Full-route bump feature mining and EDR estimation method based on airborne QAR data

Abstract

The invention relates to a full-route bump feature mining and EDR estimation method based on airborne QAR data. The method comprises the following steps: forming an EDR basic data set; mining and extraction of advanced features are realized; forming an optimized feature set; obtaining a trained neural network model; and calculating the full-route EDR value of the real route. Through combination of discrete wavelet transform and a neural network model, the invention provides an accurate and reliable EDR estimation method, and compared with a traditional single-parameter method, detail description is more obvious and is completely and automatically calculated by a computer without manual reasoning. According to the method, after model training and optimization, the EDR value on a single course can be estimated in real time, and the accuracy and practicability of analyzing an airplane bumping event through flight data in the aviation field are improved.

Description

Technical Field

[0001] The invention belongs to the technical field of aviation meteorology, and in particular relates to a method for mining full-route turbulence characteristics and estimating EDR based on airborne QAR data. Background Art

[0002] As global temperatures continue to rise and the civil aviation industry develops rapidly, flights are increasingly encountering turbulence. Turbulence not only affects flight comfort, but can also pose a threat to flight safety. In extreme cases, turbulence can even cause damage to the aircraft and injuries to passengers and crew members, so it is particularly important to effectively identify and respond to turbulence.

[0003] There are many methods for turbulence identification, such as the root mean square of vertical overload (RMS-g) and equivalent vertical gust speed (DEVG). However, these methods are easily affected by factors such as aircraft weight and airspeed, resulting in redundant mixing of turbulence caused by the environment and aircraft factors, making it difficult to accurately reflect the true atmospheric turbulence intensity. In view of this, the International Civil Aviation Organization (ICAO) has determined that the eddy dissipation rate (EDR) is an indicator that objectively reflects the turbulence intensity.

[0004] The current EDR estimation methods mainly include the following: EDR estimation method based on vertical acceleration, EDR estimation method based on vertical gusts, and EDR calculation algorithm that is calculated and transmitted in real time through airborne data. Compared with the airborne algorithm, the EDR values ​​estimated by the first two methods are usually lower, with insufficient ability to depict details and unclear numerical changes, so it is difficult to accurately reflect the real changes in turbulence. Summary of the invention

[0005] In order to solve the above problems, the object of the present invention is to provide a method for full-route turbulence feature mining and EDR estimation based on airborne QAR data.

[0006] In order to achieve the above object, the method for full-route turbulence feature mining and EDR estimation based on airborne QAR data provided by the present invention comprises the following steps performed in sequence:

[0007] 1) Collect a large number of airborne QAR data from flights, and then extract multiple flight parameters related to turbulence. Then match the existing EDR (airborne eddy dissipation rate) samples with the above flight parameters. By dividing the time window, the flight parameters within a period of time are associated with the corresponding EDR values. All EDR values ​​constitute the EDR basic data set of the complete route airborne QAR data time window.

[0008] 2) The flight parameters in the time window of the above EDR basic data set are regarded as discrete time series signals, and then the discrete time series signals are transformed and decomposed by discrete wavelet transform method to extract the time-frequency localization information or frequency domain features of the discrete time series signals, and then the time-frequency localization information or frequency domain features are used to obtain high-level features, thereby realizing the mining and extraction of high-level features;

[0009] 3) All the above high-level features are input into the random forest model, and the model is used to evaluate and rank the high-level features, and several high-ranking high-level features are selected to form the optimized feature set;

[0010] 4) constructing a neural network model with hyperparameter optimization suitable for EDR fitting, and training the neural network model using the advanced features in the above-mentioned optimization feature set and the EDR values ​​associated in step 1), thereby determining the weights and bias parameters of the neural network model, and obtaining a trained neural network model;

[0011] 5) Collect the airborne QAR data of the actual route, and then process the airborne QAR data according to the methods of step 1) and step 2) to obtain high-level features, and then input the high-level features into the above-mentioned trained neural network model, and use the weight and bias parameters of the model to calculate the full route EDR value of the route.

