Machine Learning-Based Low-Altitude Aircraft Risk Management Method and System

Abstract

The present invention discloses a risk management method and system for low-altitude aircraft based on machine learning. S1. Generate a preprocessed flight trajectory data set; S2. Apply the fractal dimension analysis method to the multi-scale flight trajectory data set to extract the complexity description vector of the flight trajectory; S3. Construct a multi-dimensional risk feature data set; S4. Automatically select the optimal risk evaluation model with the minimum description length and the best data fitting accuracy; S5. Use the selected optimal risk evaluation model to perform real-time risk assessment on the newly collected flight trajectory data set and its corresponding multi-dimensional risk feature data set to generate a flight trajectory risk evaluation output; S6. Implement dynamic risk monitoring and flight path adjustment, and provide a decision-making basis for low-altitude aircraft risk management. The present invention accurately captures high-risk feature segments with high turning rates or sharp speed changes in the trajectory, greatly improving the recognition accuracy of local risk events.

Description

Technical Field

[0001] The present invention relates to the technical field of low-altitude flight, and in particular, to a risk management method and system for low-altitude aircraft based on machine learning. Background Art

[0002] With the gradual opening of low-altitude airspace and the rapid development of unmanned aerial vehicle technology, low-altitude aircraft are increasingly widely used in the fields of urban logistics, agricultural plant protection, power inspection, and emergency rescue. However, when operating in complex environments, low-altitude aircraft face many potential risk factors, such as sudden meteorological changes, terrain interference, communication delays, and fluctuations in the aircraft's own performance, which are likely to cause flight safety problems such as flight yaw, attitude instability, and abnormal trajectories. Therefore, how to effectively monitor the flight state of low-altitude aircraft and conduct risk assessment has become an important issue that urgently needs to be solved in current low-altitude flight safety management.

[0003] Existing low-altitude aircraft risk management technologies mainly rely on rule-based library determination and traditional model evaluation. Most systems are based on static risk indicators and judge whether the flight trajectory crosses the boundary or has abnormal points through preset rules, but there are obvious deficiencies: on the one hand, static rules are difficult to adapt to the constantly changing dynamic characteristics during the flight process and cannot effectively identify the interaction risks between trajectory complexity and flight parameters; on the other hand, currently commonly used risk assessment models usually rely on manually set features or single-scale analysis, lacking the ability to finely model the local changes of flight trajectories and the complex spatial structure, resulting in low accuracy in identifying abnormal behaviors or potential risks. In addition, although some systems integrate basic flight trajectory monitoring modules, they lack the ability to jointly analyze multi-dimensional dynamic flight parameters and trajectory shape changes, and are prone to misjudgment or missed judgment, seriously affecting the timeliness and accuracy of risk response.

[0004] In summary, there is an urgent need for a full-process risk management method that can integrate multi-scale trajectory information and flight dynamic characteristics and has intelligent evaluation capabilities to improve the situation. Summary of the Invention

[0005] An object of the present invention is to propose a risk management method and system for low-altitude aircraft based on machine learning. The present invention accurately captures high-risk feature segments with high turning rates or sharp speed changes in the trajectory, greatly improving the accuracy of identifying local risk events.

[0006] A risk management method for low-altitude aircraft based on machine learning according to an embodiment of the present invention includes the following steps:

[0007] S1. Real-time obtain the original three-dimensional flight trajectory data set of the low-altitude aircraft in the specified low-altitude area and perform preprocessing to generate a preprocessed flight trajectory data set;

[0008] S2. Perform multi-scale segmentation processing on the preprocessed flight trajectory dataset to form a flight trajectory dataset at different scales, and apply the fractal dimension analysis method to the flight trajectory dataset at different scales to extract the complexity description vector of the flight trajectory;

[0009] S3. Perform data fusion on the flight trajectory complexity description vector and the dynamic flight parameter data during the operation of the low-altitude aircraft to construct a multi-dimensional risk feature dataset;

[0010] S4. Based on the minimum description length principle, screen the candidate models of the risk assessment model for the multi-dimensional risk feature dataset. By evaluating the model coding length and data coding length of each candidate model, automatically select the optimal risk assessment model with the minimum description length and the best data fitting accuracy;

[0011] S5. Use the selected optimal risk assessment model to perform real-time risk assessment on the newly collected flight trajectory dataset and its corresponding multi-dimensional risk feature dataset, and generate a flight trajectory risk assessment output;

[0012] S6. Feed back the flight trajectory risk assessment output to the flight control system and the monitoring terminal to achieve dynamic risk monitoring and flight path adjustment, and provide a decision-making basis for the risk management of low-altitude aircraft.

[0013] Optionally, the S1 includes the following steps:

[0014] S11. Set the flight monitoring time window of the low-altitude aircraft in the specified low-altitude flight area , and through the inertial measurement unit, global positioning system module and attitude sensor deployed on the low-altitude aircraft, collect the original flight trajectory data in real time to construct a three-dimensional original flight trajectory dataset , the three-dimensional original flight trajectory dataset consists of several sampling points, and each sampling point includes the three-dimensional space coordinates of the aircraft at a specific moment and the corresponding timestamp:

[0015] ;

[0016] Among them, represents the th flight trajectory point, are respectively the three-dimensional space coordinates of the low-altitude aircraft at the moment , is the corresponding timestamp, is the total number of samples;

[0017] S12. For the three-dimensional original flight trajectory dataset ​Perform time series alignment processing. By resampling the time axis of the flight trajectory points, the time interval between all adjacent sampling points is kept as a fixed time difference, and the flight trajectory data set after time normalization is obtained. ;

[0018] S13. For the flight trajectory data set after time normalization Perform outlier removal processing. Use the speed change and position deviation within the sliding window for judgment. When the speed change of the flight trajectory point within the local window exceeds the speed change threshold or the spatial position of the flight trajectory point deviates beyond the position deviation threshold, then the flight trajectory point is determined as an outlier and removed from the flight trajectory. After processing, a flight trajectory data set after outlier removal is generated. ;

[0019] S14. For the flight trajectory data set after outlier removal Perform noise filtering processing. Use the sliding window mean filtering method to smooth the flight trajectory points. Calculate the mean value with a fixed number of adjacent front and rear flight trajectory points at each flight trajectory point position and replace the original flight trajectory point position to generate a preprocessed flight trajectory data set. .

