seconds), and the speed control curve requires the aircraft to accelerate to 280 knots between T0+600 seconds and T0+660 seconds, the peak period (acceleration phase) overlaps with the resource arrival window, and it is necessary to verify whether the acceleration command causes the aircraft to occupy the conflicting airspace during the resource arrival period.

[0113] Step S49: When a conflicting section, an incompatible vertical height restriction condition or a peak period exceeding the fault tolerance boundary is detected, a parameter adjustment instruction set is generated, which includes a heading deviation angle compensation value, a resource placement height offset or a smoothing correction coefficient of a speed change curve.

[0114] The parameter adjustment instruction set is generated according to the conflict type: for the heading conflict section, the heading offset angle compensation value is calculated to eliminate the risk of opposite movement, for example, the original heading +15 degrees is adjusted to +18 degrees, so that the aircraft trajectory deviates from the conflict grid G7; for the altitude incompatibility state, the resource placement altitude offset is adjusted, for example, the navigation beacon deployment altitude is changed from 600-700 meters to 650-750 meters to avoid overlapping with the aircraft's 550-meter altitude layer; for the speed peak period conflict, the smoothing correction coefficient is introduced to adjust the gradient of the speed change curve, for example, the gradient of the deceleration stage is reduced from 5 knots / second to 3 knots / second, the deceleration time window is extended to T0+180 seconds, and the resource arrival period is staggered. The adjustment instruction set is generated through optimization algorithms (such as linear programming or genetic algorithms) to ensure that the adjusted parameters meet the emergency goals while minimizing the impact on the original strategy.

[0115] Step S410: Synchronously feed back the parameter adjustment instruction set to the first type of flight emergency prediction model and the second type of flight emergency prediction model, triggering both to regenerate the flight path adjustment strategy and emergency resource allocation strategy based on the adjusted parameters until all conflicting sections, incompatible states and boundary overflow conditions are eliminated.

[0116] The feedback mechanism injects the course deviation angle compensation value into the input parameters of the first type of model, triggering it to recalculate the path trajectory; the resource delivery height offset is input into the second type of model to adjust the resource deployment logic; the speed smoothing correction coefficient is updated to both types of models simultaneously to coordinate the spatio-temporal rhythm of the aircraft and the resources. For example, the first type of model regenerates a loitering path based on a +18-degree course to avoid grid G7; the second type of model modifies the navigation beacon delivery height to 700 meters and extends the resupply interval to 200 minutes. During the iteration process, the conflict detection module continuously verifies the compatibility of the new strategy. If there are still residual conflicts (such as the adjusted aircraft trajectory overlapping with the path of another resource), then repeat steps S45 to S49 until all conflicts are eliminated. The iteration termination condition is that there are no new conflicts in the verification results for two consecutive times, and all historical conflict flags are cleared.

[0117] Step S411: Record the logical association rules between the flight path adjustment strategy and the emergency resource allocation strategy after the conflicts are finally eliminated, and embed the logical association rules into the output parameter verification mechanism of the flight emergency prediction model.

[0118] The logical association rules encapsulate the constraint conditions of the coordination strategy through structured descriptions. For example, "when the course deviation angle ≥ +15 degrees, the resource delivery height should be ≥ 600 meters" or "the peak speed period must be at least 60 seconds earlier than the resource arrival window". The rules are encoded as executable verification conditions and embedded into the pre-verification module of the model output interface. For example, in subsequent predictions, when the first type of model generates a +15-degree course command, it automatically triggers the verification module to retrieve the associated rules and forces the second type of model to apply the ≥ 600-meter limit to the resource delivery height parameter. The rule library is dynamically expanded through case learning. For example, a new conflict type, "joint avoidance rules when visibility is insufficient due to sudden weather changes", is added to continuously enhance the model's collaborative prediction ability and conflict prevention efficiency.

[0119] Step S50: During the iterative training process, dynamically adjust the model parameters according to the difference between the simulation results and the actual processing results of each initial flight emergency prediction model for historical cases until the difference is lower than the preset convergence threshold, then stop the training, and mark the trained initial flight emergency prediction model as the flight emergency prediction model.

