A Deep Learning Method for EDR Estimation Based on Aircraft Response Features

Abstract

This invention introduces a deep learning method for EDR estimation based on aircraft response characteristics. By combining staged calculation and neural network learning, this method provides an accurate and reliable EDR estimation approach. Unlike traditional EDR estimation methods, the method described in this invention features clearer stage division and dynamic adjustment capabilities. The method includes the following steps: systematically integrating flight data collected from Quick Access Recorders (QARs) of all flights nationwide, constructing time windows, and performing intelligent hierarchical sampling; reading flight data from CSV files and dividing flight stages using a state machine; calculating takeoff and landing stages and cruise stages separately, using raw QAR data to calculate high-level features; finally, training the calculated high-level features using a ConvLSTM-ELM neural learning network model to obtain the optimal parameters of the model. After neural network training and learning, this method can accurately estimate the EDR value of the entire flight route, improving the accuracy and real-time performance of EDR calculation using recorded QAR data in the aviation field.

Description

[0001] This invention belongs to the field of aviation meteorology technology, specifically involving a deep learning method for EDR estimation based on aircraft response characteristics. Background Technology

[0002] Atmospheric turbulence remains a major challenge in aviation, as its unpredictability threatens flight safety and passenger comfort. Eddy loss rate (EDR) is recognized by the International Civil Aviation Organization (ICAO) as the standard indicator for measuring turbulence, providing a quantitative basis for assessing turbulence intensity. However, accurately estimating EDR in real-world flight scenarios is highly complex, as it reflects not only atmospheric conditions but also the aircraft's dynamic response to turbulent disturbances.

[0003] Current EDR estimation methods, broadly categorized into those based on vertical acceleration and those based on vertical wind speed, face significant limitations. These methods typically rely on simplifying assumptions, such as isotropic turbulence or the symmetry of the aircraft response, failing to adequately account for the complex coupling between aerodynamics, flight operations, and environmental disturbances. Furthermore, their dependence on high-frequency sensors limits scalability, while the widely used Fast Access Recorder (QAR) has a low data sampling rate, making it difficult to capture transient turbulence signals. While existing data-driven models show promise, they rarely integrate the physical mechanisms of aircraft response, such as inertial delay or control system feedback, resulting in suboptimal accuracy under varying flight conditions. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to provide a deep learning method for EDR estimation based on aircraft response features.

[0005] To achieve the above objectives, the deep learning method for EDR estimation based on aircraft response features provided by this invention includes the following steps performed in sequence:

[0006] 1) Systematically integrate flight data collected by Quick Access Recorders (QARs) from all flights nationwide, and perform spatiotemporal correlation analysis with existing aircraft eddy current dissipation (EDR) monitoring samples. By constructing time windows and performing intelligent hierarchical sampling, the dynamic characteristics of turbulence can be captured more effectively while ensuring the generalization ability of the trained model;

[0007] 2) Develop an adaptive flight phase segmentation algorithm that processes QAR data in CSV format and uses IVV and altitude data to divide flight phases, thereby estimating EDR data for each flight phase, effectively isolating the turbulent response of each flight phase, thus improving the accuracy of EDR estimation, clearly displaying each flight phase through visualization and performing specific analysis on each flight phase;

[0008] 3) During aircraft takeoff and landing, the QAR parameters reflecting turbulence caused by aerodynamic characteristics, maneuvering, and environmental disturbances contain a significant non-turbulent component. To accurately extract turbulent features, a decoupling method based on vertical wind speed is proposed to extract auxiliary features;

[0009] 4) During the cruise phase, the aerodynamic state is stable, and the changes in sensor parameters are mainly affected by turbulent disturbances. Therefore, a set of combined parameters based on QAR data was developed specifically for the cruise phase to obtain more accurate data.

[0010] 5) ConvLSTM-ELM neural network model is constructed and a high EDR value training mechanism is introduced to focus on more dangerous strong turbulence events;

[0011] 6) Set different evaluation criteria for different types of turbulence to overcome the tendency of learning models to learn more concentrated patterns in the data and improve the value of model application.

[0012] In step 1), airborne QAR data for various types of flights nationwide is acquired and matched with existing airborne eddy current dissipation rate (EDR) samples. Specific steps include collecting airborne QAR data involving parameters such as vertical overload, radio altitude, indicated airspeed, angle of attack, flaps, pitch rate, roll angle, vertical speed, total weight, supersonic airspeed, Mach number, geographical location, wind speed, wind direction, and static air temperature. Based on the aircraft flight number and flight time, this data is compared and matched with EDR samples, and a 15-second time window is defined. Changes in QAR data within the window are labeled with corresponding EDR values, constructing a basic EDR dataset for airborne QAR data time windows across all air routes.