[0020] Optionally, the S2 includes the following steps:

[0021] S21. According to the requirement that the low-altitude aircraft trajectory risk assessment is sensitive to the local complexity of the trajectory, construct an adaptive local scale division function jointly based on the spatial curvature change rate and average flight speed of the flight trajectory:

[0022] ;

[0023] Among them, is the adaptive local scale at the flight trajectory point , is the preset basic segmentation length, is the spatial curvature change rate at the flight trajectory point , reflecting the severity of the spatial turning of the flight trajectory, is the instantaneous flight speed of the aircraft at the flight trajectory point , is the average flight speed of the flight trajectory, , are parameters for controlling the sensitivity of adaptive local scale adjustment;

[0024] S22. Use the adaptive local scale division function to perform local dynamic segmentation on the preprocessed flight trajectory data set to generate a multi-scale flight trajectory data set , any flight trajectory segment in the multi-scale flight trajectory dataset is defined as:

[0025] ;

[0026] Among them, the starting moment of the flight trajectory segment is , and the flight trajectory point is the -th point within the flight trajectory segment , is the total number of flight trajectory segments after segmentation, represents the adaptive local scale size dynamically determined based on the current curvature change rate and instantaneous velocity with the starting point of the flight trajectory as the reference point;

[0027] S23. For each flight trajectory segment , use the box dimension algorithm in the fractal dimension analysis method to calculate the local fractal dimension eigenvalue of the flight trajectory segment , obtain the local fractal dimension feature set. The local fractal dimension eigenvalue quantitatively characterizes the local irregularity and structural complexity of the low-altitude aircraft flight trajectory within the flight trajectory segment, reflects the local risk characteristics of the flight trajectory, and construct a flight trajectory complexity description vector based on the local fractal dimension feature set .

[0028] Optionally, the S3 includes the following steps:

[0029] S31. For each sampling point in the preprocessed flight trajectory dataset , obtain the corresponding instantaneous flight speed of the sampling point. The instantaneous flight speed is calculated from the spatial distance between adjacent sampling points, the instantaneous acceleration of the aircraft, and the instantaneous acceleration of the aircraft is calculated from the absolute value of the speed difference between adjacent sampling points, and the flight attitude angle . The flight attitude angle is directly collected from the output of the aircraft attitude sensor, and calculate the instantaneous flight speed change value , the instantaneous acceleration change value of the aircraft and the flight attitude angle change value based on the difference between the parameters of adjacent sampling points, and form a dynamic flight parameter dataset ;

[0030] S32. For each flight trajectory segment , use the sliding window statistical method from the dynamic flight parameter dataset Extract the aggregated dynamic flight parameter vector corresponding to the flight trajectory segmentation In this aggregated dynamic flight parameter vector The instantaneous flight speed change value Is the difference between the maximum instantaneous speed and the minimum instantaneous speed of each sampling point within the flight trajectory segmentation The instantaneous acceleration change value of the aircraft Is the difference between the maximum instantaneous acceleration and the minimum instantaneous acceleration of each sampling point within the flight trajectory segmentation The flight attitude angle change value Is the average value of the attitude angle change values of consecutive sampling points within the flight trajectory segmentation ;

[0031] S33. Perform vector concatenation data fusion on the flight trajectory complexity description vector And the aggregated dynamic flight parameter vector of the corresponding flight trajectory segmentation To form a multi-dimensional risk feature vector The multi-dimensional risk feature vector simultaneously includes the complexity description of the flight trajectory at each local scale and the instantaneous flight speed change value, the instantaneous acceleration change value of the aircraft, and the flight attitude angle change value within the flight trajectory segmentation ;

[0032] S34. Combine the multi-dimensional risk feature vectors corresponding to all flight trajectory segments In sequence according to the sampling order to form a multi-dimensional risk feature data set .

[0033] Optionally, the S4 includes the following steps:

[0034] S41. Construct a set of candidate risk assessment models Each candidate model in the set of candidate risk assessment models Is defined as a prediction function that maps the multi-dimensional risk feature data set To the corresponding set of risk level labels :

[0035] ;

[0036] Wherein, Is the multi-dimensional risk feature vector of the th flight trajectory segment, Is the predicted risk level prediction value predicted by the candidate model , Is the candidate model parameter vector, ​​​Represents a candidate model 's decision mapping structure, specifically a non - linear model function that maps the input multi - dimensional risk feature vector to the risk level prediction value ; 's structure can be a decision tree, support vector machine, shallow neural network, or other pre - set trainable risk scoring functions, and its structural complexity is measured by ;

[0037] S42. Calculate the model encoding length and data encoding length for the candidate model . The model encoding length , combining the complexity of the model structure and the parameter vector metric information complexity, is defined as:

[0038] ;

[0039] Among them, represents the total number of parameters included in the candidate model , represents the structural depth of the decision mapping structure or the number of piece - wise functions, , are pre - set structural penalty factors; the data encoding length is defined as the total prediction error encoding loss of the candidate model for all multi - dimensional risk feature data points. Combining the true risk level label and the risk level prediction value , it is expressed using the weighted error encoding method as:

[0040] ;

[0041] Among them, is the prior importance weight of the risk label, is the probability value of the model predicting correctly under the multi - dimensional risk feature vector and the total number of parameters ;

[0042] S43. Combine the model encoding length and data encoding length, and determine the optimal risk assessment model based on the minimum description length criterion. The optimal risk assessment model satisfies the following optimization objective:

[0043] ;

[0044] The optimal risk assessment model realizes the dynamic balance among the model structure complexity, prediction accuracy, and the importance of risk weights in the low - altitude aircraft flight path risk assessment scenario.