[0120] The degree of difference is calculated by comparing the deviation between the model prediction result and the actual disposal record. For example, if the predicted heading angle of the path adjustment strategy is +12 degrees and the actual correction value is +15 degrees, the degree of difference is 3 degrees; if the predicted number of resource requirements is 3 aircraft and the actual dispatch is 4 aircraft, the degree of difference is 1 aircraft. The degree-of-difference index is normalized to the 0-1 interval (e.g., 3 degrees / maximum allowable deviation of 15 degrees = 0.2), and is weighted and integrated into the total model degree of difference (e.g., path difference weight of 0.6 and resource difference weight of 0.4). Adaptive optimization algorithms are used for dynamic parameter adjustment. For example, the Adam algorithm adjusts the learning rate according to the gradient of the degree of difference, and the decision tree model reduces overfitting through pruning. When the fluctuation range of the total degree of difference for N consecutive rounds (e.g., N = 5) of iteration is less than the preset convergence threshold (e.g., 0.05), the model is determined to have converged. For example, if the degree-of-difference sequence of a certain model is 0.15 → 0.12 → 0.11 → 0.10 → 0.09, with a fluctuation range of 0.06 (0.15 - 0.09) higher than the threshold of 0.05, further iteration is required; if the sequence is 0.08 → 0.07 → 0.06 → 0.055 → 0.053, with a fluctuation range of 0.027 (0.08 - 0.053) lower than the threshold, training is stopped. Finally, the verified model is marked as a flight emergency prediction model and injected into the model library for real-time invocation. For example, model M-1024 is marked as "Flight Path Adjustment Model - Strong Wind Scenario", and model M-1025 is marked as "Resource Scheduling Model - Airspace Congestion Scenario".

[0121] Step S400: Generate an emergency command instruction according to the flight emergency prediction model to output a flight path adjustment strategy and an emergency resource allocation strategy.

[0122] The flight path adjustment strategy is a dynamic path planning scheme generated by the model based on real-time abnormal data and historical disposal rules. Exemplarily, it specifically includes the heading offset angle, target flight altitude layer, speed regulation curve, and waypoint sequence; the emergency resource allocation strategy covers the rescue equipment dispatch path, temporary navigation beacon deployment coordinates, activation timing of emergency communication relay nodes, and fuel replenishment cycle. When generating the emergency command instruction, first analyze the heading parameter in the path adjustment strategy and the spatial trajectory of the resource dispatch path, and detect potential conflict points between the two in the airspace coordinate system. For example, the risk of spatio-temporal collision caused by the overlap of the heading correction direction and the resource transportation route; after eliminating the conflict by adjusting the heading offset angle or the curvature of the resource path, further verify whether the target flight altitude covers the vertical safety interval of the resource delivery area (e.g., the drone delivery altitude needs to avoid high-voltage lines). If there is a height mismatch, dynamically correct the flight altitude parameter; at the same time, calibrate the synchronization rate between the speed adjustment curve and the resource replenishment time window to ensure that the acceleration or deceleration phase matches the resource arrival time period. Finally, integrate the correction parameters of heading, altitude, and speed to generate a flight control instruction set, and encapsulate the resource dispatch instruction set into a standardized signal queue, and output an executable emergency command instruction after verification by airspace control rules.