[0013] In step 2), the QAR data is loaded using the pandas library and checked to ensure that the IVV and Alt columns exist. The statistical averages (mean μ and standard deviation σ) of IVV and Alt are calculated. Then, a state machine is used to classify each data point as climb, cruise, or descent, starting from the initial climb state at takeoff.

[0014] In step 3), to better accommodate the turbulence characteristics during takeoff and landing, it is necessary to decouple the QAR parameters. The turbulence components are separated from the takeoff and landing phases and treated as the primary features. A high-pass filter is then used to remove low-frequency maneuvering components while preserving the turbulence frequency range. During the decomposition process:

[0015] To separate the turbulent components, the vertical wind speed of the aircraft relative to the atmosphere is calculated using the following formula:

[0016] ω z=--IVV-TAS(cos(Pitch)·sin(AOA)·cos(Roll)-cos(AOA)sin(Pitch))(1)

[0017] The high-pass filter is set to a fourth-order Baltworth high-pass filter with a frequency of 0.1 Hz. The filtering process can be represented as follows:

[0018] ω′ z =highpass(ω z ,0.1,fs) (2)

[0019] Where fs is the sampling frequency. This method uses a sliding window average subtraction (10-second window), which is very suitable for low sampling rate data.

[0020] To enhance the robustness of the model in turbulence identification, dynamic and static auxiliary features are extracted, and vertical acceleration is calculated based on VRTG:

[0021] a z =(VRTG-1)·9.81 (3)

[0022] VRTG is the normal acceleration gravity coefficient.

[0023] The same high-pass filter (cutoff frequency of 0.1 Hz) was used to extract high-frequency components related to turbulence, reflecting acceleration fluctuations caused by turbulence; for pitch a′ z The pitch rate, pitch angle, and roll angle were subjected to the same high-pass filtering to extract attitude changes caused by turbulence; and interpolations of AOAL and AOAR were calculated to capture aerodynamic disturbances. The same high-pass filtering (cutoff frequency of 0.1 Hz) was used to extract high-frequency components related to turbulence, reflecting acceleration fluctuations caused by turbulence; for pitch rate a′ z The pitch and roll angles are subjected to the same high-pass filtering to extract attitude changes caused by turbulence; and the interpolation of AOAL and AOAR is calculated to capture aerodynamic disturbances.

[0024] In step 4), during the cruise phase, the aerodynamic state is stable, and the changes in sensor parameters are mainly affected by turbulent disturbances. Therefore, a set of combined parameters based on QAR data is developed specifically for the cruise phase, including energy altitude, Pitch-CAS coupling, MACH-SAT coupling, and turbulent energy attenuation AOA difference.

[0025] Energy level: Energy level combines potential energy (Alt) and kinetic energy (Tas). 2 / 2g), equivalent to the total energy of altitude in Bernoulli's equation (Formula 4), after removing the static pressure term, represents the total energy per unit mass of the aircraft, and incorporates altitude changes into the energy measurement, thus correcting the altitude dependence error implicit in traditional methods. The calculation formula is as follows:

[0026]

[0027] Pitch-CAS coupling: Pitch is the pitch angle, and CAS is the calibrated airspeed. The Pitch-CAS coupling term (Equation 5) amplifies the synergistic effect between control requirements and turbulent excitation through a product term, compensating for the deficiencies of the default autonomous driving closed-loop control. The calculation formula is as follows:

[0028] Pitch and Cas Coupling=pitch×Cas (5)

[0029] MACH-SAT coupling describes the synergistic amplification of velocity-temperature fluctuations. The Mach-SAT coupling (Equation 6) corrects the conventional spectral model through temperature fluctuations, which improves the accuracy of EDR estimation at high Mach numbers. The calculation formula is as follows:

[0030] Mach and SAT Coupling=mach×SAT (6)

[0031] Turbulent Energy Decay: Cornman's EDR calculation is essentially an energy flux density (Equation 7), which characterizes the average energy transfer by integrating in the frequency domain. TED represents the vertical overload (G 2 By combining the instantaneous power density (P∝a) with the airspeed (Tas) (Formula 8), the instantaneous power density (P∝a) can be directly constructed. 2 V), which is closer to the dynamics of the turbulence-aircraft interaction than Equation 7. The calculation formula is as follows:

[0032]

[0033] TED = G 2 ×Tas (8)

[0034] AOA Difference: AOA represents the aircraft's angle of attack, with L and R representing the left and right wings, respectively. The change in angle of attack quantifies the spatial asymmetry of turbulent vortices (e.g., the risk of unilateral stall), breaking the traditional assumption of symmetry. This parameter is related to the change in vertical acceleration. These two technologies complement each other and are used for early warning of lateral energy gradients. The calculation formula is as follows:

[0035] ΔAOA=AOA L -AOA R (9)

[0036] The above parameter design not only inherits the rigor of the Cornman method, but also upgrades EDR estimation from a "turbulence detection tool" to an "aircraft-atmosphere energy interaction diagnostic system" through multi-physics coupling modeling.

