A route EDR estimation method based on a multi-head attention mechanism model

Abstract

The application discloses a route EDR estimation method based on a multi-head attention mechanism model, which comprises the following steps: obtaining vertical, transverse and longitudinal three-dimensional gust data; composing a turbulence data set; determining four hyperparameters of the multi-head attention mechanism model; obtaining fused three-dimensional gust; and identifying the position of turbulence in the route, and the like.The application has the advantages and effective benefits that: the application has strong comprehensive data capacity, has the advantages of low false negative rate and high identification rate, can effectively and accurately identify the turbulence of the route, and provides reliable guarantee for the safe flight of the airplane.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of mining of machine learning, and particularly relates to a route EDR estimation method based on a multi-head attention mechanism model. BACKGROUND

[0002] Atmospheric turbulence is a complex fluid motion superimposed by vortices of various scales, which can cause sudden jolts of an airplane and seriously affect the flight quality, ride comfort and flight safety of a civil aviation airplane. When the airplane is flying at high altitude and high speed, the airplane's force balance is broken due to the influence of rapidly changing turbulence, which can induce jolts of the airplane and even cause accidents in extreme cases. Therefore, in order to reduce the risk related to turbulence and ensure flight safety, it is crucial to accurately identify turbulence.

[0003] There are various methods for identifying turbulence, and so far there is no unified method, and these methods have certain limitations: turbulence can be identified by using vertical acceleration root mean square RMS-g and equivalent vertical gust speed (DEVG), but these two indicators are often related to parameters such as airplane weight and airspeed, and the measured acceleration is easily affected by airplane maneuvering flight, so it is difficult to accurately reflect the objective atmospheric turbulence intensity. On this basis, according to the von Karman and Kolmogorov energy spectrum theory, the energy of atmospheric turbulence is only related to the turbulence dissipation rate (EDR), which can reflect the objective turbulence intensity, so it is selected by the International Civil Aviation Organization (ICAO) as a detection index of atmospheric turbulence.

[0004] There are mainly two common EDR estimation methods, which are EDR estimation based on vertical acceleration and EDR estimation based on vertical gust. The EDR estimation based on vertical acceleration uses the input-output relationship of the acceleration response function to obtain the severity of turbulence from the vertical acceleration response, but the acceleration response function model is usually determined by fitting, which is difficult to guarantee the accuracy of the algorithm. The EDR estimation based on vertical gust avoids the need for acceleration response function, but only relies on the turbulence information in the vertical gust, which is difficult to estimate the correct EDR, so it is difficult to accurately identify turbulence. SUMMARY

[0005] In order to solve the above problems, the purpose of the present application is to provide a route EDR estimation method based on a multi-head attention mechanism model.

[0006] In order to achieve the above purpose, the route EDR estimation method based on the multi-head attention mechanism model provided by the present application comprises the following steps performed in sequence:

[0007] Step 1, collect QAR data, calculate vertical, lateral and longitudinal wind speeds with the airplane body axis as the coordinate axis, and then obtain vertical, lateral and longitudinal three-dimensional gust data through detrending operation;

[0008] Step 2, the vertical, lateral and longitudinal three-dimensional gust data and the vertical overload data in the QAR data are characterized, and the turbulence data samples are constructed with the turbulence level reported by the pilot as the label, and the turbulence data set is composed of multiple turbulence samples;

[0009] Step 3, according to the turbulence data set, combining the Bayesian hyperparameter optimization algorithm, four hyperparameters of the multi-head attention mechanism model are determined;

[0010] Step 4, according to the four hyperparameters obtained in step 3, a multi-head attention mechanism model is constructed, the three-dimensional gust and vertical overload data are brought into the model, and the vertical, lateral and longitudinal gusts are fused according to the correlation between the gust and the vertical overload to obtain the fused three-dimensional gust;

[0011] Step 5, determine the window size, bring the fused gust obtained in step 4 into the EDR calculation formula, obtain the EDR value of each window, and calculate the EDR value of the entire route through the sliding window method, according to the corresponding relationship between the EDR and the turbulence level published by the International Civil Aviation Organization, and combined with the latitude and longitude information in the QAR data, the position of the turbulence in the route is identified.