[0012] In step 1), the airborne QAR data of various flights are collected, and then a plurality of flight parameters related to turbulence are extracted therefrom, and then the existing EDR samples are matched with the above flight parameters; by dividing the time window, the flight parameters within a period of time are associated with the corresponding EDR values, and the method of forming the EDR basic data set of the time window of the airborne QAR data of the complete route by all EDR values ​​is: collecting the airborne QAR data of various flights, and then extracting a plurality of flight parameters related to turbulence including vertical overload, radio altitude, indicated airspeed, angle of attack, flaps, pitch rate, roll angle, vertical speed, gross weight, true airspeed, Mach number, geographical location, wind speed, wind direction and static air temperature therefrom; then matching the existing EDR samples with the above flight parameters based on the flight number and flight time of the aircraft, and dividing the time window of 15 seconds, marking the flight parameter changes within the time window with the corresponding EDR values, thereby establishing an association between the flight parameters and the corresponding EDR values, and forming the EDR basic data set of the time window of the airborne QAR data of the complete route by all EDR values.

[0013] In step 2), the flight parameters in the time window of the EDR basic data set are regarded as discrete time series signals, and then the discrete time series signals are converted and decomposed by discrete wavelet transform method to extract the time-frequency localization information or frequency domain features of the discrete time series signals, and then the time-frequency localization information or frequency domain features are used to obtain high-level features, so as to realize the method of mining and extracting high-level features:

[0014] The flight parameters within the time window are regarded as discrete time series signals, and the discrete wavelet transform method is used to transform and decompose the discrete time series signals to extract the time-frequency localization information and frequency domain characteristics of the signals.

[0015] For the one-dimensional discrete time series signal x[n], the Daubechies1 wavelet has an associated low-pass filter g[n] and high-pass filter g[n], which are generated by the wavelet function ψ(t) and its scaling function φ(t). These filters are usually called "analysis filters": the approximation coefficient cA is obtained by convolving the one-dimensional discrete time series signal x[m] with the low-pass filter h[n], which represents the approximation of the discrete time series signal in the low-frequency part. The calculation formula is as follows:

[0016] cA j [n]=(x*h)[n]=∑ k x[k]·h[nk] (1)

[0017] The detail coefficient cD is obtained by convolving the one-dimensional discrete time series signal x[n] with the high-pass filter g[n], which represents the details of the high-frequency part of the discrete time series signal. The calculation formula is as follows:

[0018]

[0019] In discrete wavelet transform, the discrete time series signal x[n] is decomposed into different levels of approximation coefficients cA and detail coefficients cD belonging to time-frequency localization information or frequency domain features. Each level corresponds to information in a specific frequency range. Then the energy, standard deviation, maximum and mean of the approximation coefficients cA and detail coefficients cD are calculated and used as high-level features.

[0020] In step 3), the formula for evaluating the importance of the high-level features is as follows:

[0021]

[0022] Among them, N T is the number of trees in the random forest, nodes(t) is the set of all nodes in the t-th tree, p(i,t) is the probability that node i is in tree t, Δ(i,t,X j ) is the node i in feature X jImprovements on the splitting criteria include changes in information gain or Gini coefficient.

[0023] In step 4), a neural network model with hyperparameter optimization suitable for EDR fitting is constructed, and the neural network model is trained using the advanced features in the above-mentioned optimized feature set and the EDR value associated in step 1), thereby determining the weight and bias parameters of the neural network model. The method for obtaining the trained neural network model is:

[0024] Constructing a neural network model with hyperparameter optimization suitable for EDR fitting, including an input layer, a plurality of hidden layers and an output layer; inputting the high-level features in the above-mentioned optimized feature set and the EDR values ​​associated in step 1) into the neural network model, setting the optimized hyperparameters, and training the neural network model through a back propagation algorithm and an optimizer to reduce the loss function, so that the model can accurately fit the EDR value from the airborne QAR data, thereby determining the weight and bias parameters of the neural network model, and obtaining a trained neural network model;

[0025] The optimizer adopts the Radam optimizer, and a correction term is added when calculating the adaptive learning rate;

[0026] The loss function adopts the Huber loss function, which has less impact on outliers (i.e., predicted values ​​far from the true value) and is more robust than the mean square error (MSE). The expression is as follows:

[0027]

[0028] Among them, y is the true value, is the predicted value, δ is a hyperparameter used to control the turning point of square loss and absolute loss;

[0029] Hyperparameter optimization is an important step before training a neural network model. It aims to find the best combination of model hyperparameters to improve the performance and generalization ability of the model and prevent overfitting. The optimized hyperparameters include learning rate, batch size, number of nodes in the hidden layer, L1 and L2 regularization coefficients, number of layers, and output dropout rate.