[0037] 7. In step 5), a ConvLSTM-ELM model is constructed. The model input is specially configured, multi-scale spatiotemporal feature extraction is introduced, residual connections are enhanced, the ELM is strengthened, and an attention mechanism is introduced. Furthermore, a high EDR value training mechanism is adopted to make the model focus on more dangerous strong turbulence events. The following is a detailed explanation:

[0038] Input design: The model accepts a three-dimensional spacetime tensor X∈R H×W×C This includes H=15 (15-second time window), W=5 (multi-dimensional airborne sensor parameters), and C=1 (channel). This input structure preserves the dynamic response of the aircraft when encountering 15×5 turbulence and can clearly show the changes in airborne sensor data caused by turbulence.

[0039] Multi-scale spatiotemporal feature extraction: Two parallel ConvLSTM branches

[22] with different kernel sizes (2×2) and (4×4) are used to capture turbulence patterns at different scales. The fine-grained branch detects local fluctuations (e.g., sudden oscillations), while the coarse-grained branch identifies persistent disturbances (e.g., crosswind effects). The two branches are then merged, as shown in the following formula:

[0040] F merged =[F 2×2 F 4×4 ]∈R H×W×64 (10)

[0041] Enhanced residual connections: While spatiotemporally dense layers project merged features into the latent space, residual connections alleviate the vanishing gradient problem in deeper layers. Residual connection F res A lightweight transformation is introduced to stabilize training; the specific formula is as follows:

[0042] F out =Dropout(F+F res (11)

[0043] F res =Dropout(Dense(F)) (12)

[0044] To prevent overfitting, the Dropout value for each branch is set separately: the Dropout value for the main path is 0.3 (strong regularization), and the Dropout value for the residual path is 0.1 (weak perturbation).

[0045] Enhanced ELM: ELM-inspired layers freeze initial weights to random values, simulating the stochastic feature mapping W of an extreme learning machine. elm Then, the trainable attention gate reweights the ELM output to emphasize key turbulence metrics:

[0046] A = Softmax(W a ·F elm (13)

[0047] F attn =A⊙F elm (14)

[0048] Compared to traditional ELM models, trainable attention mechanisms overcome the static limitations of ELM.

[0049] The bottleneck layer compresses the attention-enhanced features to 64 dimensions, reducing the risk of overfitting while preserving high-value EDR features. The final screening rate (0.25) balances underfitting and overfitting in the imbalanced dataset.

[0050] High EDR value training mechanism: Initial weight calculation, high weights force the model to prioritize fitting high-risk turbulence samples, even at the cost of sacrificing the accuracy of some low-value samples, in accordance with the principle of "better to miss than to overshoot" in aviation safety.

[0051]

[0052] Subsequently, the cosine annealing algorithm was applied to adjust the weights. The cosine annealing algorithm maintains high weights at the beginning of training to quickly adapt to high-value samples, and gradually reduces the weights in the later stage to moderately prevent overfitting. Its smooth transition characteristics (continuity of the first derivative) enhance the optimization stability.

[0053]

[0054] The Huber loss function is chosen because it can better focus on significant values, enabling the model to capture high EDR values ​​during training.

[0055]

[0056] The RectifiedAdam optimizer is chosen as the optimizer. Compared to the Adam optimizer, it solves the gradient bias problem under dynamic weights, while the hot restart mechanism allows the learning rate to be reset periodically to avoid local optima (combined with cosine annealing). The dynamic weight callback module interacts with the ConvLSTM-ELM model through a weighted loss function, forming an end-to-end robust optimization framework.

[0057] In step 6), different evaluation criteria are set for different types of turbulence, as follows: For the low value area (EDR < 0.1), an absolute error limit of 0.01 is set, corresponding to stable airflow, and the error needs to be strictly controlled to avoid false alarms interfering with pilot operations; For the medium value area (0.1 ≤ EDR ≤ 0.3), moderate turbulence, attention is paid to overall regression performance (such as MSE, MAE, etc.); For the high value area (EDR > 0.3), high-risk events, the error limit (0.1) is relaxed, priority is given to ensuring high value capture rate, and no strong turbulence events are missed.