[0045] Optionally, S5 includes the following steps:

[0046] S51. Collect the original 3D flight trajectory data set generated in real time during the operation of the low-altitude aircraft to form a new multi-dimensional risk feature data set , and input the new multi-dimensional risk feature data set into the optimal risk assessment model , predict the risk level for each flight trajectory segment, and generate a set of flight trajectory risk prediction results ;

[0047] S52. According to the set of flight trajectory risk prediction results , perform flight trajectory risk assessment output in combination with the preset risk level mapping rule, and the risk level mapping rule is set according to the segmented risk level value and the corresponding multi-dimensional risk feature parameter threshold;

[0048] S53. Synchronously feedback the flight trajectory risk assessment output result to the aircraft control system and the remote monitoring terminal in the form of risk level labels and warning signals. The warning signals include the trajectory segmentation position, risk level, trigger parameters, and over-limit values, which are used to assist in the dynamic adjustment of the flight path and the real-time management of flight safety.

[0049] Optionally, the risk level division rule includes:

[0050] When the fractal dimension eigenvalue is greater than the set upper threshold, and the instantaneous flight speed change value or acceleration change value exceeds the corresponding aircraft safety tolerance range, it is determined that this trajectory segment is a high-risk segment, and a high-risk level is output;

[0051] When the fractal dimension eigenvalue is in the middle range, and the flight attitude angle change value is within the normal fluctuation range, but there is an abnormal fluctuation in a single parameter, it is determined that this trajectory segment is a medium-risk segment, and a medium-risk level is output;

[0052] When the change values of all dynamic flight parameters are within the safety threshold range and the fractal dimension eigenvalue is below the set lower limit range, it is determined that this trajectory segment is a low-risk segment, and a low-risk level is output.

[0053] A low-altitude aircraft risk management system based on machine learning for implementing the above-mentioned low-altitude aircraft risk management method based on machine learning, includes the following modules:

[0054] A trajectory data acquisition module for collecting the 3D flight trajectory data set of the aircraft within a specified area;

[0055] A data preprocessing module for performing spatio-temporal alignment, outlier removal, and filtering processing on the flight trajectory data set to generate a preprocessed flight trajectory data set;

[0056] A multi-scale feature extraction module, which is used to segment a flight trajectory based on adaptive scales and extract local fractal dimensions to form a description vector of the complexity of the flight trajectory;

[0057] A dynamic parameter fusion module, which is used to extract the velocity change value, acceleration change value and attitude angle change value of each segmented flight trajectory, and fuse them with the complexity description vector to generate a multi-dimensional risk feature dataset;

[0058] A model selection module, which is used to select the optimal risk assessment model with the minimum description length from candidate models based on the minimum description length principle;

[0059] A risk assessment module, which is used to predict the risk level of the flight trajectory based on the selected model and output early warning information;

[0060] A feedback module, which is used to feedback the risk assessment result to the aircraft control system and the monitoring terminal to realize the dynamic adjustment of the flight path and the linkage of risk early warning.

[0061] The beneficial effects of the present invention are as follows:

[0062] (1) The present invention introduces the local fractal dimension of the flight trajectory as the core descriptor into the risk management system, extracts the fractal dimension features through the segmented flight trajectories after multi-scale segmentation, effectively characterizes the local complexity and irregularity of the flight trajectory at different scales. In addition, it is vector-cascaded and fused with the dynamic flight parameters such as the instantaneous velocity change value, acceleration change value and attitude angle change value to form a multi-dimensional risk feature vector, significantly enhancing the sensitivity of the model to small trajectory perturbations and potential risk signs, and improving the accurate recognition rate of high-risk trajectory segments.

[0063] (2) The present invention breaks through the traditional way of relying on manually setting the model structure or fixed templates, and uses the minimum description length principle in information theory to jointly optimize and evaluate the coding complexity and data error of the candidate model set, comprehensively considering the information loss caused by the model structure complexity and prediction error, and automatically screening out the optimal model with both a compact structure and high risk discrimination ability from many candidate models, ensuring that the selected model has stable and reliable risk assessment performance in different flight environments and trajectory modes.

[0064] (3) The present invention designs an adaptive division function jointly driven by the spatial curvature change rate of flight trajectory points and the flight speed deviation degree to realize non-uniform dynamic segmentation of the flight trajectory, which not only effectively avoids the problems of feature loss or redundancy caused by fixed time window segmentation, but also can accurately capture the high-risk feature segments with high turning rates or drastic speed changes in the trajectory, greatly improving the recognition accuracy of local risk events. Description of the Drawings

[0065] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the accompanying drawings:

[0066] Figure 1 It is a flowchart of a risk management method and system for low-altitude aircraft based on machine learning proposed by the present invention. Detailed implementation manners

[0067] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0068] Reference Figure 1 , a risk management method for low-altitude aircraft based on machine learning, includes the following steps:

[0069] S1. Real-time obtain the original three-dimensional flight trajectory data set of the low-altitude aircraft in the specified low-altitude area and perform preprocessing to generate a preprocessed flight trajectory data set;

[0070] S2. Perform multi-scale segmentation processing on the preprocessed flight trajectory data set to form a multi-scale flight trajectory data set, and apply the fractal dimension analysis method to the multi-scale flight trajectory data set to extract the complexity description vector of the flight trajectory;

[0071] S3. Perform data fusion on the flight trajectory complexity description vector and the dynamic flight parameter data during the operation of the low-altitude aircraft to construct a multi-dimensional risk feature data set;

[0072] S4. Based on the minimum description length principle, screen the candidate models of the risk assessment model for the multi-dimensional risk feature data set. By evaluating the model coding length and data coding length of each candidate model, automatically select the optimal risk assessment model with the minimum description length and the best data fitting accuracy;

[0073] S5. Use the selected optimal risk assessment model to perform real-time risk assessment on the newly collected flight trajectory data set and its corresponding multi-dimensional risk feature data set to generate a flight trajectory risk assessment output;

[0074] S6. Feed back the flight trajectory risk assessment output to the flight control system and the monitoring terminal to realize dynamic risk monitoring and flight path adjustment, and provide a decision-making basis for the risk management of low-altitude aircraft.

