Deep learning-based close-range parallel runway wake flow dynamic prediction method

Abstract

The invention relates to the technical field of trailing vortex dynamic monitoring, in particular to a short-distance parallel runway wake flow dynamic prediction method based on deep learning, which comprises the following steps of: acquiring disturbance energy level and coordinate direction and fitting propagation state, extracting disturbance source features and combining with spatial direction constraint, fitting boundaries according to propagation included angle and speed change rate and classifying, and obtaining a dynamic prediction result. Acquiring a track coordinate course angle and a flight speed, calculating a track vector and a boundary normal included angle, analyzing an interference risk factor in combination with a ratio of the speed to a propagation direction, dividing disturbance levels, and outputting a preliminary prediction result. According to the method, the state vector and the disturbance data cooperatively extract features and constrain the direction, the disturbance source positioning precision is improved, the boundary is described by combining the speed change rate and the propagation included angle, the recognition robustness is enhanced, the risk level is evaluated by fusing multi-dimensional parameters, the disturbance intensity is dynamically corrected, the track deviation is introduced, and the prediction accuracy and the response efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of wake dynamic monitoring technology, and in particular to a method for dynamic prediction of wake vortices from close-range parallel runways based on deep learning. Background Technology

[0002] The field of wake vortex dynamic monitoring technology encompasses technologies related to the detection, modeling, identification, and prediction of wake vortices generated during aircraft operation. Its core lies in the dynamic acquisition and temporal modeling of wake vortex flow field characteristics generated by aircraft during takeoff, landing, or flight through sensing equipment and data processing technology. This supports air traffic control needs such as trajectory management, flight scheduling, and flight safety early warning. This field primarily involves wake vortex aerodynamic characteristic modeling, wake spatial distribution monitoring, and multi-source meteorological and flight data fusion modeling. Combining physical modeling and data-driven prediction mechanisms, a dynamic description system of wake vortex evolution is formed, supporting flight conflict avoidance and operational efficiency optimization. It is widely applied in the operational monitoring and risk control aspects of civil aviation air traffic control systems.

[0003] Among them, the deep learning-based wake dynamic prediction method for close-range parallel runways refers to a method for predicting the wake interference problem of aircraft under conditions of small distance between parallel runways by constructing a deep neural network model to predict the wake trajectory and intensity in a time series. This method involves extracting core wake parameters based on historical flight radar data and high-frequency meteorological information, using graph convolutional networks to learn the structural dependency between aircraft spatial position information and wake propagation path, combining long short-term memory networks to dynamically model the wake disturbance sequence, and constructing a nested model based on a gating mechanism to achieve multi-step prediction of wake spatial distribution. This method explicitly uses the aircraft's three-dimensional trajectory point set and real-time wind profile as input to complete the wake path time series reasoning and interference area distribution mapping.

[0004] Existing technologies primarily focus on modeling and predicting aircraft wake trajectories, generally using historical radar and meteorological data as the main inputs for wake parameter extraction and employing neural networks to model the evolution trajectory of disturbances. However, they exhibit significant shortcomings in the precise characterization of spatial boundaries and disturbance risk assessment. Most existing models extrapolate disturbance regions based on the wake center path, lacking accurate characterization of the spatial diffusion boundary of disturbance intensity. This can easily lead to misidentification of weak interference areas outside the boundary, resulting in misjudgments or missed detections. Risk assessment is also limited in scope, relying solely on the matching relationship between aircraft trajectory points and wind field data, ignoring the dynamic changes in the spatial relationship between disturbances and tracks, making accurate early warnings difficult in high-frequency changing environments. Even in areas where multiple aircraft converge, and where the disturbance boundary and track direction are asymmetrical, existing methods struggle to accurately determine risk levels, easily exhibiting response lag issues. Furthermore, their ability to correct for disturbance evolution trends is weak. Models generally lack a coupling mechanism for dynamic adjustment of disturbance intensity and track deviation analysis when predicting future wake behavior, resulting in decreased prediction accuracy and an inability to effectively support flight safety scheduling under complex meteorological conditions. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a deep learning-based method for dynamic prediction of wake turbulence in close-range parallel runways.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a deep learning-based method for dynamic prediction of wake turbulence in close-range parallel runways, comprising the following steps:

[0007] S1: Collect and organize the disturbance energy level, center coordinates, propagation direction and time step of the aircraft wake in the close parallel runway, fit the wake disturbance propagation process through a convolutional neural network and analyze the wake disturbance state vector to generate disturbance propagation state data.

[0008] S2: Based on the disturbance propagation state data, the disturbance state vector is input into the residual network algorithm to extract the disturbance source features, and constraint calculations are performed by combining the wake disturbance center coordinates and the diffusion direction to generate the disturbance source feature vector;

[0009] S3: Based on the feature vector of the disturbance source and the disturbance propagation state data, fit the disturbance space boundary by the angle relationship between the propagation direction and the velocity change in the disturbance propagation coordinates, calculate the disturbance velocity change rate, classify the disturbance boundary, and generate disturbance region boundary data;

[0010] S4: Based on the boundary data of the disturbance zone, obtain the corresponding aircraft track coordinates, heading angle and velocity, calculate the angle between the track vector and the normal of the disturbance space boundary, analyze the interference risk factors and classify the disturbance risk level by combining the ratio of the aircraft speed to the propagation direction, and generate preliminary prediction results of wake interference.

[0011] S5: Based on the preliminary prediction results of the wake interference, obtain the corresponding disturbance energy level in the disturbance boundary as the disturbance intensity, calculate the continuous change rate of the disturbance boundary and the stability of the track direction, and correct the disturbance intensity. Combine the track deviation to re-judge the disturbance risk level and predict the evolution trend of the wake disturbance, and output the dynamic prediction results of the wake of the close-range parallel runway.

[0012] As a further aspect of the present invention, the disturbance propagation state data includes a disturbance energy level sequence, a disturbance propagation path, and a state vector set; the disturbance source feature vector specifically includes a disturbance intensity scalar, a center point coordinate set, and a direction constraint vector; the disturbance zone boundary data includes a spatial boundary coordinate set, a disturbance velocity change rate value, and a boundary type label; the preliminary prediction result of the wake interference specifically includes a risk level distribution, the location of the track intersection point, and the interference factor index value; and the dynamic prediction result of the wake of the close-range parallel runway includes a disturbance level change trend, a disturbance intensity correction sequence, and a risk level sequence within a time period.