[0012] In step 1, the method for collecting QAR data and calculating vertical, lateral and longitudinal wind speed with the aircraft body axis as the coordinate axis is:

[0013] The QAR data is collected from the quick storage recorder, and the vertical, lateral and longitudinal wind speed with the aircraft body axis as the coordinate axis is calculated by using the parameters such as the heading angle, angle of attack and vertical speed corresponding to each point in the cruise stage; then the de-trending operation is performed, which is to respectively fit the functions of the vertical, lateral and longitudinal wind speed of the aircraft changing with time, and subtract the corresponding fitted functions to eliminate the influence of inertial wind, to obtain the vertical, lateral and longitudinal three-dimensional gust data.

[0014] In step 2, the method for constructing turbulence data samples with the above vertical, lateral and longitudinal three-dimensional gust data and vertical overload data in the QAR data as features, and with the turbulence level reported by the pilot as the label, and the turbulence data set is composed of multiple turbulence samples is:

[0015] The turbulence level is divided according to the EDR value specified by the International Civil Aviation Organization, and there are three levels, which are light, medium and heavy, and the corresponding labels are 1, 2 and 3 respectively;

[0016] The range of N km before and after the position of the turbulence occurrence reported by the pilot is defined as the turbulence range, the three-dimensional gust data in the vertical, lateral and longitudinal directions and the vertical overload data in the QAR data in the turbulence range are taken as the features, and the turbulence level reported by the pilot is taken as the label to construct a turbulence data sample, and a plurality of turbulence data samples form a turbulence data set.

[0017] In step 3, the method for determining four hyperparameters of the multi-head attention mechanism model according to the turbulence data set in combination with the Bayesian hyperparameter optimization algorithm is:

[0018] The four hyperparameters are: the number of heads h, the scaling factor dk, the output dropout rate dropout, and the dimension of the hidden layer hid_dim.

[0019] The processing method of the Bayesian hyperparameter optimization algorithm is as follows:

[0020] The proportion of each turbulence level in the turbulence data set is kept unchanged, the turbulence data set is randomly divided into a training set and a validation set according to a ratio of 4:1, this operation is repeated 5 times, and five training sets and five validation sets are obtained; the cross-entropy loss value is selected as the objective function, and the calculation formula of the cross-entropy loss value is:

[0021]

[0022] Wherein, C represents three turbulence levels, C=3; y i is the probability of the i-th class in the true label; is the probability of the i-th class output by the multi-head attention mechanism model;

[0023] A hyperparameter combination is composed of four hyperparameters; the number of randomly initialized points of the Bayesian hyperparameter optimization algorithm is set to an integer m, m different hyperparameter combinations are obtained, and these hyperparameter combinations are respectively brought into the multi-head attention mechanism model, m multi-head attention mechanism models are obtained, then the five training sets are used to train each multi-head attention mechanism model respectively, and the cross-entropy loss value of each multi-head attention mechanism model is tested by the five validation sets, and the average value of the cross-entropy loss value is calculated.

[0024] X is regarded as a hyperparameter combination, then a hyperparameter combination set X=[x1,x2,...,x m ] is composed of m hyperparameter combinations, the calculated m cross-entropy loss values form a cross-entropy loss value set L=[L1,L2,...,L m ], then a data set D={(x1,L1),(x2,L2),...,(x m ,L m);establish a Gaussian regression model according to the data set D;

[0025] Based on the established Gaussian regression model, select the improved probability PI as the sampling function, and pre-select a new hyperparameter combination x m+1 , and the corresponding cross-entropy loss value L m+1 to form a new array (x m+1 , L m+1 ), and the calculation formula of the improved probability PI is as follows:

[0026]

[0027] Where Φ is the cumulative distribution function of the standard normal distribution; μ(x), σ(x) are the expected value and variance of the Gaussian regression model respectively; L(x1) is the current best cross-entropy loss value.

[0028] When the cross-entropy loss value L m+1 is greater than the set threshold value, the array (x m+1 , L m+1 ) is added to the data set D, the Gaussian regression model is reconstructed, and a new hyperparameter combination is pre-selected until the cross-entropy loss value is less than the threshold value, and the corresponding hyperparameter combination, i.e. four hyperparameters, are output.

