Aircraft buffet prediction method, system and device based on decoupled physical constraints

Abstract

The present application relates to the technical field of aviation safety and weather prediction, and particularly relates to a method, system and device for predicting aircraft jolt based on decoupled physical constraints; the pre-collected multi-time sequence weather forecast data and aircraft EDR data are aligned on the same space-time grid and converted into probabilistic features containing mean and standard deviation; for sparse aircraft EDR data, an atmospheric physical model correction term based on the relationship between wind shear and turbulent energy dissipation rate is introduced to fill in, and a probabilistic input tensor is obtained; the probabilistic input tensor is input into a self-correlation attention model with fused physical constraints which is pre-constructed and trained; a contribution matrix is extracted from the attention weight of the trained model, the contribution proportion of each prediction time point to the prediction result is quantified, and the probabilistic prediction result of the aircraft jolt and the corresponding contribution matrix are output as the final prediction result; a reliable basis is provided for flight decision, and the accuracy and robustness of aircraft jolt prediction are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of aviation safety and meteorological prediction, and in particular to a method, system and device for predicting aircraft turbulence based on decoupled physical constraints. BACKGROUND

[0002] Aircraft turbulence is mainly caused by atmospheric turbulence, which poses a significant threat to flight safety, passenger comfort and operational efficiency. Traditional turbulence prediction methods mainly rely on numerical weather prediction (NWP) models and diagnostic models based on empirical rules. However, these methods have the following limitations: (1) NWP models require large computational resources and are difficult to provide real-time predictions with high temporal and spatial resolution; (2) empirical models lack accurate descriptions of complex nonlinear atmospheric processes; (3) existing methods usually provide deterministic predictions and cannot quantify the uncertainty of the predictions, making it difficult to support risk refinement and control decisions.

[0003] In recent years, with the development of artificial intelligence technology, especially the successful application of deep learning in the field of time series prediction, a new approach to turbulence prediction has been provided. For example, models based on recurrent neural networks (RNN), long short-term memory networks (LSTM) and Transformers have been tried for weather prediction. Among them, the Autoformer model effectively captures long-term dependencies and periodic patterns in time series through self-correlation attention mechanisms, and performs well in time series prediction tasks.

[0004] However, directly applying pure data-driven AI models to aviation turbulence prediction faces great challenges. First, the EDR (turbulence energy dissipation rate) data reported by aircraft is extremely sparse in space and time, and simple data interpolation will introduce large errors and violate physical laws. Second, atmospheric motion follows strict physical laws (such as the Navier-Stokes equation), and pure data-driven models may learn false correlations that violate physical laws, resulting in physically unreliable predictions, especially in areas with scarce training data. Finally, the aviation industry has extremely high requirements for safety and explainability, and the single prediction value provided by black box models cannot meet the needs of flight decision-making for risk confidence.

[0005] Therefore, there is an urgent need for a new aircraft turbulence prediction method that can integrate physical prior knowledge, handle data uncertainty, and provide explainable prediction results. SUMMARY

[0006] In order to overcome the deficiencies in the prior art, the application provides an aircraft bump prediction method, system and device based on decoupled physical constraints.

[0007] To achieve the above object, the application adopts the following technical solutions:

[0008] In a first aspect, the application provides an aircraft bump prediction method based on decoupled physical constraints, which comprises:

[0009] Aligning the pre-collected multi-time series weather forecast data and aircraft EDR data on the same spatio-temporal grid, and converting them into probabilistic features containing mean and standard deviation;

[0010] For sparse aircraft EDR data, introduce an atmospheric physical model correction term based on the relationship between wind shear and turbulent energy dissipation rate to fill in, and obtain a probabilistic input tensor;

[0011] Input the probabilistic input tensor into a pre-established self-correlation attention model that integrates physical constraints to perform aircraft bump probabilistic prediction, and use the Gaussian negative log-likelihood as the loss function to train the model;

[0012] The self-correlation attention model that integrates physical constraints comprises an encoder, an attention layer and a decoder.