[0013] As a further aspect of the present invention, the specific steps of S1 include:

[0014] S101: Collect the energy level, disturbance center coordinates, propagation direction and time step of the wake disturbance of the aircraft in the close parallel runway. Pair the disturbance energy level with the time step, perform vector transformation and normalization on the disturbance center coordinates and propagation direction, merge the data according to the time step order, and generate the wake disturbance sequence.

[0015] S102: Based on the wake perturbation sequence, the sequence is divided into time segments with fixed step lengths by a convolutional neural network. The perturbation energy level in each segment is convolved to calculate the local trend features. The features are then fused with the corresponding propagation direction vector to generate the perturbation trend change rate.

[0016] S103: Based on the disturbance trend change rate, combined with the propagation direction and disturbance center changes of multiple time segments, construct a disturbance propagation state vector sequence and perform time axis extension prediction to obtain the expansion path of the disturbance in the propagation direction and the degree of disturbance intensity fluctuation, and generate disturbance propagation state data.

[0017] As a further aspect of the present invention, the specific steps of S2 include:

[0018] S201: Based on the disturbance propagation state data, the disturbance state vector in the disturbance propagation state data is calculated through a residual network. The response feature values ​​of the disturbance state vector in the temporal and spatial dimensions are extracted and made into jump connections. The nonlinear responses of multiple levels are integrated to generate a disturbance response feature value distribution map.

[0019] S202: Based on the disturbance response characteristic value distribution map, combined with the spatial position parameters of the wake disturbance center coordinates, the change trend of the disturbance response characteristic value in the central region is determined by the sensitivity coefficient threshold and directional analysis is performed. The distribution structure of the disturbance intensity along the diffusion direction is extracted, and the directional normalized vector group is obtained.

[0020] S203: Call the direction normalization vector group and the disturbance response feature value distribution map, construct the projection mapping of the disturbance feature space by the disturbance diffusion direction and the wake center coordinates, calculate the density aggregation rate of the response values ​​of multiple feature channels and complete the coordinate adjustment, and generate the disturbance source feature vector.

[0021] As a further aspect of the present invention, the sensitivity coefficient threshold is determined by taking the coordinates of the disturbance center in the disturbance intensity distribution map as the base point, selecting a set of disturbance response values ​​within a fixed radius, calculating the mean and standard deviation of the set respectively, and weighting them to form the upper bound of the interval as the sensitivity coefficient threshold.

[0022] As a further aspect of the present invention, the specific steps of S3 include:

[0023] S301: Based on the feature vector of the disturbance source and the disturbance propagation state data, extract the propagation direction vector and velocity in the disturbance propagation coordinates, calculate the angle between the velocity change vector and the propagation direction vector, smooth the trend of the angle in the spatial distribution, and generate the disturbance angle distribution result.

[0024] S302: Call the disturbance angle distribution results, calculate the velocity change rate formed by dividing the velocity difference between propagation velocities by the time interval, determine the boundary change trend segment of the disturbance diffusion area based on the fluctuation gradient of the velocity change rate, and obtain the disturbance velocity change rate distribution map;

[0025] S303: Based on the disturbance velocity change rate distribution map and the disturbance angle distribution curve set, the disturbance space is divided into multiple boundary blocks, the angle variation range and velocity change gradient value within the block are called, the disturbance boundary judgment threshold is set, and clustering classification is performed on multiple blocks to generate disturbance area boundary data.

[0026] As a further aspect of the present invention, the disturbance boundary determination threshold is constructed by using the mean of the angle change interval in the disturbance angle distribution curve group and the median value of the gradient amplitude in the disturbance velocity change rate distribution diagram to construct a bivariate joint interval, and the upper limit of the joint interval is used as the determination threshold.

[0027] As a further aspect of the present invention, the specific steps of S4 include:

[0028] S401: Based on the boundary data of the disturbance zone, obtain the flight track coordinates, heading angle and velocity of the aircraft at the corresponding time, construct the flight track direction vector in multiple time periods and analyze the corresponding normal vector with the disturbance boundary surface, calculate and arrange the angle between the flight track direction and the boundary, and generate a sequence of flight track boundary angles.

[0029] S402: Call the track boundary angle sequence, compare the magnitude of the corresponding propagation direction in the disturbance propagation state data, calculate the ratio between the aircraft speed and the disturbance propagation speed, and combine the angle value to determine the degree of deviation of the aircraft's running direction, and obtain the interference risk factor set;

[0030] S403: Based on all ratio points and angle points in the interference risk factor set, construct a two-dimensional joint numerical distribution map, classify the interference risk factor set into different levels according to intervals, assign risk level labels accordingly and mark them to multiple track time nodes, and generate preliminary prediction results for wake interference.

[0031] As a further aspect of the present invention, the specific steps of S5 include:

[0032] S501: Based on the preliminary prediction results of the wake interference, obtain the disturbance energy level of the corresponding region in the disturbance boundary, collect the disturbance intensity value sequence in a continuous time period, calculate the ratio of the disturbance intensity difference to the time interval in adjacent time periods, calculate the continuous change rate of the disturbance boundary, and establish a disturbance intensity change rate sequence.

[0033] S502: Call the disturbance intensity change rate sequence and the track vector data of the corresponding time period of the aircraft, calculate the track direction difference and standardize it to obtain the directional stability in the whole sequence, and use it as an adjustment factor to correct the disturbance intensity and generate a stable corrected disturbance intensity sequence.

[0034] S503: Based on the track deviation between the stable corrected disturbance intensity sequence and the preliminary prediction result of the wake interference, the deviation level boundary is re-judged according to the disturbance risk level, the disturbance risk level interval is divided, a disturbance level change trend map is constructed and superimposed with a time axis, and the dynamic prediction result of the wake of the close-range parallel runway is output.