[0029] In step 4, the four hyperparameters obtained in step 3 are used to construct a multi-head attention mechanism model, the three-dimensional gust and vertical overload data are input into the model, and the vertical, horizontal and longitudinal gusts are fused according to the correlation between the gust and the vertical overload. The method for obtaining the fused three-dimensional gust is:

[0030] The gust matrix composed of three-dimensional gust data is denoted as Q or V, and the vector composed of vertical overload data is denoted as K. The matrix Q, V and vector K are projected to a lower dimension by head transformation.

[0031]

[0032] Wherein, corresponding to the projection matrix of the matrix Q, V and vector K, which needs to be learned by training; d head is the vector dimension in the projection space; h is the number of heads.

[0033] The calculation output of each head is:

[0034]

[0035] Wherein, softmax is a normalization function, and set is able to fuse the calculation output head i of each head in the feature dimension, then the expression of the fused gust is:

[0036] T g =(head) i ,...,head h W 0

[0037] Fusion Gust T g Its length is the same as that of a single gust of wind.

[0038] In step 5, the window size is determined, and the merged gust wind obtained in step 4 is substituted into the EDR calculation formula to obtain the EDR value for each window. The EDR value for the entire route is then calculated using the sliding window method. Based on the correspondence between EDR and turbulence level published by the International Civil Aviation Organization (ICAO), and combined with the latitude and longitude information in the QAR data, the method for identifying the location of turbulence in the route is as follows:

[0039] The EDR is calculated using the following formula:

[0040]

[0041] In the formula, γ is a correction factor affected by the aircraft type, and k h and k l These are the high-frequency and low-frequency coefficients for the selected frequency range, respectively. and These are the actual and theoretical energy spectral densities of turbulence, respectively.

[0042] Wherein, the actual energy spectral density is the fused gust in the frequency domain, and the formula is:

[0043]

[0044] In the formula, f s denoted as the sampling frequency of the gusts, and m as the length scale of the input wind vector. The windowed fusion gust is designed to reduce spectral leakage caused by the periodic truncation of the fusion gust due to the application of Fast Fourier Transform. The specific windowing formula is as follows:

[0045]

[0046] In the formula, For normalized graph basis-Hanning window functions;

[0047] The theoretical energy spectral density is the von Kármán autocorrelation function with a window function, as shown in the following formula:

[0048]

[0049] In the formula, Re is a constant related to the aircraft model; is the autocorrelation function of the von Karman model; L is the length scale of turbulence, usually 669m; T gk is the fusion gust T at k moment g ; G and G D are diffusion functions, which describe the energy transfer rate of turbulence; W k is the normalized partial autocorrelation function of the GIBBS-HANNING window function , and the specific formula is as follows:

[0050]

[0051] According to the calculation formula of the above EDR, the EDR value of each window is calculated, and a series of EDR values are obtained through the sliding window method, then according to the corresponding relationship between EDR and turbulence level published by the International Civil Aviation Organization shown in Table 1, combined with the latitude and longitude information in the QAR data, the position of turbulence in the air route is identified;

[0052] Table 1

[0053] Turbulence level Turbulence description EDR(m 2 / 3 s -1 )]]> 1 Light turbulence 0.1—0.4 2 Moderate turbulence 0.4—0.7 3 Heavy turbulence 0.7—0.8

[0054] The advantages and effective benefits of the present application are:

[0055] It has the advantages of strong comprehensive data capacity, low false alarm rate and high recognition rate, and can effectively and accurately identify turbulence in the air route, providing reliable protection for safe flight of the aircraft. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 is the principle framework diagram of the method of the present application.

[0057] Figure 2 is the multi-head attention mechanism model processing flowchart in the present application. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application.

[0059] As Figure 1 shown, the air route EDR estimation method based on the multi-head attention mechanism model provided by the present application includes the following steps performed in sequence:

[0060] Step 1, collect QAR data, calculate vertical, lateral and longitudinal wind speed with the aircraft body axis as the coordinate axis, and then obtain three-dimensional gust data in the vertical, lateral and longitudinal directions through the detrending operation;

[0061] The vertical, lateral and longitudinal wind speeds refer to the wind speeds in the cruise stage of the flight path, and the cruise stage is regarded as a line connected by infinite points. The vertical, lateral and longitudinal wind speeds can be calculated by using the parameters corresponding to the heading angle, angle of attack and vertical speed of each point.