[0013] The encoder converts the probabilistic input tensor into one or more of a query matrix, a key matrix and a value matrix through linear mapping for self-correlation mapping;

[0014] The attention layer calculates the cross-correlation function based on the query matrix and the key matrix, constructs an attention score matrix, and introduces decoupled physical constraints, which include a global physical attenuation matrix calculated according to the time difference and the spherical distance, and a trainable local fine-tuning matrix for capturing local weather bias;

[0015] The decoder takes the physical constraint self-correlation feature sequence as input and outputs the probabilistic prediction results of the mean and standard deviation of the EDR at multiple future time steps;

[0016] Extract the contribution matrix from the attention weight of the trained model, quantify the contribution proportion of each prediction time point to the prediction result, and output the probabilistic prediction result of the aircraft bump and the corresponding contribution matrix as the final result of the prediction.

[0017] Optionally, the aligning the pre-collected multi-temporal weather forecast data and the aircraft EDR data on the same spatio-temporal grid and converting the multi-temporal weather forecast data and the aircraft EDR data into the probabilistic features including the mean and the standard deviation comprises: converting a time stamp of the multi-temporal weather forecast data into a UTC time; taking a recording time of the aircraft EDR data as a reference, matching weather forecast data with a time difference less than a preset threshold for each EDR record point; using spherical linear interpolation on the matched weather forecast data, interpolating a spatial resolution of the weather forecast data to a latitude and longitude position of the EDR record point to obtain the aligned original data; and performing probabilistic feature representation on the aligned original data to obtain a probabilistic feature vector on each spatio-temporal grid point.

[0018] The probabilistic feature vector includes an EDR mean component, an EDR standard deviation component, a weather mean component, and a weather standard deviation component.

[0019] Optionally, the aligning the pre-collected multi-temporal weather forecast data and the aircraft EDR data on the same spatio-temporal grid and converting the multi-temporal weather forecast data and the aircraft EDR data into the probabilistic features including the mean and the standard deviation comprises: converting a time stamp of the multi-temporal weather forecast data into a UTC time; taking a recording time of the aircraft EDR data as a reference, matching weather forecast data with a time difference less than a preset threshold for each EDR record point; using spherical linear interpolation on the matched weather forecast data, interpolating a spatial resolution of the weather forecast data to a latitude and longitude position of the EDR record point to obtain the aligned original data; and performing probabilistic feature representation on the aligned original data to obtain a probabilistic feature vector on each spatio-temporal grid point.

[0020] The EDR mean component, the EDR standard deviation component, the weather mean component, and the weather standard deviation component are spliced to form the probabilistic feature vector.

[0021] Optionally, the introducing an atmospheric physics model correction term based on the relationship between wind shear and turbulent energy dissipation rate to fill in the sparse aircraft EDR data to obtain the probabilistic input tensor comprises:

[0022] The probabilistic feature vectors of the known points are interpolated to obtain an interpolation mean and an interpolation standard deviation.

[0023] The unknown point probabilistic feature vectors with missing EDR values are filled in based on the correction term of the atmospheric physics model; wherein the correction term of the atmospheric physics model includes a physical relationship between wind shear and turbulent energy dissipation rate.

[0024] The Richardson number is calculated according to the wind shear and the temperature gradient, and the Richardson number is mapped to a turbulent energy dissipation rate estimate; the turbulent energy dissipation rate estimate is superimposed on the interpolation mean, and based on the interpolation mean and the interpolation standard deviation of the superimposed turbulent energy dissipation rate estimate, the probabilistic input tensor is obtained.

[0025] Optionally, the training of the self-correlation attention model fused with physical constraints comprises:

[0026] A global physical attenuation matrix is defined based on the atmospheric vortex time and space exponential decay law;

[0027] A local fine-tuning matrix for capturing local meteorological bias is established;

[0028] An autocorrelation attention mechanism is improved by taking Autoformer as a basic framework and taking the global physical attenuation matrix and the local fine-tuning matrix as decoupled physical constraints to construct an autocorrelation attention model fused with physical constraints;

[0029] In the model training stage, the model parameters are optimized through back propagation and gradient update to obtain the trained autocorrelation attention model fused with physical constraints.