[0035] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0036] In this invention, disturbance source features are extracted through the collaborative processing of state vectors and propagation state data. Structural relationships within the propagation path are introduced for directional constraints, making disturbance source localization more spatially targeted. The joint calculation of the disturbance velocity change rate and propagation angle makes the spatial boundary description clearer, while probabilistic classification further divides the disturbance intensity region, effectively improving the accuracy and robustness of disturbance region identification. In interference assessment, the track vector and the normal angle of the disturbance boundary are combined with the ratio of propagation direction to flight speed to form a multi-dimensional interference risk index system. This enriches the representation of the impact of wake vortices on aircraft and enhances the adaptability of risk level classification. During risk level correction, the dynamic changes in disturbance intensity are combined with track direction stability, supplemented by the continuous change rate of the disturbance boundary for interference intensity adjustment. Simultaneously, track deviation is introduced for risk reassessment, improving prediction accuracy under dynamic disturbance region conditions. In the above process, driven by multiple factors such as spatial direction quantity, disturbance velocity gradient, and trajectory behavior stability, not only is the accuracy of wake boundary identification optimized, but the time series response capability of disturbance assessment is also enhanced. This provides more continuous and dynamic information support for the short- and medium-term prediction of wake vortex evolution trends, effectively improving flight safety assurance and conflict avoidance efficiency. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the main steps of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0039] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0040] Please see Figure 1 This invention provides a technical solution: a deep learning-based method for dynamic prediction of wake turbulence in close-range parallel runways, comprising the following steps:

[0041] S1: Collect and organize the disturbance energy level, center coordinates, propagation direction and time step of the aircraft wake in the close parallel runway, fit the wake disturbance propagation process through a convolutional neural network and analyze the wake disturbance state vector to generate disturbance propagation state data.

[0042] S2: Based on the disturbance propagation state data, the disturbance state vector is input into the residual network algorithm to extract the disturbance source features. Constraint calculations are performed by combining the wake disturbance center coordinates and the diffusion direction to generate the disturbance source feature vector.

[0043] S3: Based on the feature vector of the disturbance source and the disturbance propagation state data, the disturbance space boundary is fitted by the angle relationship between the propagation direction and the velocity change in the disturbance propagation coordinates, the disturbance velocity change rate is calculated, the disturbance boundary is classified, and the disturbance region boundary data is generated.

[0044] S4: Based on the boundary data of the disturbance zone, obtain the corresponding aircraft track coordinates, heading angle and velocity, calculate the angle between the track vector and the normal of the disturbance space boundary, analyze the interference risk factors and classify the disturbance risk level by combining the ratio of the aircraft speed to the propagation direction, and generate preliminary prediction results of wake interference.

[0045] S5: Based on the preliminary prediction results of wake interference, obtain the corresponding disturbance energy level in the disturbance boundary as the disturbance intensity, calculate the continuous change rate of the disturbance boundary and the stability of the track direction, and correct the disturbance intensity. Combine the track deviation to re-judge the disturbance risk level and predict the evolution trend of the wake disturbance, and output the dynamic prediction results of the wake of the near-parallel runway.

[0046] The disturbance propagation state data includes the disturbance energy level sequence, the disturbance propagation path, and the state vector set. The disturbance source feature vector specifically includes the disturbance intensity scalar, the center point coordinate set, and the direction constraint vector. The disturbance zone boundary data includes the spatial boundary coordinate set, the disturbance velocity change rate value, and the boundary type label. The preliminary prediction results of the wake interference specifically include the risk level distribution, the location of the track intersection point, and the interference factor index. The dynamic prediction results of the wake of the close-range parallel runways include the disturbance level change trend, the disturbance intensity correction sequence, and the risk level sequence within the time period.

[0047] Please see Figure 1 The specific steps for obtaining S1 are as follows:

[0048] S101: Collect the energy level, disturbance center coordinates, propagation direction and time step of the wake disturbance of the aircraft in the close parallel runway. Pair the disturbance energy level with the time step, perform vector transformation and normalization on the disturbance center coordinates and propagation direction, merge the data according to the time step order, and generate the wake disturbance sequence.

[0049] When collecting wake disturbance data of aircraft on a near-parallel runway, a Doppler lidar was used to acquire three-dimensional wind speed components at a sampling frequency of 10Hz. The geographic coordinates of the disturbance center were recorded with an accuracy of 0.1 meters using a GPS differential positioning system. Attitude angle data were recorded using a nine-axis inertial measurement unit with a time step of 50ms. The wind speed modulus was used as an indicator of the disturbance energy level. During the landing of flight CZ3301, the energy level value v = 8.3 m / s was measured at timestamp t = 125.6s, corresponding to coordinates of 31.1435°N, 121.8052°E. The geographic coordinates were converted to the runway coordinate system, and an XY plane coordinate system was established with the runway threshold as the origin. The measured coordinates were x = 1523m, y = +37m. A vector transformation was performed on the propagation direction angle θ = 215° to obtain the direction vector (cosθ, sinθ).

[0050] θ) = (-0.8192, -0.5736), calculate the vector magnitude. The normalized vector (-0.819, -0.574) is obtained by retaining three decimal places. When merging in order of time step, the data from the five consecutive sampling points from t = 125.5s to 125.9s are arranged at 0.1-second intervals to form a data chain including the energy level value sequence [8.1, 8.3, 8.6, 8.4, 8.2] m / s and the corresponding coordinate vector.

[0051] Table 1. Examples of wake disturbance data

[0052] Timestamp (s) Energy level (m / s) X-coordinate (m) Y coordinate (m) Direction angle (°) 125.5 8.1 1518 +35 214 125.6 8.3 1523 +37 215 125.7 8.6 1529 +39 216

[0053] As shown in Table 1, the energy level value of 8.3 m / s corresponding to the timestamp 125.6 s is mapped to the coordinates (1523, +37). The direction angle of 215° is transformed into a standardized direction vector (-0.819, -0.574). During the data merging stage, the energy level values ​​of the five adjacent time steps are linearly interpolated. When there is a discontinuity in the time interval, cubic spline interpolation is used to fill in the missing data points.