[0075] In this embodiment, S1 includes the following steps:

[0076] S11. Set the flight monitoring time window of the low-altitude aircraft in the specified low-altitude flight area , the raw data of the flight trajectory is collected in real time through the inertial measurement unit, global positioning system module and attitude sensor deployed on the low-altitude aircraft, and a three-dimensional raw flight trajectory dataset is constructed. , the three-dimensional raw flight trajectory dataset consists of a number of sampling points, and each sampling point includes the three-dimensional spatial coordinates of the aircraft at a specific moment and the corresponding timestamp:

[0077] ;

[0078] Among them, represents the -th flight trajectory point, are respectively the three-dimensional spatial coordinates of the low-altitude aircraft at the moment , is the corresponding timestamp, is the total number of samplings;

[0079] S12. Perform time series alignment processing on the three-dimensional raw flight trajectory dataset , and by resampling the time axis of the flight trajectory points, make the time interval between all adjacent sampling points remain a fixed time difference, and obtain the flight trajectory dataset after time normalization ;

[0080] S13. Perform outlier removal processing on the flight trajectory dataset after time normalization , and use the speed change and position deviation within the sliding window for judgment. When the speed change of the flight trajectory point within the local window exceeds the speed change threshold or the spatial position of the flight trajectory point deviates beyond the position deviation threshold, then the flight trajectory point is determined as an outlier and removed from the flight trajectory, and the flight trajectory dataset after outlier removal is generated after processing ;

[0081] S14. Perform noise filtering processing on the flight trajectory dataset after outlier removal , and use the sliding window mean filtering method to smooth the flight trajectory points. Calculate the mean value of a fixed number of adjacent front and back flight trajectory points at the position of each flight trajectory point and replace the original flight trajectory point position to generate the preprocessed flight trajectory dataset .

[0082] In this embodiment, S2 includes the following steps:

[0083] S21. According to the requirement that the risk assessment of the low-altitude aircraft flight track is sensitive to the local complexity of the track, construct an adaptive local scale division function based on the joint of the spatial curvature change rate and the average flight speed of the flight track:

[0084] ;

[0085] Among them, is the adaptive local scale at the flight trajectory point ; is the preset basic segmentation length, is the flight trajectory point The spatial curvature change rate at, which reflects the severity of the spatial turning of the flight trajectory, is the flight trajectory point The instantaneous flight speed of the aircraft at, is the average flight speed of the flight trajectory, , are parameters for controlling the sensitivity of adaptive local scale adjustment;

[0086] S22. Use the adaptive local scale division function Perform local dynamic segmentation on the preprocessed flight trajectory dataset to generate a multi-scale flight trajectory dataset . Any flight trajectory segment in the multi-scale flight trajectory dataset is defined as:

[0087] ;

[0088] Among them, the starting time of the flight trajectory segment is , the flight trajectory point is the th point in the flight trajectory segment , is the total number of flight trajectory segments after segmentation, represents the size of the adaptive local scale dynamically determined based on the current curvature change rate and instantaneous speed with the starting point of the flight trajectory as the reference point;

[0089] S23. For each flight trajectory segment , use the box dimension algorithm in the fractal dimension analysis method to calculate the local fractal dimension eigenvalue of the flight trajectory segment , obtain the local fractal dimension feature set. The local fractal dimension eigenvalue quantitatively characterizes the local irregularity and structural complexity of the flight trajectory of the low-altitude aircraft in the flight trajectory segment, reflects the local risk characteristics of the flight trajectory, and constructs a flight trajectory complexity description vector based on the local fractal dimension feature set .

[0090] In this embodiment, S3 includes the following steps:

[0091] S31. For each sampling point in the preprocessed flight trajectory dataset , obtain the instantaneous flight speed corresponding to this sampling point , where the instantaneous flight speed is calculated from the spatial distance between adjacent sampling points, the instantaneous acceleration of the aircraft , and the instantaneous acceleration of the aircraft is calculated from the absolute value of the speed difference between adjacent sampling points, and the flight attitude angle , the flight attitude angle is directly collected from the output of the aircraft attitude sensor, and the instantaneous flight speed change value is calculated based on the difference in parameters between adjacent sampling points , the instantaneous acceleration change value of the aircraft , and the flight attitude angle change value , to form a dynamic flight parameter dataset ;

[0092] S32. For each flight trajectory segment , use the sliding window statistical method to extract from the dynamic flight parameter dataset the aggregated dynamic flight parameter vector corresponding to the flight trajectory segment . In this aggregated dynamic flight parameter vector , the instantaneous flight speed change value is the difference between the maximum instantaneous speed and the minimum instantaneous speed of each sampling point within the flight trajectory segment , the instantaneous acceleration change value of the aircraft is the difference between the maximum instantaneous acceleration and the minimum instantaneous acceleration of each sampling point within the flight trajectory segment , and the flight attitude angle change value is the average value of the attitude angle change values of consecutive sampling points within the flight trajectory segment ;

[0093] S33. Perform vector - level cascade data fusion on the flight trajectory complexity description vector and the aggregated dynamic flight parameter vector corresponding to the respective flight trajectory segment to form a multi - dimensional risk feature vector . The multi - dimensional risk feature vector simultaneously includes the complexity description of the flight trajectory at each local scale and the instantaneous flight speed change value, the instantaneous acceleration change value of the aircraft, and the flight attitude angle change value within the flight trajectory segment ;

[0094] S34. Combine in sequence the multi - dimensional risk feature vectors corresponding to all flight trajectory segments to form a multi - dimensional risk feature dataset .