[0058] The deep learning method for EDR estimation based on aircraft response features provided by this invention has the following beneficial effects:

[0059] 1. This method outperforms traditional methods in estimating EDR values ​​for single flight paths, achieving higher accuracy. Its estimation results are closely related to turbulence characterization parameters such as vertical overload, effectively and meticulously depicting the evolution characteristics of atmospheric turbulence.

[0060] 2. The model achieved excellent fitting results during training and also performed well on the independent test set. By employing advanced feature extraction techniques, the model effectively avoided overfitting on the validation data, thus ensuring its excellent generalization performance.

[0061] 3. This model adopts dual-scale spatiotemporal feature extraction, which enhances the ability to capture strong turbulence, and weights the ELM attention mechanism, which improves the training speed by 15 times compared to the BP network. Its unique training mechanism further reduces the error rate of scarce samples.

[0062] 4. This method achieves a significant breakthrough in strong turbulence capture capability, low false alarm control capability, and computational efficiency, and has broad application value. Attached Figure Description

[0063] Figure 1 A schematic diagram of the deep learning method for EDR estimation based on aircraft response features provided by this invention;

[0064] Figure 2 This is a schematic diagram illustrating the use of staged processing to extract features from airborne QAR data within a time window in this invention.

[0065] Figure 3 This is a flowchart illustrating the calculation of EDR values ​​using a neural network model in this invention; Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0067] like Figure 1 As shown, the deep learning method for EDR estimation based on aircraft response features provided by this invention includes the following steps performed in sequence:

[0068] 1) This involves acquiring airborne QAR data from various types of flights nationwide and matching it with existing airborne eddy current dissipation rate (EDR) samples. Specific steps include collecting airborne QAR data involving parameters such as vertical overload, radio altitude, indicated airspeed, angle of attack, flaps, pitch rate, roll angle, vertical speed, total weight, supersonic airspeed, Mach number, geographical location, wind speed, wind direction, and static air temperature. Based on the aircraft flight number and flight time, this data is compared and matched with EDR samples, and a 15-second time window is defined. Changes in QAR data within the window are labeled with corresponding EDR values, constructing a basic EDR dataset for airborne QAR data time windows across all air routes. This invention relates to a total of 31 parameters, detailed in Table 1.

[0069] Table 1

[0070]

[0071]

[0072]

[0073] 2) Load the QAR data using the pandas library and check to ensure that the IVV and Alt columns exist. Calculate the statistical averages (mean μ and standard deviation σ) of IVV and Alt. Then, use a state machine to classify each data point as climb, cruise, or descent, starting from the initial climb state at takeoff.

[0074] 3) To better accommodate the turbulence characteristics during takeoff and landing, it is necessary to decouple the QAR parameters. The turbulence components are separated from the takeoff and landing phases and treated as the primary features. A high-pass filter is then used to remove low-frequency maneuvering components while preserving the turbulence frequency range. During the decomposition process:

[0075] To separate the turbulent components, the vertical wind speed of the aircraft relative to the atmosphere is calculated using the following formula:

[0076] ω z =-IVV-TAS(cos(Pitch)·sin(AOA)·coS(Roll)-coS(AOA)sin(Pitch))(1)

[0077] The high-pass filter is set to a fourth-order Baltworth high-pass filter with a frequency of 0.1 Hz. The filtering process can be represented as follows:

[0078] ω′ Z =highpass(ω Z ,0.1,fs) (2)

[0079] Where fs is the sampling frequency. This method uses a sliding window average subtraction (10-second window), which is very suitable for low sampling rate data.

[0080] To enhance the robustness of the model in turbulence identification, dynamic and static auxiliary features are extracted, and vertical acceleration is calculated based on VRTG:

[0081] a Z =(VRTG-1)·9.81 (3)

[0082] VRTG is the normal acceleration gravity coefficient.

[0083] The same high-pass filter (cutoff frequency of 0.1 Hz) was used to extract high-frequency components related to turbulence, reflecting acceleration fluctuations caused by turbulence; for pitch a′ Z The pitch rate, pitch angle, and roll angle were subjected to the same high-pass filtering to extract attitude changes caused by turbulence; and interpolations of AOAL and AOAR were calculated to capture aerodynamic disturbances. The same high-pass filtering (cutoff frequency of 0.1 Hz) was used to extract high-frequency components related to turbulence, reflecting acceleration fluctuations caused by turbulence; for pitch rate a′ Z The pitch and roll angles are subjected to the same high-pass filtering to extract attitude changes caused by turbulence; and the interpolation of AOAL and AOAR is calculated to capture aerodynamic disturbances.