[0054] S102: Based on the wake perturbation sequence, the sequence is divided into time segments with fixed step lengths by a convolutional neural network. The perturbation energy level in each segment is convolved to calculate the local trend features. The features are then fused with the corresponding propagation direction vector to generate the perturbation trend change rate.

[0055] When constructing a convolutional neural network, set the time window length T. w =10 sampling points, number of convolution kernels D=8, width of each convolution kernel k=3, C d The weight parameters representing the d-th convolutional kernel are obtained through pre-training, with values ​​ranging from [1.5, 1.5], reflecting the sensitivity to multiple spatial frequency features. M (l)M represents the normalized perturbation energy level input of the l-th segment, obtained by normalizing the energy level values ​​output from the previous steps using the minmax method. (2) =8.3, standardized value Let represent the magnitude of the convolution kernel in the l-th segment and d-th channel. Using the parameters of convolution kernel number 1 as [0.25, 0.5, 0.25], the magnitude calculation process is as follows: δ represents the denominator correction factor, set to 0.01, to prevent numerical instability caused by small-module-length convolutional kernels. D represents the total number of convolutional kernels, set to 8 based on the number of input channels. α (l) Represents the dynamic adjustment factor, calculated as 1 / ln(N) l +2), N l The dynamic adjustment factor is calculated to determine the number of valid data points in the current segment: when time segment l = 2, including N... l When there are 10 valid data points, N l For the representative data points, a sliding convolution is performed on the energy level values ​​[8.1, 8.3, 8.6, 8.4, 8.2, 8.0, 7.9, 7.8, 7.7, 7.6] of the l=2 time segment. A denominator correction coefficient δ=0.01 is set to prevent division by zero errors. The dynamic adjustment factor α is calculated based on the time segment length. (2) =1 / ln(10+2) =1 / 2.4849 =0.402, then substituting into the local trend characteristics, Taking convolution kernel number 2 as an example, the same calculation is performed on all 8 convolution kernels: Summation: H2 = 1.359 + (-1.314) + ... + 0.883 = 2.147. The sum of the results of the remaining 6 convolutional kernels is 2.102. H2 represents the second convolutional kernel. The result 2.102 is compared with the preset threshold Q = 1.8, which exceeds the threshold by 15.7%, indicating that there is a trend change in the current time segment. In the feature fusion stage, the direction vector (-0.819, -0.574) and the H value are concatenated to generate a 384-dimensional fused feature vector.

[0056] S103: Based on the rate of change of the disturbance trend, combined with the propagation direction and the change of the disturbance center in multiple time segments, construct the disturbance propagation state vector sequence and perform time-axis extension prediction to obtain the expansion path of the disturbance in the propagation direction and the degree of disturbance intensity fluctuation, and generate disturbance propagation state data.

[0057] When constructing the disturbance propagation state vector, the fused feature vectors of three consecutive time segments (l = 1, 2, 3) are concatenated to form a 1152-dimensional state vector. The time-axis extension prediction adopts the ARIMA model, with the autoregressive order p = 5, the difference order d = 1, and the moving average order q = 2. The state changes of three time steps are predicted using the first five state vectors [V-{t-4}, V-{t-3}, V-{t-2}, V-{t-1}, Vt], where t represents the time step. When t = 125.6s, the predicted disturbance center coordinate offset Δx = +5.2m / step, Δy = -1.3m / step, and energy level decay rate γ = 0.15 / s are obtained. The propagation path angle θ = 215 ± 2° is calculated based on the propagation direction vector. The disturbance intensity fluctuation is determined by calculating the standard deviation of the energy level sequence σ = 0.32m / s, which is less than the safety threshold of 0.5m / s, and is therefore judged as a controllable fluctuation.

[0058] Please see Figure 1 The specific steps to obtain S2 are as follows:

[0059] S201: Based on the disturbance propagation state data, the disturbance state vector in the disturbance propagation state data is calculated through the residual network. The response feature values ​​of the disturbance state vector in the temporal and spatial dimensions are extracted and made into jump connections. The nonlinear responses of multiple levels are integrated to generate a disturbance response feature value distribution map.

[0060] Based on the perturbation propagation state data obtained in the previous steps, an input matrix including a 128-dimensional state vector is constructed. The residual network is set up with four residual blocks, each with two convolutional layers. The first layer uses a 3×3 convolutional kernel, and the second layer uses a 1×1 convolutional kernel. The weight parameters are obtained through transfer from an ImageNet pre-trained model. For the input vector V at time t = 125.6s... t =[v1, v2-v 128 ], V t V represents a 128-dimensional perturbation state vector, where v1 represents the state vector V. tThe parameter values ​​of the first dimension are: the first convolution kernel K1 = [0.32, -0.15, 0.28] performs a convolution operation on the first 3 features [8.3, 1523, 37] to obtain 0.32×8.3+(-0.15)×1523+0.28×37 = -215.6, which is then processed by the ReLU activation function to obtain 0. The original input vector is added element-wise to the output vector of the third residual block through a skip connection. When the feature dimensions do not match, a 1×1 convolution is used for dimensionality increase. During feature extraction, the moving average of the state vectors over 10 consecutive time steps is calculated, including the time interval from t = 125.5s to 125.9s. The sequence of 64th-dimensional feature values ​​is [72.3, 73.1, 74.5, 73.8, 72.9], and the mean is calculated as μ = (72.3 + 73.1 + 74.5 + 73.8 + 72.9) / 5 = 73.32. For spatial feature extraction, the feature variance σ of the 8 neighboring coordinate points is calculated for the coordinate parameters x = 1523 and y = 37. 2 =5.7.

[0061] Table 2. Examples of characteristic response value distribution

[0062] Timestamp (s) Feature Dimension Response value X-coordinate (m) Y coordinate (m) 125.6 64 73.1 1523 37 125.6 65 68.4 1523 37 125.6 66 71.2 1523 37

[0063] As shown in Table 2, the response value of feature dimension 64 at time t = 125.6s is 73.1, which is mapped to the spatial coordinates (1523, 37). The 128-dimensional feature is projected to a two-dimensional space by PCA dimensionality reduction, and the principal component weight vector W1 = [0.12, 0.09-0.11] is calculated. The projected coordinates are P = (73.1 × 0.12) + (68.4 × 0.09) + (71.2 × 0.11) = 15.7. When generating the scatter plot, the projected coordinates of 500 time steps are drawn at 0.1-second intervals to form the feature cluster distribution.