[0095] In this embodiment, S4 includes the following steps:

[0096] S41. Construct a set of candidate risk assessment models , and each candidate model in the set of candidate risk assessment models is defined as a prediction function that maps a multi-dimensional risk feature data set to a corresponding set of risk level labels :

[0097] ;

[0098] Among them, is the multi-dimensional risk feature vector of the th flight trajectory segment, is the predicted risk level prediction value corresponding to candidate model , is the candidate model parameter vector, represents the decision mapping structure of candidate model , specifically a non-linear model function that maps the input multi-dimensional risk feature vector to the risk level prediction value ; The structure of can be a decision tree, a support vector machine, a shallow neural network, or other predefined trainable risk scoring functions, and its structural complexity is measured by ;

[0099] S42. Calculate the model coding length and data coding length for candidate model . The model coding length is defined by combining the structural complexity of the model and the parameter vector metric information complexity as:

[0100] ;

[0101] Among them, represents the total number of parameters included in candidate model , represents the structural depth or the number of piecewise functions of the decision mapping structure , , are predefined structural penalty factors; the data coding length is defined as the total encoding loss of the prediction error of candidate model for all multi-dimensional risk feature data points. Combining the true risk level label and the risk level prediction value , it is expressed using a weighted error coding method as:

[0102] ;

[0103] Among them, is the prior importance weight of the risk label, is the probability value that the model predicts correctly under the multi-dimensional risk feature vector and the total number of parameters ;

[0104] S43. Combine the model coding length and the data coding length, and determine the optimal risk evaluation model based on the minimum description length criterion , and the optimal risk evaluation model satisfies the following optimization objective:

[0105] ;

[0106] The optimal risk evaluation model realizes the dynamic balance among the model structure complexity, prediction accuracy and the importance of risk weights in the low-altitude aircraft trajectory risk evaluation scenario.

[0107] In this embodiment, S5 includes the following steps:

[0108] S51. Collect the original three-dimensional flight trajectory dataset generated in real time during the operation of the low-altitude aircraft to form a new multi-dimensional risk feature dataset , and input the new multi-dimensional risk feature dataset into the optimal risk evaluation model , predict the risk level of each flight trajectory, and generate a set of flight trajectory risk prediction results ;

[0109] S52. According to the set of flight trajectory risk prediction results , perform flight trajectory risk evaluation output in combination with the preset risk level mapping rule, and the risk level mapping rule is set according to the segmented risk level value and the corresponding multi-dimensional risk feature parameter threshold;

[0110] S53. Synchronously feedback the flight trajectory risk evaluation output result to the aircraft control system and the remote monitoring terminal in the form of a risk level label and a warning signal. The warning signal includes the trajectory segmentation position, risk level, trigger parameter and overlimit value, which are used to assist the dynamic adjustment of the flight path and the real-time management of flight safety.

[0111] In this embodiment, the risk level division rule includes:

[0112] When the fractal dimension feature value is greater than the set upper threshold, and the instantaneous flight speed change value or acceleration change value exceeds the corresponding aircraft safety tolerance range, it is determined that the trajectory segment is a high-risk segment, and a high-risk level is output;

[0113] When the fractal dimension eigenvalue is in the middle range, and the change value of the flight attitude angle fluctuates within the normal range, but there is an abnormal fluctuation in a single parameter, it is determined that this trajectory segment is a medium-risk segment, and the medium-risk level is output.

[0114] When the change values of all dynamic flight parameters are within the safety threshold range and the fractal dimension eigenvalue is below the set lower limit range, it is determined that this trajectory segment is a low-risk segment, and the low-risk level is output.

[0115] A risk management system for low-altitude aircraft based on machine learning, which is used to execute a risk management method for low-altitude aircraft based on machine learning, includes the following modules:

[0116] The trajectory data acquisition module is used to acquire the three-dimensional flight trajectory dataset of the aircraft within the specified area.

[0117] The data preprocessing module is used to perform spatio-temporal alignment, outlier removal, and filtering on the flight trajectory dataset to generate the preprocessed flight trajectory dataset.

[0118] The multi-scale feature extraction module is used to segment the flight trajectory based on the adaptive scale and extract the local fractal dimension to form a description vector of the flight trajectory complexity.

[0119] The dynamic parameter fusion module is used to extract the speed change value, acceleration change value, and attitude angle change value of each segmented flight trajectory, and fuse them with the complexity description vector to generate a multi-dimensional risk feature dataset.

[0120] The model selection module is used to select the optimal risk assessment model with the minimum description length from the candidate models based on the minimum description length principle.

[0121] The risk assessment module is used to predict the risk level of the flight trajectory and output the early warning information based on the selected model.

[0122] The feedback module is used to feedback the risk assessment result to the aircraft control system and the monitoring terminal to realize the dynamic adjustment of the flight path and the linkage of risk early warning.

[0123] Example 1: At 07:41 on October 7, 2024, an abnormal low-altitude flight event occurred in the intelligent logistics test area of City A. At 6:55 in the early morning of the same day, an autonomous flying quadcopter drone (aircraft code ZQ-FS20241007A) with the model "Track ZQ-4" was performing a routine low-altitude route inspection task. The flight area was the south-to-north line of Shunde Innovation Port Area, with a total length of about 4.6 kilometers and the altitude controlled between 60 and 75 meters. The task objective was to conduct high-frequency interference detection on the signal coverage of the intelligent distribution channels in the innovation park.