[0123] As an implementation manner, in step S400, according to the flight path adjustment strategy and the emergency resource allocation strategy output by the flight emergency prediction model, an emergency command instruction is generated, which may specifically include: parsing the heading, altitude, and speed parameters in the flight path adjustment strategy to generate a path parameter set; parsing the resource path, position, and supply parameters in the emergency resource allocation strategy to generate a resource parameter set; detecting the trajectory intersection point of the heading in the path parameter set and the movement trajectory in the resource parameter set in the airspace coordinate system, determining the conflict time window and adjusting the heading or trajectory curvature to generate an updated path and resource parameter set; verifying whether the target altitude in the path parameter set covers the vertical safety interval of the resource delivery position, and dynamically correcting the altitude parameter to generate the final path parameter when it does not cover; matching the synchronization rate of the speed adjustment curve and the resource supply period, and calibrating the time axis of the speed change stage to generate an optimized speed parameter; integrating the corrected heading, altitude, and speed parameters to generate a flight control instruction set, and integrating the adjusted resource path, position, and cycle parameters to generate a resource scheduling instruction set; converting the two types of instruction sets into standardized signals and encapsulating them into an instruction queue, verifying whether their execution timings comply with the airspace control rules, and readjusting when there is a conflict until the verification is passed and the final emergency command instruction is output.

[0124] Specifically, parse the heading change parameters, altitude correction parameters, and speed regulation parameters in the flight path adjustment strategy to generate a set of path parameters including the heading deviation angle, target flight altitude, and speed adjustment curve. At the same time, parse the resource scheduling path parameters, resource delivery location parameters, and supply time interval parameters in the emergency resource allocation strategy to generate a set of resource parameters including the resource movement trajectory, delivery coordinate sequence, and supply cycle sequence. According to the heading deviation angle in the set of path parameters and the resource movement trajectory in the set of resource parameters, detect the trajectory intersection points of the two in the airspace coordinate system, and determine the conflict time window corresponding to the trajectory intersection points. Based on the conflict time window, adjust the path curvature of the heading deviation angle or the resource movement trajectory to generate an updated set of path parameters and an updated set of resource parameters that eliminate spatio-temporal conflicts. Perform vertical altitude alignment verification on the target flight altitude in the updated set of path parameters and the delivery coordinate sequence in the set of resource parameters to identify whether the target flight altitude covers the vertical safety interval of the resource delivery location. If not, dynamically correct the target flight altitude according to the upper and lower limit values of the vertical safety interval to generate a final set of path parameters with height compatibility. According to the speed adjustment curve in the final set of path parameters and the supply cycle sequence in the updated set of resource parameters, match the synchronization rate of the speed change period and the resource arrival period, and perform time-axis calibration on the acceleration or deceleration phase of the speed adjustment curve based on the synchronization rate to generate an optimized speed regulation parameter with time synchronization. Integrate the heading deviation angle, height-compatible target flight altitude, and optimized speed regulation parameter with time synchronization in the final set of path parameters to generate a flight control instruction set including heading instructions, altitude instructions, and speed instructions. At the same time, integrate the resource movement trajectory, delivery coordinate sequence, and supply cycle sequence in the updated set of resource parameters to generate a resource scheduling instruction set including path navigation instructions, coordinate positioning instructions, and cycle trigger instructions. Convert the formats of the flight control instruction set and the resource scheduling instruction set according to the preset instruction encoding rules to generate standardized flight control signals and standardized resource scheduling signals. Logically encapsulate the standardized flight control signals and standardized resource scheduling signals, inject the emergency event identifier and time stamp, and form an executable emergency command instruction queue. Verify whether the execution timings of the flight control signals and resource scheduling signals in the emergency command instruction queue meet the airspace control rules and resource deployment constraint conditions. If there are conflicts, return to the step of adjusting the conflict time window to regenerate the instructions until all signals pass the verification and then output the final emergency command instructions.

[0125] Step S500: Execute the emergency command instructions, and update the parameters of the flight emergency prediction model according to the real-time feedback data after execution.

[0126] Exemplarily, the execution of the emergency command instructions is completed through the coordination of the air traffic control system, the aircraft autopilot module, and the ground resource scheduling platform. The real-time feedback data includes indicators such as the response accuracy of the aircraft to the instructions (such as the deviation between the actual heading angle and the instruction value), the delay between the actual arrival time of the resource deployment and the planned time, and the dynamic flight density after the airspace change. By comparing the expected response data with the actual feedback data, a deviation indicator is generated. If the deviation exceeds the fault tolerance threshold (such as the heading angle deviation is greater than 2 degrees or the resource delay exceeds 5 minutes), the model parameter update mechanism is triggered: the incremental learning algorithm is used to inject the new feedback data into the model. Without changing the original model structure, the weight parameters are adjusted by the gradient descent method to reduce the prediction error of similar scenarios. For example, if the model underestimates the deployment time of the navigation beacon during a "low visibility forced landing" event, the incremental learning will correct the weight of the resource scheduling path planning module and optimize the resource allocation strategy for subsequent similar events. The updated model needs to be jointly verified by historical data and new data to ensure its comprehensive performance stability in complex scenarios.