[0064] S202: Based on the distribution map of disturbance response characteristic values, combined with the spatial position parameters of the wake disturbance center coordinates, the change trend of disturbance response characteristic values ​​in the central region is determined by the sensitivity coefficient threshold and directional analysis is performed. The distribution structure of disturbance intensity along the diffusion direction is extracted, and the directional normalized vector group is obtained.

[0065] The sensitivity threshold is set as θ = μ + 1.5σ, where μ is the mean of the feature response value and σ is the standard deviation. The response value of feature dimension 64 is calculated during the time period t = 125.0-125.9, yielding a mean μ = 72.8 and a standard deviation σ = 1.2. Therefore, the orientation angle θ = 72.8 + 1.5 × 1.2 = 74.6. When the eigenvalue 74.5 is detected to be close to the threshold at t = 125.7s, it is determined to be a critical state. During orientation resolution, the coordinate changes of three adjacent time steps are calculated. Δx represents the change in position of the disturbance center along the X-axis of the runway coordinate system, Δx = 1529 - 152. 3 = 6m, Δy = 39 - 37 = 2m, direction angle θ = arctan(2 / 6) = 18.4°, Δy represents the change in position of the disturbance center along the y-axis of the runway coordinate system. After normalization, the direction vector (cosθ, sinθ) = (0.949, 0.316) is obtained. When calculating the intensity distribution, 5 sampling points are set at equal intervals along the 18.4° direction with a spacing of 10m. The measured energy level values ​​are [8.6, 8.3, 8.1, 7.9, 7.7] m / s. The attenuation rate k = (7.7 - 8.6) / 40 = -0.0225 m / s·m is obtained by fitting. -1 .

[0066] S203: Call the direction normalization vector group and the disturbance response feature value distribution map, construct the projection mapping of the disturbance feature space through the disturbance diffusion direction and the wake center coordinates, calculate the density aggregation rate of the response values ​​of multiple feature channels and complete the coordinate adjustment, and generate the disturbance source feature vector.

[0067] When constructing the polar coordinate system, the perturbation center (1523, 37) is used as the origin. The direction vector group is converted into an angle value θ = 18.4°. The radial resolution Δr = 10m and the angular resolution Δθ = 5° are set. The density aggregation calculation uses the Epanechnikov kernel function with a bandwidth h = 15m. The density contribution value is calculated for the coordinate point (1533, 39) at a distance from the center point. The kernel function value K(d / h) = 0.75 × (1 - (10.2 / 15)) 2 =0.75×0.514=0.385, the density value of the three sampling points in the cumulative direction is ΣK=0.385+0.421+0.398=1.204. When adjusting the coordinates, the original X coordinate 1523 is projected according to the direction angle to obtain the new coordinate x′=1523+10×cos18.4°=1523+9.5=1532.5m. The generated feature vector includes the mapping relationship between the processed coordinates (1532.5, 38.1) and the density value 1.204.

[0068] Please see Figure 1 The specific steps to obtain S3 are as follows:

[0069] S301: Based on the feature vector of the disturbance source and the disturbance propagation state data, extract the propagation direction vector and velocity in the disturbance propagation coordinates, calculate the angle between the velocity change vector and the propagation direction vector, smooth the trend of the angle in the spatial distribution, and generate the disturbance angle distribution result.

[0070] Based on the obtained feature vector of the disturbance source, This represents the velocity change vector, calculated by the difference between the velocity vectors at two consecutive time steps. It includes the velocity vector at t = 125.6s. At t = 125.7 s but Representing adjacent velocity vectors, the magnitude is calculated as follows: θ0 represents the original included angle, calculated using the vector dot product formula. The propagation direction vector is taken. and dot product of unit vectors Substituting the numerical values, we get θ0 = arccos(-0.583) = 124°, where k represents the spatial smoothing coefficient, based on the velocity standard deviation σ. v =1.2m / s and directional stability coefficient β=0.7, calculate: k=σ v ×β 1.5 =1.2 × 0.7 1.5 =0.704, v i Representing the velocity change vector component, taking the three-dimensional spatial coordinate component v x =0.3m / s, v y =6m / s, v z = 2m / s, n=3 represents the spatial dimension (x, y, z), substituting into the formula Numerator = 6.33 × 124 + 0.704 × (|0.3| + |6| + |2|) = 791.76, Denominator = 6.33 + 0.704 = 7.034 = 112.6°, θ = 791.76 / 7.034 = 112.6°. When the calculation result is compared with the standard safe angle range [100°, 135°] in the aviation field, 112.6° is within the safe range, indicating that the calculation result meets the requirements of engineering applications.

[0071] S302: Call the disturbance angle distribution results, calculate the velocity change rate formed by dividing the velocity difference between propagation velocities by the time interval, determine the boundary change trend segment of the disturbance diffusion area based on the fluctuation gradient of the velocity change rate, and obtain the disturbance velocity change rate distribution map;

[0072] When calculating the rate of change of velocity, we take the velocity difference between t = 125.6s and 125.7s as Δv = 0.4 - 0.3 = 0.1 m / s, and the time interval as Δt = 0.1s, thus obtaining the rate of change a = Δv / Δt = 1.0 m / s.2 The fluctuation gradient was calculated using the five-point difference method, with the rate of change [0.9, 1.0, 1.1, 1.0, 0.8] m / s over five consecutive time steps. 2 Calculate the second derivative to get When a region is identified as stable, boundary trend identification is performed when the absolute value of the gradient at three consecutive points exceeds 0.5 m / s. 3 The boundary was marked as a mutation boundary, including the gradient value sequence [0.6, 0.7, 0.8] during the period from t = 125.3s to 125.5s, triggering the boundary identification. When generating the distribution map, the rate of change of velocity was set to 0.1 m / s. 2 Color levels are divided at intervals, with the red area >1.2m / s. 2 The area marked as high-risk is blue, with a speed of <0.8 m / s. 2 Marked as a stable region.