[0062] The QAR data is collected from the quick storage recorder, and the vertical, lateral and longitudinal wind speeds with the aircraft body axis as the coordinate axis are calculated by using the parameters corresponding to the heading angle, angle of attack and vertical speed of each point in the cruise stage. Then, the trend removal operation is performed. The method is to respectively fit the functions of the vertical, lateral and longitudinal wind speeds of the aircraft varying with time, and subtract the corresponding fitted functions to eliminate the influence of inertial wind, so as to obtain the vertical, lateral and longitudinal three-dimensional gust data.

[0063] Step 2, the above vertical, lateral and longitudinal three-dimensional gust data and the vertical overload data in the QAR data are characterized, and the turbulence data samples are constructed with the turbulence level reported by the pilot as the label. The turbulence data set is composed of multiple turbulence data samples.

[0064] The turbulence level is divided according to the numerical value of the EDR specified by the International Civil Aviation Organization, and there are three levels, namely light, moderate and severe, and the corresponding labels are 1, 2 and 3 respectively.

[0065] The range of N km before and after the center of the turbulence occurrence position reported by the pilot is defined as the turbulence range. The turbulence data samples are constructed with the vertical, lateral and longitudinal three-dimensional gust data and the vertical overload data in the QAR data within the turbulence range as the characteristics, and the turbulence level reported by the pilot as the label. The turbulence data set is composed of multiple turbulence data samples.

[0066] Step 3, according to the turbulence data set, the four hyperparameters of the multi-head attention mechanism model are determined by combining the Bayesian hyperparameter optimization algorithm.

[0067] The four hyperparameters are: h: the number of heads, which affects the performance and accuracy of the multi-head attention mechanism model; dk: the scaling factor, which is used to balance the size of the attention score and reduce gradient dispersion; dropout: the output dropout rate, which is used to reduce the over-reliance between different heads to increase the robustness and generalization ability of the multi-head attention mechanism model; hid_dim: the dimension of the hidden layer, which determines the feature dimension of the attention output of each head, and affects the representation ability and feature extraction ability of the multi-head attention mechanism model.

[0068] The processing method of the Bayesian hyperparameter optimization algorithm is as follows:

[0069] The turbulent flow data set is randomly divided into a training set and a validation set in a ratio of 4:1 while keeping the proportion of each turbulent flow level in the turbulent flow data set unchanged. This operation is repeated 5 times, and five training sets and five validation sets are obtained in total. A cross-entropy loss value is selected as the objective function, and the calculation formula of the cross-entropy loss value is as follows:

[0070]

[0071] where C represents three turbulent flow levels, C = 3; y i is the probability of the i-th class in the true label; is the probability of the i-th class output by the multi-head attention mechanism model.

[0072] An hyperparameter combination is composed of four hyperparameters; the number of random initialization points of the Bayesian hyperparameter optimization algorithm is set to an integer m, and m different hyperparameter combinations are obtained. These hyperparameter combinations are respectively brought into the multi-head attention mechanism model, and m multi-head attention mechanism models are obtained. Then, the five training sets are used to train each multi-head attention mechanism model, and the cross-entropy loss values of each multi-head attention mechanism model are tested by using the five validation sets, and the average value of the cross-entropy loss values is calculated.

[0073] x is regarded as an hyperparameter combination, and the hyperparameter combination set X = [x1, x2,..., x m m is composed of m hyperparameter combinations. The corresponding m cross-entropy loss values are composed of the cross-entropy loss value set L = [L1, L2,..., L m m]. Then, the data set D = {(x1, L1), (x2, L2),..., (x m m, L m m)} is composed of the hyperparameter combination set X and the cross-entropy loss value set L. A Gaussian regression model is established according to the data set D.

[0074] Based on the established Gaussian regression model, the improved probability PI is selected as the sampling function, a new hyperparameter combination x m+1 is preselected, and the corresponding cross-entropy loss value L m+1 is composed of the new array (x m+1 , L m+1 ). The calculation formula of the improved probability PI is as follows:

[0075]

[0076] where Φ is the cumulative distribution function of the standard normal distribution; μ(x), σ(x) are the expected value and variance of the Gaussian regression model, respectively; and L(x1) is the current best cross-entropy loss value.