[0030] In step 5), the airborne QAR data of the real route is collected, and then the airborne QAR data is processed according to the method of step 1) and step 2) to obtain high-level features, and then the high-level features are input into the above-mentioned trained neural network model. The method for calculating the full route EDR value of the route using the weight and bias parameters of the model is:

[0031] First, the airborne QAR data of the real route is obtained to ensure the quality and integrity of the data. Then, the airborne QAR data is divided into time windows and features are extracted according to the methods of steps 1) and 2), and finally high-level features including energy, standard deviation, maximum and mean of the corresponding approximation coefficient cA and detail coefficient cD are obtained; these high-level features are then input into the trained neural network model obtained in step 4), and through the forward propagation process of the model, the weights and bias parameters learned by the model are used to calculate the EDR value of each time window on the route.

[0032] The full-route turbulence feature mining and EDR estimation method based on airborne QAR data provided by the present invention has the following beneficial effects:

[0033] 1. This method fits the eddy dissipation rate (EDR) transmitted by the onboard algorithm of a large number of route models. Compared with the traditional EDR calculation method, it can more accurately estimate the EDR value of a single route. Its estimation results have a good correlation with parameters reflecting turbulence such as vertical overload, and can describe the changing characteristics of atmospheric turbulence in detail.

[0034] 2. The neural network model showed excellent fitting ability. It performed well in detection independent of the training data, and after advanced feature extraction, the validation data did not overfit, thus ensuring the generalization ability of the model.

[0035] 3. After high-level feature extraction and importance ranking, the high-frequency components of the airborne QAR parameters related to turbulence were evaluated as significantly important. When these parameters were incorporated into the neural network model for EDR fitting, the model showed strong interpretability, thus overcoming the "black box" problem of the machine learning model to a certain extent.

[0036] 4. By fitting the airborne QAR data with the existing airborne calculated EDR data, the relationship between the airborne QAR data and the real-time calculated EDR is learned by using the iterative calculation of the neural network, so that accurate EDR estimation results can be obtained without manually setting the calculation method. This method shows broad application prospects.

[0037] 5. By fitting the airborne QAR data of a large number of flight time windows, the EDR of a single route can be estimated, showing strong applicability and broad application potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A flow chart of the method for mining full-route turbulence characteristics and estimating EDR based on airborne QAR data provided by the present invention;

[0039] Figure 2It is a schematic diagram of the matching of airborne QAR parameters and eddy energy dissipation rate (EDR) and time window division in the present invention;

[0040] Figure 3 It is a schematic diagram of the process of extracting features from airborne QAR data within a time window using a discrete wavelet transform method in the present invention;

[0041] Figure 4 Flow chart of the EDR value fitting process calculated using the hyperparameter optimized neural network model in the present invention. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solutions and advantages of the present invention clearer, the implementation scheme of the present invention is further described in detail below in conjunction with the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0043] like Figure 1 As shown, the method for full-route turbulence feature mining and EDR estimation based on airborne QAR data provided by the present invention includes the following steps performed in sequence:

[0044] 1) Obtain airborne QAR data of a large number of flights and extract 18 flight parameters related to turbulence, as shown in Table 1; then match these flight parameters with existing airborne eddy dissipation rate (EDR) samples according to the aircraft flight number and flight time.

[0045] Table 1

[0046]

[0047]

[0048] Divide the time window into 15 seconds, and the flight parameters of each flight form a parameter group, and the parameter group within 15 seconds forms a time window. The flight parameter changes in each time window are marked as corresponding EDR values. The specific operation is as follows Figure 2 As shown, the flight parameters are associated with the corresponding EDR values, and all EDR values ​​constitute the EDR basic data set of the complete route airborne QAR data time window.