[0073] S303: Based on the distribution map of the rate of change of the disturbance velocity and the distribution curve of the disturbance angle, the disturbance space is divided into multiple boundary blocks. The angle variation range and velocity change gradient value within the block are called, the disturbance boundary judgment threshold is set, and clustering classification is performed on multiple blocks to generate disturbance area boundary data.

[0074] For spatial partitioning, the disturbance area was divided into 36 blocks using a 50m × 50m grid, with the runway centerline as the reference. The calculated angle variation range within block B5 was 112.5° and 114.5°, with a mean velocity gradient μ = 0.9 m / s. 2 The threshold value is set as Q = μ + 2σ = 0.9 + 2 × 0.2 = 1.3 m / s 2 σ represents the standard deviation of the velocity gradient. Cluster analysis is performed on the 8 sampling points within block B5. When 4 points exceed the threshold, they are marked as active blocks, including point P1 with a velocity gradient of 1.4 m / s. 2 The included angle is 113.2°, and it is close to the adjacent point P2 (1.35m / s). 2 Euclidean distance (113.5°) Clusters smaller than 0.1 are grouped into the same category, generating three types of boundary data. Category I blocks consist of 12 grids with an average velocity gradient of 1.2 m / s. 2 The included angle fluctuates within ±1.5°.

[0075] Please see Figure 1 The specific steps to obtain S4 are as follows:

[0076] S401: Based on the boundary data of the disturbance zone, obtain the trajectory coordinates, heading angle and velocity of the aircraft at the corresponding time, construct the trajectory direction vector in multiple time periods and analyze the corresponding normal vector with the disturbance boundary surface, calculate and arrange the angle between the trajectory direction and the boundary, and generate a sequence of trajectory boundary angles.

[0077] Based on the acquired disturbance zone boundary data, the track coordinate sequence of flight CZ3301 during the period from t = 125.5s to 125.7s was extracted. The coordinate changes from 31.1435°N to 31.1442°N and 121.8052°E ​​to 121.8048°E were recorded at 0.1-second intervals and converted to the runway coordinate system, resulting in X coordinates [1518, 1523, 1529]m and Y coordinates [+35, +37, +39]m. The heading angle was calculated using the three-point difference method, calculating the direction vector for three consecutive coordinate points, including the vector at time t = 125.6s. The heading angle θ = arctan(2 / 5) = 21.8°. The velocity is calculated by dividing the displacement modulus by the time interval. (This is the course speed.) When obtaining the boundary normal vector, take the boundary points (1530, 40) and (1535, 38) of the B5 block of the disturbance region, and calculate the boundary vector. normal vector Normalization The included angle is calculated using the vector dot product formula, where the track direction vector is used. Then the included angle φ = arccos(0.928×0.371+0.371×0.928) = arccos(0.688) = 46.6°. When generating the sequence, the included angle values ​​of the three time steps [45.2°, 46.6°, 47.8°] are sorted according to the timestamp.

[0078] Table 3 Examples of Track Boundary Angles

[0079]

[0080]

[0081] As shown in Table 3, the heading angle of 21.8° at t = 125.6s is calculated to be 46.6° with the normal vector (0.371, 0.928). When the direction of the normal vector changes by more than 10°, the boundary recalculation mechanism is triggered, and the normal vector is updated to the mean vector of the adjacent blocks.

[0082] S402: Call the track boundary angle sequence, compare the magnitude of the corresponding propagation direction in the disturbance propagation state data, calculate the ratio between the aircraft speed and the disturbance propagation speed, and combine the angle value to determine the degree of deviation of the aircraft's running direction, and obtain the interference risk factor set;

[0083] Extract the perturbation propagation velocity v at time t = 125.6s. p = 8.3 m / s, aircraft speed v a =54.1m / s, velocity ratio r=v a / v p=6.52, the included angle value φ = 46.6°, the risk factor calculation uses a piecewise function. When r>5 and φ>45°, the risk value R = 0.7r + 0.3φ / 10. Substituting the values, the value is R = 0.7×6.52 + 0.3×46.6 / 10 = 4.564 + 1.398 = 5.962. When the R value of three consecutive time steps exceeds the threshold Q = 5.0, it is marked as a high-risk period, including the R value sequence from t = 125.5s to 125.7s [5.12, 5.96, 6.34], all of which exceed the threshold and are judged as a continuous risk state.

[0084] S403: Based on all ratio points and angle points in the interference risk factor set, construct a two-dimensional joint numerical distribution map, classify the interference risk factor set into intervals and classify the risk level labels accordingly, and mark them to the multi-track time nodes to generate preliminary prediction results of wake interference.

[0085] When constructing a two-dimensional coordinate system, the horizontal axis represents the velocity ratio r (0-10), and the vertical axis represents the included angle φ (0°-90°). 500 data points are plotted as a scatter plot, and risk levels are divided as follows: green area (r<3 and φ<30°), yellow area (3≤r<5 or 30°≤φ<45°), orange area (5≤r<7 or 45°≤φ<60°), and red area (r≥7 or φ≥60°). The point (6.52, 46.6°) falling into the orange area is labeled as risk level 3. When labeling time points, the coordinates (6.52, 46.6) at t=125.6s are connected to the preceding and following time steps to form a broken line. When the slope of the broken line exceeds 1.5, a warning indicator is added. The generated prediction results include four risk level areas, with the red area accounting for 12%, mainly distributed within 3km of the runway entrance.

[0086] Please see Figure 1 The specific steps to obtain S5 are as follows:

[0087] S501: Based on the preliminary prediction results of wake interference, obtain the disturbance energy level of the corresponding region in the disturbance boundary, collect the disturbance intensity value sequence in a continuous time period, calculate the ratio of the disturbance intensity difference to the time interval in adjacent time periods, calculate the continuous change rate of the disturbance boundary, and establish a disturbance intensity change rate sequence.