[0084] 4) During the cruise phase, the aerodynamic state is stable, and the changes in sensor parameters are mainly affected by turbulent disturbances. Therefore, a set of combined parameters based on QAR data was developed specifically for the cruise phase, including energy altitude, Pitch-CAS coupling, MACH-SAT coupling, and turbulent energy attenuation AOA difference.

[0085] Energy level: Energy level combines potential energy (Alt) and kinetic energy (Tas). 2 / 2g), equivalent to the total energy of altitude in Bernoulli's equation (Formula 4), after removing the static pressure term, represents the total energy per unit mass of the aircraft, and incorporates altitude changes into the energy measurement, thus correcting the altitude dependence error implicit in traditional methods. The calculation formula is as follows:

[0086]

[0087] Pitch-CAS Coupling: Pitch is the pitch angle, and CAS is the calibrated airspeed. The Pitch-CAS coupling term (Equation 5) amplifies the synergistic effect between control requirements and turbulent excitation through a product term, compensating for the deficiencies of the default autonomous driving closed-loop control. The calculation formula is as follows:

[0088] Pitch and Cas Coupling=pitch×Cas (5)

[0089] MACH-SAT coupling describes the synergistic amplification of velocity-temperature fluctuations. The Mach-SAT coupling (Equation 6) corrects the conventional spectral model through temperature fluctuations, which improves the accuracy of EDR estimation at high Mach numbers. The calculation formula is as follows:

[0090] Mach and SAT Coupling=mach×SAT (6)

[0091] Turbulent Energy Decay: Cornman's EDR calculation is essentially an energy flux density (Equation 7), which characterizes the average energy transfer by integrating in the frequency domain. TED represents the vertical overload (G 2 By combining the instantaneous power density (P∝a) with the airspeed (Tas) (Formula 8), the instantaneous power density (P∝a) can be directly constructed. 2 V), which is closer to the dynamics of the turbulence-aircraft interaction than Equation 7. The calculation formula is as follows:

[0092]

[0093] TED = G 2 ×Tas (8)

[0094] AOA Difference: AOA represents the aircraft's angle of attack, with L and R representing the left and right wings, respectively. The change in angle of attack quantifies the spatial asymmetry of turbulent vortices (e.g., the risk of unilateral stall), breaking the traditional assumption of symmetry. This parameter is related to the change in vertical acceleration. These two technologies complement each other and are used for early warning of lateral energy gradients. The calculation formula is as follows:

[0095] ΔAOA=AOA L -AOA R (9)

[0096] The above parameter design not only inherits the rigor of the Cornman method, but also upgrades EDR estimation from a "turbulence detection tool" to an "aircraft-atmosphere energy interaction diagnostic system" through multi-physics coupling modeling.

[0097] 5) ConvLSTM-ELM model is constructed, with special settings for the model input, multi-scale spatiotemporal feature extraction, enhanced residual connections, enhanced ELM, and an attention mechanism introduced. Furthermore, a high EDR value training mechanism is employed to focus the model on more dangerous strong turbulence events. The following is a detailed explanation:

[0098] Input design: The model accepts a three-dimensional spacetime tensor X∈R H×W×C This includes H=15 (15-second time window), W=5 (multi-dimensional airborne sensor parameters), and C=1 (channel). This input structure preserves the dynamic response of the aircraft when encountering 15×5 turbulence and can clearly show the changes in airborne sensor data caused by turbulence.

[0099] Multi-scale spatiotemporal feature extraction: Two parallel ConvLSTM branches

[22] with different kernel sizes (2×2) and (4×4) are used to capture turbulence patterns at different scales. The fine-grained branch detects local fluctuations (e.g., sudden oscillations), while the coarse-grained branch identifies persistent disturbances (e.g., crosswind effects). The two branches are then merged, as shown in the following formula:

[0100] F merged =[F 2×2 F 4×4 ]∈R H×W×64 (10)

[0101] Enhanced residual connections: While spatiotemporally dense layers project merged features into the latent space, residual connections alleviate the vanishing gradient problem in deeper layers. Residual connection F res A lightweight transformation is introduced to stabilize training; the specific formula is as follows:

[0102] F out =Dropout(F+F res (11)

[0103] F res =Dropout(Dense(F)) (12)

[0104] To prevent overfitting, the Dropout value for each branch is set separately: the Dropout value for the main path is 0.3 (strong regularization), and the Dropout value for the residual path is 0.1 (weak perturbation).