[0077] When the cross-entropy loss value Lm+1 When the cross-entropy loss value is greater than the set threshold value, the array (x m+1 , L m+1 ) is added to the data set D, the Gaussian regression model is reconstructed, and a new combination of hyperparameters is preselected until the cross-entropy loss value is less than the threshold value, and the corresponding combination of hyperparameters, i.e., four hyperparameters, is output.

[0078] Step 4: According to the four hyperparameters obtained in step 3, a multi-head attention mechanism model is constructed, the three-dimensional gust and vertical overload data are brought into the model, and the vertical, lateral and longitudinal gusts are fused according to the correlation between the gust and the vertical overload to obtain a fused three-dimensional gust.

[0079] The framework flowchart of the multi-head attention mechanism model is shown in Figure 2 .

[0080] The gust matrix composed of three-dimensional gust data is denoted as Q or V, and the vector composed of vertical overload data is denoted as K. The matrix Q, V and the vector K are projected to a lower dimension by head-by-head transformation.

[0081]

[0082] wherein, respectively correspond to the projection matrices of the matrix Q, V and the vector K, which need to be obtained by training and learning; d head is the vector dimension in the projection space; h is the number of heads;

[0083] The calculation output of each head is:

[0084]

[0085] wherein, softmax is a normalization function, and set each head i 's calculation output in the feature dimension can be fused, and the expression of the fused gust is:

[0086] T g = (head i ,..., head h ) W 0

[0087] The length of the fused gust T g is consistent with that of a single gust.

[0088] Step 5, determining the window size, bringing the fusion gust obtained in step 4 into the EDR calculation formula to obtain the EDR value of each window, and calculating the EDR value in the entire route through the sliding window method, identifying the position of the turbulence in the route according to the corresponding relationship between the EDR and the turbulence level published by the International Civil Aviation Organization, and combining the latitude and longitude information in the QAR data.

[0089] The sliding window method is to set a window size, which is set to 10s in the application, and then calculate an EDR value according to the QAR data in the window. The window time is continuously slid to finally calculate a series of EDR values in the route.

[0090] The calculation formula of the EDR is as follows:

[0091]

[0092] In the formula, γ is a correction factor affected by the aircraft model, k h and k l are high-frequency and low-frequency coefficients of the selected frequency interval, respectively, and are the actual energy spectrum density and the theoretical energy spectrum density of the turbulence, respectively;

[0093] The actual energy spectrum density is the fusion gust in the frequency domain, and the formula is as follows:

[0094]

[0095] In the formula, f s is the sampling frequency of the gust, m is the length scale of the input wind vector, is the fusion gust after windowing, and the windowing processing is to reduce the frequency spectrum leakage phenomenon caused by the periodic truncation of the fusion gust by using the fast Fourier transform. The specific windowing formula is as follows:

[0096]

[0097] In the formula, is the normalized graph-Hanning window function;

[0098] The theoretical energy spectrum density is the von Karman autocorrelation function with a window function, and the formula is as follows:

[0099]

[0100] In the formula, Re is a constant related to the aircraft model; is the autocorrelation function of the von Karman model; L is the turbulence length scale, usually 669m; T gk is the fusion gust T g at k; G and G Drespectively, are diffusion functions describing the energy transfer rate of the turbulence; W k is the normalized Rife-Hann window function is the biased autocorrelation function, and the specific formula is as follows:

[0101]

[0102] According to the calculation formula of the EDR, the EDR values of each window are calculated, and a series of EDR values are obtained through the sliding window method. Then, according to the corresponding relationship between the EDR and the turbulence level published by the International Civil Aviation Organization shown in Table 1, and in combination with the latitude and longitude information in the QAR data, the positions of the turbulence in the air route are identified.