[0124] At 07:41:12, when cruising to the 8th flight segment of the mission (position GPS coordinates: xxx.260138°E, xx.876322°N), the ground control system received an abnormal prompt: The aircraft had three consecutive non-instructional yaw operations within a short period of time. The yaw amplitude of the second time reached 29.6 degrees, far higher than the upper limit of the emergency heading adjustment allowed for this type of aircraft body (22 degrees). At the same time, the flight control backhaul data showed that the instantaneous speed Δv was 5.91 m / s, the acceleration Δa was 7.24 m / s², the attitude angle change Δψ was 16.7°, and the fractal dimension soared to 1.894 within 3 seconds.

[0125] The on-duty administrator immediately invoked the intelligent management system for flight trajectory risks. The system uses the "Risk Management Method for Low-altitude Aircraft Based on Machine Learning" proposed in the present invention to quickly extract features and judge risks for the current flight segment. The processing flow is as follows:

[0126] The system extracts the 49th to 53rd trajectory points from the trajectory data for local processing. Each point contains an accurate timestamp, three-dimensional coordinates, speed, acceleration, and attitude angle change value.

[0127] The preprocessing module first performs time series alignment and sliding window anomaly recognition on the data of these five segments of trajectory points. It is found that the speed change Δv exceeds 4.8 m / s at the 52nd point, and the spatial curvature change rate is severe, so it is judged as a high-order perturbation trajectory.

[0128] The multi-scale analysis module performs segmentation on this segment of the trajectory based on the adaptive function λ(κ,v), and finally forms a local flight segment with a length of about 4.2 seconds, containing 31 micro-trajectory points.

[0129] The system further uses the fractal dimension calculation algorithm to calculate the dimension eigenvalue DB = 1.894 of this segment, which is far higher than the high-risk threshold of 1.73.

[0130] At the same time, the dynamic parameter extraction module obtains: Δv = 5.91 m / s, Δa = 7.24 m / s², Δψ = 16.7°, which together form the multi-dimensional risk feature vector R.

[0131] This risk feature vector is input into the optimal XGBoost risk discrimination model constructed by the present invention in real time. The model outputs a risk level of "high" and locates the risk trigger logic as: (Δa > 6.0) ∧ (DB > 1.73). The system immediately generates the following automated report:

[0132] Event time: 2024-10-07 07:41:12; Aircraft number: ZQ-FS20241007A; Position coordinates: xxx.260138°E, xx.876322°N; Fractal dimension: 1.894; Speed change value: 5.91m / s; Acceleration change value: 7.24m / s²; Attitude angle change: 16.7°; Risk level: high;

[0133] Recommended handling: Initiate emergency adjustment, reduce the flight speed to 2.5m / s, shift the trajectory to 8.7° in the left yaw direction, and remove the interference band.

[0134] The ground administrator manually intervened in the flight control system according to the system's suggestion and implemented path correction operations, completing the control intervention at 07:41:23. The aircraft readjusted its flight attitude as instructed, and no risk warnings appeared in the subsequent flight segment. The mission was successfully completed at 07:58:37.

[0135] In order to verify the system performance of the present invention, we used the case data for comparative experiments, selected a total of 16,870 trajectory segments collected in area A from June to September 2024, and constructed a training set, which included 756 manually annotated high-risk trajectory segments. We compared the traditional static threshold-based method (rule base), the standard random forest method and the machine learning method adopted by the present invention. The specific experimental results are as follows:

[0136] Table 1 Comparison of index data of the method of the present invention, the static threshold method and the standard random forest method

[0137]

[0138] Further analysis found that the traditional rule-based method mainly relies on single-dimensional super-threshold judgment of speed and acceleration, ignoring the trajectory structure information, resulting in a low recognition rate; although the random forest has the ability to judge multi-dimensional features, it is insufficient in modeling the complexity of trajectory morphology. The method of the present invention innovatively integrates trajectory complexity and dynamic flight characteristics, constructs a stable and robust risk feature model, and can maintain a high recognition rate and timely response capability even in a high-disturbance environment.

[0139] Therefore, judging from the entire process of this real event, the method proposed in the present invention can accurately capture abnormal low-altitude flight status, quickly analyze potential risks and make intervention suggestions, which greatly improves the safety and mission success rate of aircraft in complex low-altitude environments, and has extremely high practicality and promotion value.

[0140] The present invention introduces the local fractal dimension of the flight trajectory as the core descriptor into the risk management system. By extracting the fractal dimension features from the trajectory segments after multi-scale segmentation, it effectively characterizes the local complexity and irregularity of the flight trajectory at different scales. In addition, through vector concatenation and fusion with dynamic flight parameters such as the instantaneous speed change value, acceleration change value, and attitude angle change value, a multi-dimensional risk feature vector is formed, significantly enhancing the sensitivity of the model to minor trajectory perturbations and potential risk signs, and improving the accurate recognition rate of high-risk trajectory segments.

[0141] The present invention breaks through the traditional way of relying on manually setting the model structure or fixed templates. It uses the minimum description length principle in information theory to jointly optimize and evaluate the coding complexity and data error of the candidate model set, comprehensively considering the information loss caused by the model structure complexity and prediction error, and automatically selects the optimal model with a compact structure and high-risk discrimination ability from numerous candidate models, ensuring that the selected model has stable and reliable risk assessment performance in different flight environments and trajectory patterns.

[0142] The present invention designs an adaptive segmentation function jointly driven by the spatial curvature change rate of flight trajectory points and the flight speed deviation degree to achieve non-uniform dynamic segmentation of the flight trajectory. It not only effectively avoids the problems of feature loss or redundancy caused by fixed time window segmentation, but also can accurately capture the high-risk feature segments with high turning rates or rapid speed changes in the trajectory, greatly improving the recognition accuracy of local risk events.

[0143] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, should be covered by the protection scope of the present invention.