[0105] Enhanced ELM: ELM-inspired layers freeze initial weights to random values, simulating the stochastic feature mapping W of an extreme learning machine. elm Then, the trainable attention gate reweights the ELM output to emphasize key turbulence metrics:

[0106] A = Softmax(Wa ·F elm (13)

[0107] F attn =A⊙F elm (14)

[0108] The bottleneck layer compresses the attention-enhanced features to 64 dimensions, reducing the risk of overfitting while preserving high-value EDR features. The final screening rate (0.25) balances underfitting and overfitting in the imbalanced dataset.

[0109] High EDR value training mechanism: Initial weight calculation, high weights force the model to prioritize fitting high-risk turbulence samples, even at the cost of sacrificing the accuracy of some low-value samples, in accordance with the principle of "better to miss than to overshoot" in aviation safety.

[0110]

[0111] Subsequently, the cosine annealing algorithm was applied to adjust the weights. The cosine annealing algorithm maintains high weights at the beginning of training to quickly adapt to high-value samples, and gradually reduces the weights in the later stage to moderately prevent overfitting. Its smooth transition characteristics (continuity of the first derivative) enhance the optimization stability.

[0112]

[0113] The Huber loss function is chosen because it can better focus on significant values, enabling the model to capture high EDR values ​​during training.

[0114]

[0115] The RectifiedAdam optimizer is chosen as the optimizer. Compared to the Adam optimizer, it solves the gradient bias problem under dynamic weights, while the hot restart mechanism allows the learning rate to be reset periodically to avoid local optima (combined with cosine annealing). The dynamic weight callback module interacts with the ConvLSTM-ELM model through a weighted loss function, forming an end-to-end robust optimization framework.

[0116] 6) Set different evaluation criteria for different types of turbulence, as follows: For the low value area (EDR < 0.1), set an absolute error limit of 0.01, corresponding to stable airflow, and strictly control the error to avoid false alarms interfering with pilot operations; For the medium value area (0.1 ≤ EDR ≤ 0.3), moderate turbulence, focus on overall regression performance (such as MSE, MAE, etc.); For the high value area (EDR > 0.3), high-risk events, relax the error limit (0.1), prioritize ensuring high value capture rate, and do not miss strong turbulence events.

[0117] Next, these high-level features are fed into a multilayer neural network model with optimal weights and biases to obtain an estimate of the eddy current dissipation rate (EDR).

[0118] The above description is a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications and substitutions made by those skilled in the art within the scope of the technology disclosed in the present invention should be covered within the protection scope of the present invention.

Claims

1. A deep learning method for EDR estimation based on aircraft response features, characterized in that: The method for mining full-route turbulence features and estimating EDR using airborne QAR data includes the following steps performed in sequence: S1: Systematically integrate flight data collected by Quick Access Recorders (QARs) from all flights nationwide, and perform spatiotemporal correlation analysis with existing aircraft eddy current dissipation (EDR) monitoring samples. By constructing time windows and performing intelligent hierarchical sampling, the dynamic characteristics of turbulence can be captured more effectively while ensuring the generalization ability of the trained model. S2: Develop an adaptive flight phase segmentation algorithm that processes QAR data in CSV format and uses IW and altitude data to divide flight phases, thereby estimating EDR data for each flight phase, effectively isolating the turbulent response of each flight phase, thus improving the accuracy of EDR estimation, clearly displaying each flight phase through visualization and performing specific analysis on each flight phase; S3: During aircraft takeoff and landing, the QAR parameters of the response to turbulence caused by aerodynamic characteristics, maneuvering, and environmental disturbances contain a significant non-turbulent component. To accurately extract turbulent features, a decoupling method based on vertical wind speed is proposed to extract auxiliary features; S4: During the cruise phase, the aerodynamic state is stable, and the changes in sensor parameters are mainly affected by turbulent disturbances. Therefore, a set of combined parameters based on QAR data is developed specifically for the cruise phase to obtain more accurate data. S5: Construct a ConvLSTM-ELM neural network model and introduce a high EDR value training mechanism to focus on more dangerous strong turbulence events; S6: Set different evaluation criteria for different types of bumps to overcome the tendency of learning models to learn more concentrated patterns in the data and improve the value of model application.

2. The deep learning method for EDR estimation based on aircraft response features as described in claim 1, characterized in that: In S1, airborne QAR data for various types of flights across the country is acquired and matched with existing airborne eddy current dissipation rate (EDR) samples. Specific steps include collecting airborne QAR data involving parameters such as vertical overload, radio altitude, indicated airspeed, angle of attack, flaps, pitch rate, roll angle, vertical speed, total weight, supersonic airspeed, Mach number, geographical location, wind speed, wind direction, and static air temperature. Based on the aircraft flight number and flight time, this data is compared and matched with EDR samples, and a 15-second time window is defined. Changes in QAR data within the window are labeled with corresponding EDR values, constructing a basic EDR dataset for airborne QAR data time windows across all air routes.