[0088] Based on the preliminary prediction results of wake interference, the perturbation energy level sequence [8.1, 8.3, 8.6] m / s of the red risk area R3 during the period from t = 125.5s to 125.7s was extracted. The intensity difference between adjacent time steps ΔE1 = 8.3 - 8.1 = 0.2 m / s, the time interval Δt = 0.1s, and the rate of change r1 = ΔE1 / Δt = 2.0 m / s 2 The subsequent difference ΔE2 = 8.6 - 8.3 = 0.3 m / s, and the rate of change r2 = 0.3 / 0.1 = 3.0 m / s2 A rate of change sequence [2.0, 3.0] was established. The boundary continuous rate of change was calculated using a three-point moving average. The mean μ for the sequence [2.0, 3.0, 2.8] was obtained as μ = (2.0 + 3.0 + 2.8) / 3 = 2.6 m / s. 2 When the rate of change of three consecutive points exceeds the threshold Q r =2.5m / s 2 When the change rate exceeds the threshold during the period from t=125.6s to 125.8s [3.0, 2.8, 3.2], it is marked as a rapidly changing boundary and is determined to be a continuously active region.

[0089] Table 4 Examples of Disturbance Intensity Change Rate

[0090]

[0091] As shown in Table 4, the rate of change at t = 125.6 s is 3.0 m / s. 2 The active flag is triggered. When the interval between two adjacent active regions is less than 0.2s, they are merged into the same event. The start time is recorded as 125.6s and the duration is 0.3s.

[0092] S502: Call the disturbance intensity change rate sequence and the track vector data of the corresponding time period of the aircraft, calculate the track direction difference and standardize it to obtain the directional stability in the whole sequence, and use it as an adjustment factor to correct the disturbance intensity and generate a stable corrected disturbance intensity sequence.

[0093] Extract the heading angle sequence [20.3°, 21.8°, 23.1°] of flight CZ3301 from t = 125.5s to 125.7s. Calculate the direction differences Δθ1 = 21.8 - 20.3 = 1.5° and Δθ2 = 23.1 - 21.8 = 1.3°. After standardization, obtain z1 = (1.5 - μ) / (20.3 - 20.3 ... θ ) / σ θ z1 represents the standardized directional difference, where the mean μ θ =1.4°, standard deviation σ θ =0.1°, calculate z1 = (1.5-1.4) / 0.1 = 1.0, z1 and z2 represent the normalized directional difference, directional stability S = 1 / (1+|z1|+|z2|) = 1 / (1+1.0+0.9) = 0.345, when correcting for the disturbance strength, multiply the energy level value of 8.3m / s at t = 125.6s by the stability to get E. adj =8.3 × 0.345 = 2.86 m / s, E adj The value represents the disturbance energy level after stability correction, generating a correction sequence [2.75, 2.86, 3.01] m / s. When the stability is below 0.4, a heading stability warning is triggered.

[0094] S503: Based on the track deviation in the stable correction disturbance intensity sequence and the preliminary prediction results of wake interference, the deviation level boundary is re-judged according to the disturbance risk level, the disturbance risk level interval is divided, a disturbance level change trend map is constructed and superimposed with a time axis, and the dynamic prediction results of the wake of the close-range parallel runway are output.

[0095] When reassessing the risk level, the corrected intensity value of 2.86 m / s is compared with the original threshold range. The original red area threshold E red =3.0m / s, adjusted threshold E′ red =E red ×S=3.0×0.345=1.035m / s, where S represents directional stability. The correction value of 2.86m / s exceeds the new threshold, maintaining the red level. When constructing the trend chart, the horizontal axis represents time (125.5s-125.9s), and the vertical axis represents the correction intensity value (0-5m / s). The data from 500 time steps are plotted as a step chart, with the warning line set at 1.035m / s. When 5 consecutive points exceed the line, a red background is filled. The output results include 3 red warning periods, with the longest lasting 0.4s, corresponding to the runway coordinate intervals X=1520-1540m and Y=35-40m.

[0096] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A deep learning-based method for dynamic prediction of wake turbulence in close-range parallel runways, characterized in that, Includes the following steps: S1: Collect and organize the disturbance energy level, center coordinates, propagation direction and time step of the aircraft wake in the close parallel runway, fit the wake disturbance propagation process through a convolutional neural network and analyze the wake disturbance state vector to generate disturbance propagation state data. S2: Based on the disturbance propagation state data, the disturbance state vector is input into the residual network algorithm to extract the disturbance source features, and constraint calculations are performed by combining the wake disturbance center coordinates and the diffusion direction to generate the disturbance source feature vector; S3: Based on the feature vector of the disturbance source and the disturbance propagation state data, fit the disturbance space boundary by the angle relationship between the propagation direction and the velocity change in the disturbance propagation coordinates, calculate the disturbance velocity change rate, classify the disturbance boundary, and generate disturbance region boundary data; S4: Based on the boundary data of the disturbance zone, obtain the corresponding aircraft track coordinates, heading angle and velocity, calculate the angle between the track vector and the normal of the disturbance space boundary, analyze the interference risk factors and classify the disturbance risk level by combining the ratio of the aircraft speed to the propagation direction, and generate preliminary prediction results of wake interference.

2. The method for dynamic prediction of wake turbulence in close-range parallel runways based on deep learning according to claim 1, characterized in that, The disturbance propagation state data includes a disturbance energy level sequence, a disturbance propagation path, and a state vector set. The disturbance source feature vector specifically includes a disturbance intensity scalar, a center point coordinate set, and a direction constraint vector. The disturbance zone boundary data includes a spatial boundary coordinate set, a disturbance velocity change rate value, and a boundary type label. The preliminary prediction results of the wake interference specifically include a risk level distribution, the location of the track intersection point, and an interference factor index.

3. The method for dynamic prediction of wake turbulence in close-range parallel runways based on deep learning according to claim 1, characterized in that, The specific steps of S1 include: S101: Collect the energy level, disturbance center coordinates, propagation direction and time step of the wake disturbance of the aircraft in the close parallel runway. Pair the disturbance energy level with the time step, perform vector transformation and normalization on the disturbance center coordinates and propagation direction, merge the data according to the time step order, and generate the wake disturbance sequence. S102: Based on the wake perturbation sequence, the sequence is divided into time segments with fixed step lengths by a convolutional neural network. The perturbation energy level in each segment is convolved to calculate the local trend features. The features are then fused with the corresponding propagation direction vector to generate the perturbation trend change rate. S103: Based on the disturbance trend change rate, combined with the propagation direction and disturbance center changes of multiple time segments, construct a disturbance propagation state vector sequence and perform time axis extension prediction to obtain the expansion path of the disturbance in the propagation direction and the degree of disturbance intensity fluctuation, and generate disturbance propagation state data.