[0103] Table 1

[0104] Turbulence level Turbulence description EDR(m 2 / 3 s -1 )]]> 1 Light turbulence 0.1—0.4 2 Moderate turbulence 0.4—0.7 3 Heavy turbulence 0.7—0.8

Claims

1. A route EDR estimation method based on a multi-head attention mechanism model, characterized in that: The route EDR estimation method based on the multi-head attention mechanism model comprises the following steps performed in sequence: Step 1, collect QAR data, calculate vertical, lateral and longitudinal wind speeds with the aircraft body axis as the coordinate axis, and then obtain vertical, lateral and longitudinal three-dimensional gust data through detrending operation; Step 2, take the above-mentioned vertical, lateral and longitudinal three-dimensional gust data and vertical overload data in the QAR data as features, and take the turbulence level reported by the pilot as a label to construct a turbulence data sample, and a plurality of turbulence data samples form a turbulence data set; Step 3, according to the turbulence data set, determine four hyperparameters of the multi-head attention mechanism model in combination with the Bayesian hyperparameter optimization algorithm; Step 4, according to the four hyperparameters obtained in step 3, construct a multi-head attention mechanism model, input the three-dimensional gust and vertical overload data into the model, and fuse the vertical, lateral and longitudinal gusts according to the correlation between the gust and the vertical overload to obtain the fused gust; Step 5, determine the window size, input the fused gust obtained in step 4 into the EDR calculation formula, obtain the EDR value of each window, and calculate the EDR value of the entire route through the sliding window method, according to the corresponding relationship between the EDR and the turbulence level published by the International Civil Aviation Organization, and in combination with the latitude and longitude information in the QAR data, identify the position of the turbulence in the route.

2. The route EDR estimation method based on the multi-head attention mechanism model according to claim 1, characterized in that: In step 1, the method for collecting QAR data, calculating vertical, lateral and longitudinal wind speeds with the aircraft body axis as the coordinate axis, and then obtaining vertical, lateral and longitudinal three-dimensional gust data through detrending operation is as follows: QAR data is collected from the quick storage recorder, and the vertical, lateral and longitudinal wind speeds with the aircraft body axis as the coordinate axis are calculated by using the parameters such as the heading angle, angle of attack and vertical speed corresponding to each point in the cruise stage; then, detrending operation is performed, that is, functions of the vertical, lateral and longitudinal wind speeds of the aircraft varying with time are fitted respectively, and the corresponding fitted functions are subtracted to eliminate the influence of inertial wind, thereby obtaining the vertical, lateral and longitudinal three-dimensional gust data. 3.The route EDR estimation method based on the multi-head attention mechanism model according to claim 1, characterized in that: In step 2, the method for taking the above-mentioned vertical, lateral and longitudinal three-dimensional gust data and vertical overload data in the QAR data as features, taking the turbulence level reported by the pilot as a label to construct a turbulence data sample, and forming a turbulence data set from a plurality of turbulence data samples is as follows: The turbulence level is divided according to the EDR value specified by the International Civil Aviation Organization, and there are three levels, namely, light, medium and heavy, and the corresponding labels are 1, 2 and 3 respectively; A range of N km expanded before and after the position of the turbulence reported by the pilot is defined as the turbulence range, and the vertical, lateral and longitudinal three-dimensional gust data and the vertical overload data in the QAR data in the turbulence range are taken as features, and the turbulence level reported by the pilot is taken as a label to construct a turbulence data sample, and a plurality of turbulence data samples form a turbulence data set. 4.The route EDR estimation method based on the multi-head attention mechanism model according to claim 1, characterized in that: In step 3, the method for determining four hyperparameters of the multi-head attention mechanism model according to the turbulence data set in combination with the Bayesian hyperparameter optimization algorithm is as follows: The four hyperparameters are respectively: the number of heads h, the scaling factor dk, the output dropout rate dropout and the dimension of the hidden layer hid_dim; The processing method of the Bayesian hyperparameter optimization algorithm is as follows: The proportions of each turbulence level in the turbulence data set are kept unchanged, the turbulence data set is randomly divided into a training set and a validation set at a ratio of 4:1, the operation is repeated 5 times, and five training sets and five validation sets are obtained; the cross-entropy loss value is selected as the objective function, and the calculation formula of the cross-entropy loss value is: wherein C represents a turbulent flow level number, C = 3; y i is the probability of the i-th class in the real label; is the probability of the i-th class output by the multi-head attention mechanism model; A hyperparameter combination is composed of four hyperparameters; the number of random initialization points of the Bayesian hyperparameter optimization algorithm is set to an integer m, m different hyperparameter combinations are obtained, and the hyperparameter combinations are respectively substituted into the multi-head attention mechanism model, m multi-head attention mechanism models are obtained, then each multi-head attention mechanism model is trained using the five training sets, and the cross-entropy loss value of each multi-head attention mechanism model is tested using the five validation sets, and the average value of the cross-entropy loss value is calculated; X is a hyperparameter combination, and a hyperparameter combination set X is composed of m hyperparameter combinations [x1,x2,...,x m The calculated m cross-entropy loss values ​​are then combined into a cross-entropy loss value set L = [L1, L2, ..., L]. m Then, the dataset D = {(x1,L1),(x2,L2),...,(x...} is constructed from the hyperparameter combination set X and the cross-entropy loss value set L. m ,L m )}; Build a Gaussian regression model based on dataset D; Based on the established Gaussian regression model, the improved probability PI is selected as the sampling function, and a new hyperparameter combination x m+1 is preselected m+1 , and the corresponding cross-entropy loss value L m+1 is composed into a new array (x m+1 , L m+1 ), and the calculation formula of the improved probability PI is as follows: Wherein, Φ is the cumulative distribution function of the standard normal distribution; μ(x), σ(x) are the expected value and variance of the Gaussian regression model respectively; L(x1) is the current best cross-entropy loss value; When the cross-entropy loss value L m+1 When the value exceeds the set threshold, the array (x) will be... m+1 ,L m+1 Add dataset D, rebuild the Gaussian regression model, and pre-select new hyperparameter combinations until the cross-entropy loss value is less than the threshold. Output the corresponding hyperparameter combination, i.e., four hyperparameters.