3. The deep learning method for EDR estimation based on aircraft response features according to claim 1, characterized in that: In S2, the pandas library is used to load the QAR data and check to ensure that the IVV and Alt columns exist. The statistical averages (mean μ and standard deviation σ) of IVV and Alt are calculated. Then, a state machine is used to classify each data point as climb, cruise, or descent, starting from the initial climb state at takeoff.

4. The deep learning method for EDR estimation based on aircraft response features according to claim 1, characterized in that: In S3, to better accommodate the turbulence characteristics during takeoff and landing, it is necessary to decouple the QAR parameters. The turbulence components are separated from the takeoff and landing phases and treated as the primary features. A high-pass filter is then used to remove low-frequency maneuvering components while preserving the turbulence frequency range. During the decomposition process: To separate the turbulent components, the vertical wind speed of the aircraft relative to the atmosphere is calculated using the following formula: ω z =-IVV-TAS(coS(Pitch)·sin(AOA)·cos(Roll)-cos(AOA)sin(Pitch))(1) The high-pass filter is set to a fourth-order Baltworth high-pass filter with a frequency of 0.1 Hz. The filtering process can be represented as follows: ω′ z <highpass(ω) z ,0.1,fs) (2) Where fS is the sampling frequency. This method uses a sliding window average subtraction (10-second window), which is very suitable for low sampling rate data. To enhance the robustness of the model in turbulence identification, dynamic and static auxiliary features are extracted, and vertical acceleration is calculated based on VRTG: the z =(VRTG-1)·9.81 (3) VRTG is the normal acceleration gravity coefficient. The same high-pass filter (cutoff frequency of 0.1 Hz) was used to extract high-frequency components related to turbulence, reflecting acceleration fluctuations caused by turbulence; for pitch a′ z The pitch rate, pitch angle, and roll angle were subjected to the same high-pass filtering to extract attitude changes caused by turbulence; and interpolations of AOAL and AOAR were calculated to capture aerodynamic disturbances. The same high-pass filtering (cutoff frequency of 0.1 Hz) was used to extract high-frequency components related to turbulence, reflecting acceleration fluctuations caused by turbulence; for pitch rate a′ z The pitch and roll angles are subjected to the same high-pass filtering to extract attitude changes caused by turbulence; and the interpolation of AOAL and AOAR is calculated to capture aerodynamic disturbances.

5. The deep learning method for EDR estimation based on aircraft response features according to claim 1, characterized in that: In S4, during the cruise phase, the aerodynamic state is stable, and the changes in sensor parameters are mainly affected by turbulent disturbances. Therefore, a set of combined parameters based on QAR data was developed specifically for the cruise phase, including energy altitude, Pitch-CAS coupling, MACH-SAT coupling, and turbulent energy attenuation AOA difference. Energy level: Energy level combines potential energy (Alt) and kinetic energy (Tas). 2 / 2g), equivalent to the total energy of altitude in Bernoulli's equation (Formula 4), after removing the static pressure term, represents the total energy per unit mass of the aircraft, and incorporates altitude changes into the energy measurement, thus correcting the altitude dependence error implicit in traditional methods. The calculation formula is as follows: Pitch-CAS Coupling: Pitch is the pitch angle, and CAS is the calibrated airspeed. The Pitch-CAS coupling term (Equation 5) amplifies the synergistic effect between control requirements and turbulent excitation through a product term, compensating for the deficiencies of the default autonomous driving closed-loop control. The calculation formula is as follows: Pitch and Cas Coupling=pitch × Cas (5) MACH-SAT coupling describes the synergistic amplification of velocity-temperature fluctuations. The Mach-SAT coupling (Equation 6) corrects the conventional spectral model through temperature fluctuations, which improves the accuracy of EDR estimation at high Mach numbers. The calculation formula is as follows: Mach and SAT Coupling=mach×SAT (6) Turbulent Energy Decay: Cornman's EDR calculation is essentially an energy flux density (Equation 7), which characterizes the average energy transfer by integrating in the frequency domain. TED represents the vertical overload (G 2 By combining the instantaneous power density (P∝a) with the airspeed (Tas) (Equation 8), the instantaneous power density (P∝a) can be directly constructed. 2 V), which is closer to the dynamics of the turbulence-aircraft interaction than Equation 7. The calculation formula is as follows: TED=G 2 × Tas (8) AOA Difference: AOA represents the aircraft's angle of attack, with L and R representing the left and right wings, respectively. The change in angle of attack quantifies the spatial asymmetry of turbulent vortices (e.g., the risk of unilateral stall), breaking the traditional assumption of symmetry. This parameter is related to the change in vertical acceleration. These two technologies complement each other and are used for early warning of lateral energy gradients. The calculation formula is as follows: ΔAOA=AOA L -Yes R (9) The above parameter design not only inherits the rigor of the Cornman method, but also upgrades EDR estimation from a "turbulence detection tool" to an "aircraft-atmosphere energy interaction diagnostic system" through multi-physics coupling modeling.