4. The method for dynamic prediction of wake turbulence in close-range parallel runways based on deep learning according to claim 1, characterized in that, The specific steps of S2 include: S201: Based on the disturbance propagation state data, the disturbance state vector in the disturbance propagation state data is calculated through a residual network. The response feature values ​​of the disturbance state vector in the temporal and spatial dimensions are extracted and made into jump connections. The nonlinear responses of multiple levels are integrated to generate a disturbance response feature value distribution map. S202: Based on the disturbance response characteristic value distribution map, combined with the spatial position parameters of the wake disturbance center coordinates, the change trend of the disturbance response characteristic value in the central region is determined by the sensitivity coefficient threshold and directional analysis is performed. The distribution structure of the disturbance intensity along the diffusion direction is extracted, and the directional normalized vector group is obtained. S203: Call the direction normalization vector group and the disturbance response feature value distribution map, construct the projection mapping of the disturbance feature space by the disturbance diffusion direction and the wake center coordinates, calculate the density aggregation rate of the response values ​​of multiple feature channels and complete the coordinate adjustment, and generate the disturbance source feature vector.

5. The deep learning-based wake dynamic prediction method for near-distance parallel runways according to claim 4, characterized in that, The sensitivity coefficient threshold is determined by using the coordinates of the disturbance center in the disturbance intensity distribution map as the base point, selecting a set of disturbance response values ​​within a fixed radius, calculating the mean and standard deviation of the set respectively, and then weighting them to form the upper bound of the interval as the sensitivity coefficient threshold.

6. The method for dynamic prediction of wake turbulence in close-range parallel runways based on deep learning according to claim 1, characterized in that, The specific steps of S3 include: S301: Based on the feature vector of the disturbance source and the disturbance propagation state data, extract the propagation direction vector and velocity in the disturbance propagation coordinates, calculate the angle between the velocity change vector and the propagation direction vector, smooth the trend of the angle in the spatial distribution, and generate the disturbance angle distribution result. S302: Call the disturbance angle distribution results, calculate the velocity change rate formed by dividing the velocity difference between propagation velocities by the time interval, determine the boundary change trend segment of the disturbance diffusion area based on the fluctuation gradient of the velocity change rate, and obtain the disturbance velocity change rate distribution map; S303: Based on the disturbance velocity change rate distribution map and the disturbance angle distribution curve set, the disturbance space is divided into multiple boundary blocks, the angle variation range and velocity change gradient value within the block are called, the disturbance boundary judgment threshold is set, and clustering classification is performed on multiple blocks to generate disturbance area boundary data.

7. The deep learning-based wake dynamic prediction method for near-distance parallel runways according to claim 6, characterized in that, The disturbance boundary determination threshold is constructed by using the mean of the angle change interval in the disturbance angle distribution curve group and the median value of the gradient amplitude in the disturbance velocity change rate distribution diagram to construct a bivariate joint interval, and the upper limit of the joint interval is used as the determination threshold.

8. The method for dynamic prediction of wake turbulence in close-range parallel runways based on deep learning according to claim 1, characterized in that, The specific steps of S4 include: S401: Based on the boundary data of the disturbance zone, obtain the flight track coordinates, heading angle and velocity of the aircraft at the corresponding time, construct the flight track direction vector in multiple time periods and analyze the corresponding normal vector with the disturbance boundary surface, calculate and arrange the angle between the flight track direction and the boundary, and generate a sequence of flight track boundary angles. S402: Call the track boundary angle sequence, compare the magnitude of the corresponding propagation direction in the disturbance propagation state data, calculate the ratio between the aircraft speed and the disturbance propagation speed, and combine the angle value to determine the degree of deviation of the aircraft's running direction, and obtain the interference risk factor set; S403: Based on all ratio points and angle points in the interference risk factor set, construct a two-dimensional joint numerical distribution map, classify the interference risk factor set into different levels according to intervals, assign risk level labels accordingly and mark them to multiple track time nodes, and generate preliminary prediction results for wake interference.

9. The method for dynamic prediction of wake turbulence in close-range parallel runways based on deep learning according to claim 1, characterized in that, The method further includes: S5: Based on the preliminary prediction results of the wake interference, obtain the corresponding disturbance energy level in the disturbance boundary as the disturbance intensity, calculate the continuous change rate of the disturbance boundary and the stability of the track direction, and correct the disturbance intensity. Combine the track deviation to re-judge the disturbance risk level and predict the evolution trend of the wake disturbance, and output the wake dynamic prediction results of the close-range parallel runway. The dynamic prediction results of the wake of the close-range parallel runway include the trend of disturbance level change, the disturbance intensity correction sequence, and the risk level sequence within the time period.

10. The deep learning-based wake dynamic prediction method for near-distance parallel runways according to claim 9, characterized in that, The specific steps of S5 include: S501: Based on the preliminary prediction results of the wake interference, obtain the disturbance energy level of the corresponding region in the disturbance boundary, collect the disturbance intensity value sequence in a continuous time period, calculate the ratio of the disturbance intensity difference to the time interval in adjacent time periods, calculate the continuous change rate of the disturbance boundary, and establish a disturbance intensity change rate sequence. S502: Call the disturbance intensity change rate sequence and the track vector data of the corresponding time period of the aircraft, calculate the track direction difference and standardize it to obtain the directional stability in the whole sequence, and use it as an adjustment factor to correct the disturbance intensity and generate a stable corrected disturbance intensity sequence. S503: Based on the track deviation between the stable corrected disturbance intensity sequence and the preliminary prediction result of the wake interference, the deviation level boundary is re-judged according to the disturbance risk level, the disturbance risk level interval is divided, a disturbance level change trend map is constructed and superimposed with a time axis, and the dynamic prediction result of the wake of the close-range parallel runway is output.

Images