[0049] 2) The flight parameters in each flight time window in the EDR basic data set are regarded as discrete time series signals, and the discrete time series signals are processed using the discrete wavelet transform method to mine the features. The specific operations are as follows: Figure 3As shown in the figure. The approximation coefficient cA obtained after discrete wavelet transform mainly reflects the overall trend of low-frequency signals and describes the overall trend of airborne QAR parameter changes; the detail coefficient cD mainly captures the details of high-frequency signals and reflects the sudden changes in airborne QAR parameter changes. The analysis of these two coefficients can not only reveal the trend of airborne QAR data, but also capture the sudden changes in airborne QAR data when encountering environmental changes.

[0050] For the one-dimensional discrete time series signal x[n], the Daubechies1 wavelet has an associated low-pass filter h[n] and high-pass filter g[n], which are generated by the wavelet function ψ(t) and its scaling function φ(t). These filters are usually called "analysis filters". The approximation coefficient cA is obtained by convolving the one-dimensional discrete time series signal x[n] with the low-pass filter h[n], which represents the approximation of the discrete time series signal in the low-frequency part. The calculation formula is as follows:

[0051] cA j [n]=(x*h)[n]=∑ k x[k]·h[nk] (1)

[0052] The detail coefficient cD is obtained by convolving the one-dimensional discrete time series signal x[n] with the high-pass filter g[n], which represents the details of the high-frequency part of the discrete time series signal. The calculation formula is as follows:

[0053] cD j [n]=(x*g)[n]=∑ k x[k]·g[nk] (2)

[0054] The mean, energy, maximum value and standard deviation of the approximation coefficient cA and detail coefficient cD are calculated as advanced features of the eddy dissipation rate (EDR) fitting;

[0055] 3) Use the random forest algorithm to screen the high-level features obtained after discrete wavelet transform. The high-level features obtained by the approximation coefficient cA and the detail coefficient cD are used as input features, and the random forest model is used to evaluate their importance and sort them, and several high-level features with higher rankings are selected to form the optimized feature set;

[0056] In this step, the random forest model can effectively determine which features are most important for the eddy dissipation rate (EDR) fitting, thereby optimizing the feature set. The formula for evaluating the importance of high-level features is as follows:

[0057]

[0058] Among them, N Tis the number of trees in the random forest, nodes(t) is the set of all nodes in the t-th tree, p(i,t) is the probability that node i is in tree t, Δ(i,t,X j ) is the node i in feature X j Improvements on the splitting criteria include changes in information gain or Gini coefficient.

[0059] 4) Using the optimized feature set obtained in step 3), the hyperparameter optimized neural network model is trained for the eddy current dissipation rate (EDR). The training process of the hyperparameter optimized neural network model is as follows: Figure 4 As shown. First, the associated EDR value obtained in step 1) is input into the multi-layer neural network model. The hyperparameter optimization library combines a series of parameters, including the number of layers of the neural network model, the ratio of the output discard rate, and the L1 and L2 regularization coefficients, to configure the multi-layer neural network model. Subsequently, the EDR value is fitted according to these settings. The output layer of the neural network model will return the fitted EDR value, and the fitted EDR value of the sample will be calculated through the Huber loss function and the true EDR value of the sample. The specific calculation formula is as follows:

[0060]

[0061] Among them, y is the true value, is the predicted value, δ is a hyperparameter used to control the turning point of square loss and absolute loss;

[0062] This loss function has little impact on outliers (i.e., predicted values ​​far from the true value), allowing the neural network model to focus on reducing the loss value to achieve the best fitting effect. During the training process, the neural network model will iterate multiple times, adjusting the weights and bias parameters to find the parameter configuration that can minimize the loss function. The hyperparameter optimization library records the best loss value under each set of parameter configurations to guide the optimization process of the next set of parameters.