[0127] As an implementation manner, in step S500, the emergency command instructions are executed, and the parameters of the flight emergency prediction model are updated according to the real-time feedback data after execution, which may specifically include: step S510: Monitor the aircraft response data, resource scheduling status data, and airspace change data during the execution of the emergency command instructions.

[0128] Exemplarily, during the execution of the emergency command instructions, three types of key data are collected in real time through the multi-source sensor network and the data link: The aircraft response data includes the actual execution parameters of the aircraft for the heading correction, altitude adjustment, and speed control instructions. For example, the actual heading angle deviation is +14.5 degrees (expected +15 degrees), the actual altitude climbs to 798 meters (expected 800 meters), and the actual speed drops to 248 knots (expected 250 knots); The resource scheduling status data covers the deployment progress and operation status of emergency resources (such as fire-fighting drones, navigation beacons). For example, the deviation distance (such as 50 meters) between the actual movement trajectory of the drone from the base coordinates (X1, Y1) to the target area (X2, Y2) and the planned path, and the actual activation time of the navigation beacon at the coordinates (X3, Y3) (T0 + 620 seconds, expected T0 + 600 seconds); The airspace change data records the dynamic changes in the distribution density of aircraft in the airspace after the execution of the instructions (such as the number of aircraft in the airspace grid G7 drops from 6 to 4), the boundary adjustment of the temporary control area (such as the control area radius expands from 2 kilometers to 3 kilometers), and the update of the obstacle status (such as the addition of the coordinates of a high-voltage tower). The monitoring data is aligned by timestamp and airspace coordinates to ensure spatio-temporal consistency and provide a basis for subsequent deviation analysis.

[0129] Step S520: Compare the aircraft response data with the expected response data in the flight path adjustment strategy to generate a first deviation indicator.

[0130] The expected response data is the path adjustment instruction output by the flight emergency prediction model, such as heading correction +15 degrees, climb to 800 meters, and speed reduction to 250 knots. The deviation between the actual response data and the expected value is calculated through a quantitative index: the heading angle deviation is |14.5 - 15| = 0.5 degrees, the altitude deviation is |798 - 800| = 2 meters, and the speed deviation is |248 - 250| = 2 knots. The first deviation index comprehensively considers the parameter deviations using the weighted average method. For example, the weight of the heading angle is 0.6, the weight of the altitude is 0.3, and the weight of the speed is 0.1. Then the index value is (0.5×0.6 + 2×0.3 + 2×0.1) = 0.3 + 0.6 + 0.2 = 1.1. The index threshold is set according to flight safety standards. For example, the heading angle tolerance is ±1 degree, the altitude tolerance is ±5 meters, and the speed tolerance is ±3 knots. The corresponding first deviation index threshold is 1.8 (weighted calculated value). If it exceeds the threshold, it is determined that there is a significant deviation in the instruction execution.

[0131] Step S530: Compare the resource scheduling status data with the expected resource data in the emergency resource allocation strategy to generate a second deviation index.