Claims

1. A risk management method for low-altitude aircraft based on machine learning, characterized in that, It includes the following steps: S1. Obtain the original dataset of the three-dimensional flight trajectory of the low-altitude aircraft in the specified low-altitude area in real time and perform preprocessing to generate the preprocessed flight trajectory dataset; S2. Perform multi-scale segmentation processing on the preprocessed flight trajectory dataset to form a flight trajectory dataset with different scales, and apply the fractal dimension analysis method to the flight trajectory dataset with different scales to extract the complexity description vector of the flight trajectory; S3. Perform data fusion on the flight trajectory complexity description vector and the dynamic flight parameter data during the operation of the low-altitude aircraft to construct a multi-dimensional risk feature dataset; S4. Based on the minimum description length principle, screen the candidate models of the risk assessment model for the multi-dimensional risk feature dataset. By evaluating the model coding length and data coding length of each candidate model, automatically select the optimal risk assessment model with the minimum description length and the best data fitting accuracy; S5. Use the selected optimal risk assessment model to perform real-time risk assessment on the newly collected flight trajectory dataset and its corresponding multi-dimensional risk feature dataset to generate the flight trajectory risk assessment output; S6. Feed back the flight trajectory risk assessment output to the flight control system and the monitoring terminal to realize dynamic risk monitoring and flight path adjustment, and provide a decision-making basis for the risk management of low-altitude aircraft; The S2 includes the following steps: S21. According to the requirement that the track risk assessment of the low-altitude aircraft is sensitive to the local complexity of the track, construct an adaptive local scale division function based on the joint of the spatial curvature change rate and the average flight speed of the flight track: λ(κ i ,v i ) = l base ·exp(-α·|κ i |)·exp(-β·|v i -v mean |); where λ(κ i , v i ) is the adaptive local scale at the flight trajectory point p i , l base is the preset basic segmentation length, κ i is the spatial curvature change rate at the flight trajectory point p i , reflecting the severity of the spatial turning of the flight trajectory, v i is the instantaneous flight speed of the aircraft at the flight trajectory point p i , v mean is the average flight speed of the flight trajectory, and α and β are parameters for controlling the sensitivity of the adaptive local scale adjustment; S22. Use the adaptive local scale division function λ(κ i , v i ) to perform local dynamic segmentation on the preprocessed flight trajectory dataset T pre to generate a multi-scale flight trajectory dataset T scale . Any flight trajectory segment S in the multi-scale flight trajectory dataset j is defined as: S j = {p i | t j ≤ t i < t j + λ(κ j , v j )}, j = 1, 2, …, M; Among them, the flight trajectory segment S j starts at time t j , and the flight trajectory point p i is the i-th point within the flight trajectory segment S j . M is the total number of flight trajectory segments after segmentation. λ(κ j , v j ) represents the adaptive local scale size dynamically determined based on the current curvature change rate and instantaneous velocity with the starting point p j of the flight trajectory as the reference point; S23. For each flight trajectory segment S j , calculate the local fractal dimension eigenvalue D of the flight trajectory segment S j by using the box dimension algorithm in the fractal dimension analysis method B,j , and obtain the local fractal dimension feature set. The local fractal dimension eigenvalue quantitatively characterizes the local irregularity and structural complexity of the flight trajectory of the low-altitude aircraft within the flight trajectory segment, reflects the local risk characteristics of the flight trajectory, and based on the local fractal dimension feature set D B , construct the flight trajectory complexity description vector C traj .

2. The risk management method for low-altitude aircraft based on machine learning according to claim 1, wherein The S1 includes the following steps: S11. Set the flight monitoring time window [t0, t N for the low-altitude aircraft within the specified low-altitude flight area. Through the inertial measurement unit, global positioning system module and attitude sensor deployed on the low-altitude aircraft, collect the original flight trajectory data in real time and construct the three-dimensional original flight trajectory dataset T raw . The three-dimensional original flight trajectory dataset T raw consists of a number of sampling points. Each sampling point includes the three-dimensional space coordinates (x i , y i , z i ) of the aircraft at a specific moment and the corresponding timestamp t i ; S12. Perform time series alignment processing on the original three-dimensional flight trajectory dataset T raw by resampling the time axis of the flight trajectory points so that the time interval between all adjacent sampling points remains a fixed time difference, obtaining the flight trajectory dataset T after time normalization align ; S13. The flight trajectory dataset T after time standardization align Perform outlier removal processing. Use the speed change and position deviation within the sliding window for judgment. When the speed change of the flight trajectory point within the local window exceeds the speed change threshold or the spatial position deviation of the flight trajectory point exceeds the position deviation threshold, then determine the flight trajectory point as an outlier and remove it from the flight trajectory. After processing, generate the flight trajectory dataset T after outlier removal clean ; S14. The flight trajectory dataset T after abnormal point rejection clean is subjected to noise filtering. The sliding window mean filtering method is used to smooth the flight trajectory points. The mean value is calculated using a fixed number of adjacent front and rear flight trajectory points at each flight trajectory point position, replacing the original flight trajectory point position to generate the preprocessed flight trajectory dataset T pre .