[0063] In each iteration, the hyperparameter optimization library selects a new parameter combination based on previous experience and fits the EDR value again. The RAdam optimizer updates the weights and bias parameters of the neural network model based on the current loss value to further optimize the fitting process. Through multiple optimization and loss minimization processes, the optimal parameter combination is finally found, allowing the neural network model to better fit the eddy dissipation rate (EDR) in the airborne QAR data.

[0064] 5) After hyperparameter optimization and iterative learning of the model, the weights and bias parameters obtained under the best loss are recorded to save the optimized neural network model. According to the screening rules in step 4), the parameters of the airborne QAR data of a single route are selected, and the same airborne QAR data used in training is retained. Then, the discrete wavelet transform method described in step 3) is used to extract features from these selected parameters to generate high-level features.

[0065] Next, these high-level features are fed into a multi-layer neural network model with saved optimal weights and bias parameters to obtain the estimation results of the eddy dissipation rate (EDR).

[0066] The above description is a specific implementation of the present invention and is not intended to limit the present invention. Any modifications and substitutions made by a person skilled in the art within the technical scope disclosed in the present invention should be included in the protection scope of the present invention.

Claims

1. A method for full-route turbulence feature mining and EDR estimation based on airborne QAR data, characterized by: The method for mining full-route turbulence features and estimating EDR based on airborne QAR data comprises the following steps performed in sequence: 1) Collect a large number of airborne QAR data of flights, and then extract multiple flight parameters related to turbulence from them, and then match the existing EDR samples with the above flight parameters; by dividing the time window, associate the flight parameters within a period of time with the corresponding EDR values, and all EDR values ​​constitute the EDR basic data set of the complete route airborne QAR data time window; 2) The flight parameters in the time window of the above EDR basic data set are regarded as discrete time series signals, and then the discrete time series signals are transformed and decomposed by discrete wavelet transform method to extract the time-frequency localization information or frequency domain features of the discrete time series signals, and then the time-frequency localization information or frequency domain features are used to obtain high-level features, thereby realizing the mining and extraction of high-level features; 3) All the above high-level features are input into the random forest model, and the model is used to evaluate and rank the high-level features, and several high-ranking high-level features are selected to form the optimized feature set; 4) constructing a neural network model with hyperparameter optimization suitable for EDR fitting, and training the neural network model using the advanced features in the above-mentioned optimization feature set and the EDR values ​​associated in step 1), thereby determining the weights and bias parameters of the neural network model, and obtaining a trained neural network model; 5) Collect the airborne QAR data of the actual route, and then process the airborne QAR data according to the methods of step 1) and step 2) to obtain high-level features, and then input the high-level features into the above-mentioned trained neural network model, and use the weight and bias parameters of the model to calculate the full route EDR value of the route.

2. The method for full-route turbulence feature mining and EDR estimation based on airborne QAR data according to claim 1 is characterized by: In step 1), the airborne QAR data of a large number of flights are collected, and then a plurality of flight parameters related to turbulence are extracted therefrom, and then the existing EDR samples are matched with the above flight parameters; by dividing the time window, the flight parameters within a period of time are associated with the corresponding EDR values, and the method of forming the EDR basic data set of the time window of the airborne QAR data of the complete route by all EDR values ​​is: collecting a large number of airborne QAR data of various flights, and then extracting a plurality of flight parameters related to turbulence including vertical overload, radio altitude, indicated airspeed, angle of attack, flaps, pitch rate, roll angle, vertical speed, gross weight, true airspeed, Mach number, geographical location, wind speed, wind direction and static air temperature therefrom; then, based on the flight number and flight time of the aircraft, the existing EDR samples are matched with the above flight parameters, and the time window of 15 seconds is divided, and the flight parameter changes within the time window are marked with the corresponding EDR values, thereby establishing an association between the flight parameters and the corresponding EDR values, and forming the EDR basic data set of the time window of the airborne QAR data of the complete route by all EDR values.