[0132] The expected resource data includes resource deployment coordinates, arrival time, and quantity. For example, the fire drone should arrive at (X2, Y2) at T0 + 600 seconds, and the navigation beacon should be activated at (X3, Y3) at T0 + 600 seconds. In the actual scheduling data, the drone arrives at T0 + 620 seconds (delayed by 20 seconds), and the beacon activation time is T0 + 615 seconds (advanced by 15 seconds). The second deviation index is calculated by the weighted sum of the absolute value of the time deviation and the quantity deviation. For example, the weight of the time deviation is 0.7, and the weight of the quantity deviation is 0.3. Assuming there is no deviation in the number of drones, the index value is (20×0.7 + 0×0.3) = 14. The preset threshold is set according to the urgency of resource scheduling. For example, the time deviation threshold is ±30 seconds, and the quantity deviation threshold is ±1 unit. The corresponding second deviation index threshold is 21 (time deviation 30×0.7 + quantity deviation 1×0.3 = 21.3). The current index of 14 is not exceeded, but the trend needs to be recorded.

[0133] Step S540: When the first deviation index or the second deviation index exceeds the preset fault tolerance threshold, trigger the model parameter update mechanism and calculate the adjustment amount of the model parameters according to the current deviation index.

[0134] Exemplarily, the preset fault tolerance threshold can be dynamically adjusted according to the event priority. For example, the first deviation threshold for first-level events is set to 1.5, and the second deviation threshold is 18; for second-level events, it is relaxed to 2.0 and 24. If the current first deviation index is 1.8 (exceeding the first-level threshold of 1.5) and the second deviation is 14 (not exceeding the threshold), then parameter update for the flight path adjustment model is triggered. The adjustment amount is calculated through the error backpropagation algorithm. For example, the model weight gradient corresponding to a heading angle deviation of 0.5 degrees is Δθ = 0.5 × learning rate 0.01 = 0.005, and the gradient corresponding to a height deviation of 2 meters is Δh = 2 × 0.01 = 0.02. The adjustment amount injection mechanism determines the parameter correction symbol (such as increasing the heading weight coefficient) according to the deviation direction (such as insufficient heading correction).

[0135] Step S550: Use the incremental learning algorithm to inject the adjustment amount into the flight emergency prediction model, and re-verify the prediction accuracy of the updated model.

[0136] The incremental learning algorithm optimizes parameters iteratively with small batches of newly added data without changing the model structure. For example, the weight of the heading correction module in the flight path adjustment model was originally 0.85, and after injecting the adjustment amount +0.005, it is updated to 0.855; the weight of the height correction module is adjusted from 0.75 to 0.77. The updated model is jointly verified through the historical dataset and the newly added feedback data. For example, using the cross-validation method, it is calculated that the average heading angle error of the updated model in 100 historical cases is reduced from 1.2 degrees to 1.0 degrees, and the error in 20 newly added feedback cases is 1.1 degrees, meeting the preset performance threshold (such as the average error ≤ 1.5 degrees).

[0137] As an implementation method, in step S550, using the incremental learning algorithm to inject the adjustment amount into the flight emergency prediction model may specifically include: Step S551: Extract a new dataset from the real-time feedback data that is consistent with the type of model input parameters.

[0138] Exemplarily, the new dataset filters parameters from the real-time feedback data that are aligned with the model input features. For example, the input features of the flight emergency prediction model include flight speed deviation amount, height deviation, and airspace capacity. Then the new data needs to contain the same fields: the speed deviation amount in a certain feedback case is +20 knots, the height deviation is -5 meters, and the airspace capacity is 5 aircraft / km². Field integrity needs to be verified during data extraction. For example, invalid records with missing airspace capacity values are excluded to ensure that the feature dimensions of the new dataset are exactly the same as those of the training data.

[0139] Step S552: Perform feature alignment processing on the new dataset so that the feature dimensions of the new dataset are exactly the same as those used in the training stage.

[0140] Exemplarily, feature alignment includes data standardization and dimensionality mapping. For example, during the training phase, the flight speed deviation amount is normalized using min-max normalization (mapping the range [-50, +50] knots to [0, 1]). For the new data, +20 knots needs to be converted to (20 + 50) / 100 = 0.7. The airspace capacity is in units of aircraft per square kilometer during the training phase, and 5 aircraft in the new data can be directly retained. If the new data contains features not seen during the training phase (such as new obstacle type encodings), the dimensional consistency is maintained by filling with default values or removing the feature.