6. The deep learning method for EDR estimation based on aircraft response features according to claim 1, characterized in that: In S5, a ConvLSTM-ELM model is constructed. The model input is specially configured, multi-scale spatiotemporal feature extraction is introduced, residual connections are enhanced, the ELM is strengthened, and an attention mechanism is introduced. Furthermore, a high EDR value training mechanism is employed to focus the model on more dangerous, strong turbulence events. The following is a detailed explanation: Input design: The model accepts a three-dimensional spacetime tensor X∈R H×W×C This includes H=15 (15-second time window), W=5 (multi-dimensional airborne sensor parameters), and C=1 (channel). This input structure preserves the dynamic response of the aircraft when encountering 15×5 turbulence and can clearly show the changes in airborne sensor data caused by turbulence. Multi-scale spatiotemporal feature extraction: Two parallel ConvLSTM branches [22] with different kernel sizes (2×2) and (4×4) are used to capture turbulence patterns at different scales. The fine-grained branch detects local fluctuations (e.g., sudden oscillations), while the coarse-grained branch identifies persistent disturbances (e.g., crosswind effects). The two branches are then merged, as shown in the following formula: F merged =[F 2×2 ,F 4×4 ]∈R H×W×64 (10) Enhanced residual connections: While spatiotemporally dense layers project merged features into the latent space, residual connections alleviate the vanishing gradient problem in deeper layers. Residual connection F res A lightweight transformation is introduced to stabilize training; the specific formula is as follows: F our =Dropout(F+F res ) (11) F res =Dropout(Dense(F)) (12) To prevent overfitting, the Dropout value for each branch is set separately: the Dropout value for the main path is 0.3 (strong regularization), and the Dropout value for the residual path is 0.1 (weak perturbation). Enhanced ELM: ELM-inspired layers freeze initial weights to random values, simulating the stochastic feature mapping W of an extreme learning machine. elm Then, the trainable attention gate reweights the ELM output to emphasize key turbulence metrics: A=Softmax(W a ·F elm ) (13) F attn =A⊙F elm (14) The bottleneck layer compresses the attention-enhanced features to 64 dimensions, reducing the risk of overfitting while preserving high-value EDR features. The final screening rate (0.25) balances underfitting and overfitting in the imbalanced dataset. High EDR value training mechanism: Initial weight calculation, high weights force the model to prioritize fitting high-risk turbulence samples, even at the cost of sacrificing the accuracy of some low-value samples, in accordance with the principle of "better to miss than to overshoot" in aviation safety. Subsequently, the cosine annealing algorithm was applied to adjust the weights. The cosine annealing algorithm maintains high weights at the beginning of training to quickly adapt to high-value samples, and gradually reduces the weights in the later stage to moderately prevent overfitting. Its smooth transition characteristics (continuity of the first derivative) enhance the optimization stability. The Huber loss function is chosen because it can better focus on significant values, enabling the model to capture high EDR values ​​during training. The RectifiedAdam optimizer is chosen as the optimizer. Compared to the Adam optimizer, it solves the gradient bias problem under dynamic weights, while the hot restart mechanism allows the learning rate to be reset periodically to avoid local optima (combined with cosine annealing). The dynamic weight callback module interacts with the ConvLSTM-ELM model through a weighted loss function, forming an end-to-end robust optimization framework.

7. The deep learning method for EDR estimation based on aircraft response features according to claim 1, characterized in that: In S6, different evaluation criteria are set for different types of turbulence, as follows: For the low value area (EDR < 0.1), an absolute error limit of 0.01 is set, corresponding to stable airflow, and the error must be strictly controlled to avoid false alarms interfering with pilot operations; For the medium value area (0.1 ≤ EDR ≤ 0.3), moderate turbulence, attention is paid to overall regression performance (such as MSE, MAE, etc.); For the high value area (EDR > 0.3), high-risk events, the error limit (0.1) is relaxed, prioritizing the high value capture rate and not missing strong turbulence events.

Images