5. The route EDR estimation method based on the multi-head attention mechanism model according to claim 1, characterized in that: In step 4, the four hyperparameters obtained in step 3 are used to construct a multi-head attention mechanism model, three-dimensional gust and vertical overload data are substituted into the model, and the method of fusing vertical, horizontal and longitudinal gusts according to the correlation between gust and vertical overload is: The gust matrix composed of three-dimensional gust data is denoted as Q or V, and the vector composed of vertical overload data is denoted as K, the matrix Q, V and the vector K are projected to a lower dimension through head transformation; wherein, The projection matrix corresponding to the matrix Q, V and the vector K respectively, needs to be obtained through training and learning; d head is the vector dimension in the projection space; h is the number of sub-heads. The calculation output of each head is: Wherein, softmax is a normalization function, and let Each split head can output head i The expression of the fused gust is: T g = (head i ,...,head h )W 0 Fusion Storm T g of the same length as a single storm.

6. The route EDR estimation method based on the multi-head attention mechanism model according to claim 5, characterized in that: In step 5, the window size is determined, the fused gust obtained in step 4 is substituted into the EDR calculation formula, the EDR value of each window is obtained, and the EDR value of the entire route is calculated through the sliding window method, and the method of identifying the position of turbulence in the route according to the corresponding relationship between EDR and turbulence level published by the International Civil Aviation Organization and in combination with the latitude and longitude information in the QAR data is: The calculation formula of the EDR is as follows: where γ is a correction factor influenced by the model of the receiver, k h and k l are the high and low frequency coefficients of the selected frequency interval, respectively, and are the actual and theoretical energy spectral density of the turbulence, respectively. Wherein, the actual energy spectrum density is the fused gust in the frequency domain, and the formula is: where f is the sampling frequency of the gust, s is the sampling frequency of the gust, m is the length scale of the input wind vector, is the windowed fused gust. The windowing process is to reduce the spectrum leakage phenomenon caused by the periodic truncation of the fused gust using the fast Fourier transform. The specific windowing formula is: wherein is a normalized Rife-Vincent window function; The theoretical energy spectrum density is the von Karman autocorrelation function with a window function, and the formula is as follows: where Re is a constant related to the aircraft model; is the autocorrelation function of the von Karman model; L is the turbulence length scale, taken as 669 m; T gk is the fusion gust T g at time k; G and G D are the diffusion functions, respectively, describing the energy transfer rate of the turbulence; W k is the normalized partial autocorrelation function of the Gabor-Hanning window function , and the specific formula is as follows: According to the above EDR calculation formula, the EDR value of each window is calculated, and a series of EDR values are obtained through the sliding window method, then the position of turbulence in the route is identified according to the corresponding relationship between EDR and turbulence level published by the International Civil Aviation Organization and in combination with the latitude and longitude information in the QAR data.

Images