[0141] Step S553: Without changing the original model structure, input the new dataset into the flight emergency prediction model, and calculate the loss function value between the output of the flight emergency prediction model and the actual feedback result.

[0142] Exemplarily, the model structure is fixed. For example, it is the tree structure of a random forest or the number of layers of a neural network remains unchanged. After the new dataset is input, the model outputs a prediction instruction (such as a heading correction of +14 degrees), and the loss function value (such as the mean square error MSE = (14 - 14.5) 2 = 0.25) is calculated with the actual feedback heading correction execution value (+14.5 degrees). The loss function value reflects the fitting degree of the current model to the new data and is gradually reduced through iteration.

[0143] Step S554: According to the loss function value, use the gradient descent method to adjust the weight parameters in the flight emergency prediction model, so that the prediction error of the flight emergency prediction model for the new data is gradually reduced.

[0144] The gradient descent method updates the model weights according to the loss function gradient. For example, the weight matrix W of the heading correction layer in a neural network was originally [0.85, 0.12], the loss gradient is calculated as ∂L / ∂W = [0.02, -0.005], and the learning rate η = 0.01. Then the updated W = W - η×∂L / ∂W = [0.85 - 0.01×0.02, 0.12 - 0.01×(-0.005)] = [0.8498, 0.12005]. Through multiple rounds of iteration (such as 10 rounds), the loss value drops from 0.25 to 0.18, indicating that the model gradually adapts to the new data distribution.

[0145] Step S555: After each parameter adjustment, perform local validation on the flight emergency prediction model to ensure that the comprehensive performance of the adjusted flight emergency prediction model on the historical dataset and the new dataset is not lower than the preset performance threshold.

[0146] Exemplarily, stratified sampling can be used for local verification. For example, 50 cases are randomly selected from the historical dataset and all 20 cases of the new dataset are used to calculate the comprehensive performance index (such as the mean absolute error MAE). For example, the MAE of the historical data increases from 1.2 degrees to 1.3 degrees (the allowable threshold is 1.5), and the MAE of the new data decreases from 1.1 degrees to 1.0 degrees, and the overall performance remains qualified. If the adjustment causes the MAE of the historical data to exceed the limit (such as rising to 1.6 degrees), the parameters are rolled back and the learning rate is reduced for readjustment until the performance constraints of the dual datasets are met. After the verification is passed, the updated model parameters are persistently stored for subsequent emergency command use.

[0147] Figure 2 The following is a schematic diagram of the hardware entity of a computer system provided by an embodiment of the present invention. As Figure 2 shown, the hardware entity of the computer system 1000 includes: a processor 1001 and a memory 1002. Among them, the memory 1002 stores a computer program that can run on the processor 1001, and when the processor 1001 executes the program, it implements the steps in the method of any of the above embodiments.

Claims

1. An emergency command method for low-altitude flight, characterized in that: The method comprises: acquiring real-time flight monitoring data of a target area; the real-time flight monitoring data comprises aircraft status parameters, meteorological parameters and airspace occupancy parameters; determining the type of emergency event in the target area according to an abnormal data subset in the real-time flight monitoring data; calling a flight emergency prediction model corresponding to the type of emergency event; the flight emergency prediction model is trained based on flight characteristics and airspace characteristics screened from historical emergency event data, and the types of input parameters and output parameters of each flight emergency prediction model are consistent in a training phase and an application phase; outputting a flight path adjustment strategy and an emergency resource allocation strategy according to the flight emergency prediction model, and generating an emergency command instruction; executing the emergency command instruction, and updating the parameters of the flight emergency prediction model according to real-time feedback data after execution.