3. A risk management method for low-altitude aircraft based on machine learning according to claim 1, characterized in that The S3 includes the following steps: S31. For each sampling point in the preprocessed flight trajectory dataset T pre obtain the instantaneous flight speed v i , the instantaneous acceleration a i of the aircraft, and the flight attitude angle ψ i , and calculate the instantaneous flight speed change value Δv i , the instantaneous acceleration change value Δa i of the aircraft, and the flight attitude angle change value Δψ i based on the differences in parameters between adjacent sampling points, thus forming the dynamic flight parameter dataset T dyn ; S32. For each flight trajectory segment S j , use the sliding window statistical method to extract from the dynamic flight parameter dataset T dyn the aggregated dynamic flight parameter vector D j corresponding to the flight trajectory segment S j . In this aggregated dynamic flight parameter vector D j , the instantaneous flight speed change value Δv j is the difference between the maximum instantaneous speed and the minimum instantaneous speed of each sampling point within the flight trajectory segment S j . The aircraft instantaneous acceleration change value Δa j is the difference between the maximum instantaneous acceleration and the minimum instantaneous acceleration of each sampling point within the flight trajectory segment S j . The flight attitude angle change value Δψ j is the average value of the attitude angle change values of consecutive sampling points within the flight trajectory segment S j ; S33. Cascading the flight trajectory complexity description vector C traj with the aggregated dynamic flight parameter vector D j of the corresponding flight trajectory segment S j to perform vector - level cascaded data fusion to form a multi - dimensional risk feature vector R j . The multi - dimensional risk feature vector simultaneously includes the complexity description of the flight trajectory at each local scale and the instantaneous flight speed change value, the instantaneous acceleration change value of the aircraft, and the flight attitude angle change value within the flight trajectory segment S j ; S34. Segment all flight trajectories into segments S j The corresponding multi-dimensional risk feature vector R j Combine them in sequence according to the sampling order to form a multi-dimensional risk feature dataset T risk .

4. A risk management method for low-altitude aircraft based on machine learning according to claim 1, characterized in that The S4 includes the following steps: S41. Construct a set M of candidate risk assessment models cand , each candidate model M in the set of candidate risk assessment models k is defined as a prediction function that maps a multi-dimensional risk feature data set T risk to the corresponding risk level label set Y = {y j | j = 1, 2, …, M}: where, R j is the multi-dimensional risk feature vector of the j-th flight trajectory segment, is the predicted risk level prediction value corresponding to the candidate model M k , θ k is the candidate model parameter vector, and f k represents the decision mapping structure of the candidate model M k . S42. Calculate the model encoding length and data encoding length for the candidate model M k The model encoding length L(M k ) combined with the complexity of the model structure and the metric information complexity of the parameter vector is defined as: L(M k ) = γ1·|θ k | + γ2·dim(f k ); where |θ k | represents the total number of parameters included in the candidate model M k , dim(f k ) represents the structural depth of the decision mapping structure f k or the number of piecewise functions, and γ1, γ2 are preset structural penalty factors; The data encoding length L(T risk |M k ) is defined as the total encoding loss of the prediction error of the candidate model M k for all multi-dimensional risk feature data points, combined with the true risk level label y j and the predicted value of the risk level expressed in a weighted error encoding manner as: Among them, ω j is the prior importance weight of the risk label, is the probability value of the model predicting correctly under the multi-dimensional risk feature vector R j and the total number of parameters θ k ; S43. Combine the encoding length of the comprehensive model and the encoding length of the data, and determine the optimal risk assessment model M based on the minimum description length criterion opt .

5. A risk management method for low-altitude aircraft based on machine learning according to claim 1, characterized in that, The S5 includes the following steps: S51. Collect the original 3D flight trajectory dataset generated in real time during the operation of low-altitude aircraft to form a new multi-dimensional risk feature dataset Input the new multi-dimensional risk feature dataset into the optimal risk assessment model M opt to predict the risk level for each flight trajectory and generate a set of flight trajectory risk prediction results S52. According to the set of flight trajectory risk prediction results Perform flight trajectory risk assessment output in combination with the preset risk level mapping rule, and the risk level mapping rule is set according to the segmented risk level value and the corresponding multi-dimensional risk characteristic parameter threshold; S53. Synchronously feed back the flight trajectory risk assessment output result to the aircraft control system and the remote monitoring terminal in the form of a risk level label and a warning signal. The warning signal includes the track segment position, risk level, trigger parameter and over-limit value, which are used to assist the dynamic adjustment of the flight path and the real-time management of flight safety.

6. The risk management method for low-altitude aircraft based on machine learning according to claim 5, characterized in that, The risk level division rules include: When the fractal dimension eigenvalue is greater than the set upper threshold, and the instantaneous flight speed change value or acceleration change value exceeds the safety tolerance range of the corresponding aircraft, it is determined that this track segment is a high-risk segment, and a high-risk level is output; When the fractal dimension eigenvalue is in the middle range, and the flight attitude angle change value is within the normal fluctuation range, but there is an abnormal fluctuation in a single parameter, it is determined that this track segment is a medium-risk segment, and a medium-risk level is output; When the change values of all dynamic flight parameters are within the safety threshold range and the fractal dimension eigenvalue is below the set lower limit range, it is determined that this track segment is a low-risk segment, and a low-risk level is output.

7. A risk management system for low-altitude aircraft based on machine learning, which is used to execute a risk management method for low-altitude aircraft based on machine learning according to any one of claims 1-6, characterized in that, It includes the following modules: A track data acquisition module, which is used to acquire the three-dimensional flight trajectory dataset of the aircraft in the specified area; A data preprocessing module, which is used to perform spatio-temporal alignment, abnormal point removal and filtering processing on the flight trajectory dataset to generate the preprocessed flight trajectory dataset; A multi-scale feature extraction module, which is used to segment the flight trajectory based on the adaptive scale and extract the local fractal dimension to form the flight trajectory complexity description vector; A dynamic parameter fusion module, which is used to extract the velocity change value, acceleration change value and attitude angle change value of each segmented flight trajectory, and fuse them with the complexity description vector to generate a multi-dimensional risk feature data set; A model selection module, which is used to select the optimal risk assessment model with the minimum description length from the candidate models based on the minimum description length principle; A risk assessment module, which is used to predict the risk level of the flight trajectory and output early warning information based on the selected model; A feedback module, which is used to feedback the risk assessment result to the aircraft control system and the monitoring terminal to realize the dynamic adjustment of the flight path and the linkage of risk early warning.

Images