3. The method for full-route turbulence feature mining and EDR estimation based on airborne QAR data according to claim 1 is characterized by: In step 2), the flight parameters in the time window of the EDR basic data set are regarded as discrete time series signals, and then the discrete time series signals are converted and decomposed by discrete wavelet transform method to extract the time-frequency localization information or frequency domain features of the discrete time series signals, and then the time-frequency localization information or frequency domain features are used to obtain high-level features, so as to realize the method of mining and extracting high-level features: The flight parameters in the time window are regarded as discrete time series signals, and the discrete time series signals are transformed and decomposed using discrete wavelet transform method to extract the time-frequency localization information and frequency domain characteristics of the signals. For the one-dimensional discrete time series signal x[n], the Daubechies1 wavelet has an associated low-pass filter h[n] and high-pass filter g[n], which are generated by the wavelet function ψ(t) and its scaling function φ(t). These filters are usually called "analysis filters": the approximation coefficient cA is obtained by convolving the one-dimensional discrete time series signal x[n] with the low-pass filter h[n], which represents the approximation of the discrete time series signal in the low-frequency part. The calculation formula is as follows: cA j [n]=(x*h)[n]=Σ k x[k]·h[nk] (1) The detail coefficient cD is obtained by convolving the one-dimensional discrete time series signal x[n] with the high-pass filter g[n], which represents the details of the high-frequency part of the discrete time series signal. The calculation formula is as follows: cD j [n]=(x*g)[n]=Σ k x[k] g[nk] (2) In discrete wavelet transform, the discrete time series signal x[n] is decomposed into different levels of approximation coefficients cA and detail coefficients cD belonging to time-frequency localization information or frequency domain features. Each level corresponds to information in a specific frequency range. Then the energy, standard deviation, maximum and mean of the approximation coefficients cA and detail coefficients cD are calculated and used as high-level features.

4. The method for full-route turbulence feature mining and EDR estimation based on airborne QAR data according to claim 1 is characterized by: In step 3), the formula for evaluating the importance of the high-level features is as follows: Among them, N T is the number of trees in the random forest, nodes(t) is the set of all nodes in the t-th tree, p(i,t) is the probability that node i is in tree t, Δ(i,t,X j ) is the node i in feature X j Improvements on the splitting criteria include changes in information gain or Gini coefficient.

5. The method for full-route turbulence feature mining and EDR estimation based on airborne QAR data according to claim 1 is characterized by: In step 4), a neural network model with hyperparameter optimization suitable for EDR fitting is constructed, and the neural network model is trained using the advanced features in the above-mentioned optimized feature set and the EDR value associated in step 1), thereby determining the weight and bias parameters of the neural network model. The method for obtaining the trained neural network model is: Constructing a neural network model with hyperparameter optimization suitable for EDR fitting, including an input layer, a plurality of hidden layers and an output layer; inputting the high-level features in the above-mentioned optimized feature set and the EDR value associated in step 1) into the neural network model, setting the optimized hyperparameters, training the neural network model through a back propagation algorithm and an optimizer, thereby determining the weight and bias parameters of the neural network model, and obtaining a trained neural network model; The optimizer adopts the Radam optimizer, and a correction term is added when calculating the adaptive learning rate; The loss function adopts the Huber loss function, which is expressed as follows: Among them, y is the true value, is the predicted value, δ is a hyperparameter used to control the turning point of square loss and absolute loss; The optimized hyperparameters include learning rate, batch size, number of nodes in hidden layers, L1 and L2 regularization coefficients, number of layers, and output dropout rate.

6. The method for full-route turbulence feature mining and EDR estimation based on airborne QAR data according to claim 1 is characterized by: In step 5), the airborne QAR data of the real route is collected, and then the airborne QAR data is processed according to the method of step 1) and step 2) to obtain high-level features, and then the high-level features are input into the above-mentioned trained neural network model. The method for calculating the full route EDR value of the route using the weight and bias parameters of the model is: First, the airborne QAR data of the real route is obtained to ensure the quality and integrity of the data. Then, the airborne QAR data is divided into time windows and features are extracted according to the methods of steps 1) and 2), and finally high-level features including energy, standard deviation, maximum and mean of the corresponding approximation coefficient cA and detail coefficient cD are obtained; these high-level features are then input into the trained neural network model obtained in step 4), and through the forward propagation process of the model, the weights and bias parameters learned by the model are used to calculate the EDR value of each time window on the route.

Images