2. The method according to claim 1, characterized in that The method of determining the type of emergency event in the target area according to the abnormal data subset in the real-time flight monitoring data comprises: extracting a first data subset corresponding to aircraft state parameters, a second data subset corresponding to meteorological parameters, and a third data subset corresponding to airspace occupancy parameters from the real-time flight monitoring data; performing dynamic threshold screening on the first data subset, the second data subset, and the third data subset respectively to obtain abnormal aircraft data exceeding a preset aircraft state threshold in the first data subset, abnormal meteorological data exceeding a preset meteorological fluctuation threshold in the second data subset, and abnormal airspace data exceeding a preset airspace capacity threshold in the third data subset; inputting the abnormal aircraft data, abnormal meteorological data, and abnormal airspace data into an event type association model to determine the priority and association parameters of the emergency event type; generating an event feature label containing the emergency event type according to the priority and association parameters, and matching the event feature label with the calling condition of the flight emergency prediction model.

3. The method according to claim 2, characterized in that The calling of the flight emergency prediction model corresponding to the emergency event type includes: obtaining multiple candidate flight emergency prediction models associated with the event feature label; each candidate flight emergency prediction model corresponds to a different emergency event processing stage, and the input parameter type adopted by each candidate flight emergency prediction model in the training stage is consistent with the parameter type in the current real-time flight monitoring data; determining the response weight and error range of each candidate flight emergency prediction model according to the historical verification results of the candidate flight emergency prediction model; screening out at least one target flight emergency prediction model whose response weight is higher than a preset weight threshold and whose error range is lower than a preset error threshold from the candidate flight emergency prediction models; reorganizing the abnormal aircraft data, abnormal meteorological data and abnormal airspace data according to the input format of the target flight emergency prediction model in the training stage to generate a standardized input data set; inputting the standardized input data set into the target flight emergency prediction model to trigger the target flight emergency prediction model to perform flight path simulation and resource demand prediction, and obtain a flight path adjustment strategy and an emergency resource allocation strategy.

4. The method according to claim 3, characterized in that The training process of the flight emergency prediction model includes: obtaining a historical emergency event data set; the historical emergency event data set includes a plurality of processed emergency event cases, each emergency event case includes aircraft parameters, meteorological parameters, airspace parameters when the emergency event occurs, and path adjustment records and resource allocation records after the emergency event is processed; extracting a flight feature set related to flight speed, altitude deviation, and heading angle deviation from the aircraft parameters of each emergency event case, and extracting an airspace feature set related to airspace capacity, obstacle distribution, and temporary control area from the airspace parameters; performing redundancy analysis on the flight feature set and the airspace feature set to remove The optimized target flight feature set and target airspace feature set are obtained by removing redundant features whose correlation with emergency event handling results is lower than a preset correlation threshold. Multiple initial flight emergency prediction models are constructed based on the target flight feature set and the target airspace feature set, and each initial flight emergency prediction model is iteratively trained by cross-validation. During the iterative training process, model parameters are dynamically adjusted according to the difference between the simulation results of historical cases and the actual handling results of each initial flight emergency prediction model, and the training is stopped when the difference is lower than a preset convergence threshold, and the trained initial flight emergency prediction model is marked as the flight emergency prediction model.

5. The method according to claim 4, characterized in that The redundancy analysis of the flight feature set and the airspace feature set is performed to remove redundant features whose correlation with the emergency event processing result is lower than a preset correlation threshold, so as to obtain an optimized target flight feature set and a target airspace feature set, including: calculating a first correlation coefficient between each flight feature in the flight feature set and the emergency event processing result, and a second correlation coefficient between each airspace feature in the airspace feature set and the emergency event processing result; removing target flight features whose first correlation coefficient is lower than a first correlation coefficient threshold from the flight feature set, and removing target airspace features whose second correlation coefficient is lower than a second correlation coefficient threshold from the airspace feature set; performing a combined analysis on the remaining flight features and the remaining airspace features after the removal to determine the influence weights of different feature combinations on the processing results; selecting the core flight features ranked in the top N positions by weight from the remaining flight features according to the influence weights, and selecting the core airspace features ranked in the top M positions by weight from the remaining airspace features; and merging the core flight features and the core airspace features into the target flight feature set and the target airspace feature set.

6. The method according to claim 5, characterized in that The method constructs multiple initial flight emergency prediction models based on the target flight feature set and the target airspace feature set, including: dividing the historical emergency event data set into a training subset and a verification subset; the training subset is used to generate an initial flight emergency prediction model, and the verification subset is used to evaluate model performance; according to the target flight feature set and the target airspace feature set, training data corresponding to core flight features and core airspace features are extracted from the training subset; a supervised learning algorithm is used to perform multiple rounds of fitting on the training data to generate a first type of flight emergency prediction model; the first type of flight emergency prediction model is used to predict a flight path adjustment strategy; a time series analysis algorithm is used to perform trend fitting on the training data to generate a second type of flight emergency prediction model; the second type of flight emergency prediction model is used to predict an emergency resource allocation strategy.

7. The method according to claim 6, characterized in that The supervised learning algorithm is used to perform multiple rounds of fitting on the training data to generate a first type of flight emergency prediction model, including: determining a set of annotated results related to flight path adjustment in the training data; the annotated result set includes the path change instructions actually executed in historical cases and the corresponding execution effect scores; using the core flight characteristics and core airspace characteristics as input variables, and using the path change instructions and the execution effect scores as output variables to construct a regression analysis model; in each round of fitting, adjusting the weight coefficient in the model according to the deviation value between the predicted path instruction and the actual path instruction of the regression analysis model; when the change in the deviation value after K consecutive rounds of fitting is less than a preset change threshold, determining that the regression analysis model has reached a convergence state, and saving the model parameters at this time as the first type of flight emergency prediction model; the time series analysis is used to analyze the flight path of the regression analysis model; and the time series analysis is used to analyze the flight path of the regression analysis model. The method comprises the following steps: performing trend fitting on the training data using an analysis algorithm to generate a second type of flight emergency prediction model, including: extracting time series parameters related to resource allocation in the training data; the time series parameters include resource scheduling duration, resource consumption rate, and resource replenishment interval; performing a stationarity test on the time series parameters, and performing differential processing on the non-stationary series until a stationary time series is obtained; constructing an autoregressive integral moving average model based on the stationary time series, and determining the optimal order and lag parameters of the model; performing reverse fitting on the historical resource allocation data using the optimal order and lag parameters to generate a resource demand prediction curve; optimizing the parameters of the autoregressive integral moving average model based on the matching degree between the resource demand prediction curve and the actual resource usage curve, and marking the optimized model as the second type of flight emergency prediction model.

8. The method according to claim 1, characterized in that The executing of the emergency command instruction and updating the parameters of the flight emergency prediction model according to the real-time feedback data after the execution include: monitoring the aircraft response data, resource scheduling status data and airspace change data during the execution of the emergency command instruction; comparing the aircraft response data with the expected response data in the flight path adjustment strategy to generate a first deviation index; comparing the resource scheduling status data with the expected resource data in the emergency resource allocation strategy to generate a second deviation index; when the first deviation index or the second deviation index exceeds a preset fault tolerance threshold, triggering a model parameter update mechanism, and calculating the adjustment amount of the model parameter according to the current deviation index; using an incremental learning algorithm to inject the adjustment amount into the flight emergency prediction model, and re-verifying the prediction accuracy of the updated model.

9. The method according to claim 8, characterized in that The incremental learning algorithm is used to inject the adjustment amount into the flight emergency prediction model, including: extracting a new data set that is consistent with the model input parameter type from the real-time feedback data; performing feature alignment processing on the new data set so that the feature dimension of the new data set is exactly the same as the feature dimension used in the training phase; without changing the original model structure, inputting the new data set into the flight emergency prediction model, and calculating the loss function value between the output of the flight emergency prediction model and the actual feedback result; according to the loss function value, using the gradient descent method to adjust the weight parameters in the flight emergency prediction model so that the prediction error of the flight emergency prediction model for new data is gradually reduced; after each parameter adjustment, the flight emergency prediction model is locally verified to ensure that the comprehensive performance of the adjusted flight emergency prediction model on the historical data set and the new data set is not lower than the preset performance threshold.

10. A computer system comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, wherein: When the processor executes the program, the steps in the method according to any one of claims 1 to 9 are implemented.

Images