[0030] Optionally, the global physical attenuation matrix defined based on the atmospheric vortex time and space exponential decay law comprises:

[0031] The time difference and spherical distance between any two points in the space-time grid are read;

[0032] The time difference is divided by a preset time attenuation constant, the spherical distance is divided by a preset space attenuation constant, and the negative exponent of the sum is obtained to obtain the global physical attenuation matrix element.

[0033] Optionally, in the model training stage, the model parameters are optimized through back propagation and gradient update, including initializing the local fine-tuning matrix as a full zero matrix, calculating the attention weight through forward propagation in each iteration training, calculating the Gaussian negative log-likelihood loss by combining the real EDR value, calculating the gradient of the Gaussian negative log-likelihood loss on the full zero matrix through back propagation, and applying gradient update to the elements of the full zero matrix.

[0034] Optionally, the contribution matrix is extracted from the attention weight, the contribution proportion of each prediction time point to the prediction result is quantified, and the probabilistic prediction result of the aircraft turbulence and the contribution matrix are output, including grouping and summing the attention weight according to the prediction time point index to obtain the total attention score of each prediction time point; calculating the proportion of the total attention score of each prediction time point in the total attention score of all features to obtain the normalized contribution proportion; filling the normalized contribution proportion into the corresponding position of the contribution matrix to generate the contribution matrix; each element of the contribution matrix represents the relative contribution proportion of the corresponding prediction time point to the prediction result in the prediction time step; and the contribution matrix and the probabilistic prediction result are output.

[0035] In a second aspect, the present application provides an aircraft turbulence prediction system based on decoupled physical constraints, which comprises:

[0036] An integration module is configured to align the pre-acquired multi-time weather forecast data and the aircraft EDR data on the same spatio-temporal grid and convert them into probabilistic features containing mean and standard deviation;

[0037] A data padding module is configured to introduce an atmospheric physics model correction term based on the relationship between wind shear and turbulent energy dissipation rate to fill the sparse aircraft EDR data and obtain a probabilistic input tensor;

[0038] A model training module is configured to input the probabilistic input tensor into a pre-constructed and trained self-correlation attention model with fusion of physical constraints;

[0039] A result output module is configured to extract a contribution matrix from the attention weight of the trained model, quantify the contribution proportion of each prediction time point to the prediction result, and output the probabilistic prediction result of the aircraft buffet and the corresponding contribution matrix as the final prediction result.

[0040] In a third aspect, the present application provides an electronic device, which comprises:

[0041] at least one processor; and

[0042] a memory connected to the at least one processor in communication; wherein,

[0043] The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to execute the method of any one of the first aspect.

[0044] Compared with the closest prior art, the present application has the beneficial effects that:

[0045] The aircraft buffet prediction method, system and device based on decoupled physical constraints provided by the present application introduce decoupled physical constraints (global fixed matrix + local adjustable matrix), directly embed the physical law of atmospheric vortex decay into the learning process of the model, ensure that the model prediction follows the basic physical law, effectively avoid the physical absurd results that may be produced by the pure data-driven model, and significantly improve the prediction accuracy in the data sparse area and extreme conditions.

[0046] The present application uses probabilistic features as input and output, and uses Gaussian negative log-likelihood loss function for training, so that the model can spontaneously learn and output the prediction uncertainty (standard deviation). This provides more rich risk information for flight decision, supports more fine risk assessment and management.

[0047] The application uses a filling method based on physical prior to process sparse EDR data, which can generate more physically realistic data compared to traditional mathematical interpolation, providing a higher quality training basis for the model and improving overall performance.

[0048] The application can clearly quantify the contribution proportion of information of different forecast time to the final prediction result by extracting the contribution matrix from the attention mechanism optimized from the physical constraint. This interpretability enables pilots and forecasters to understand the basis for the model's decision-making, enhancing the trust in the prediction results.

[0049] In the application, the decoupled design is used to fix the priori knowledge in the global matrix, providing a stable optimization direction for the model and alleviating the gradient instability problem in long sequence prediction. At the same time, the local adjustable matrix gives the model flexibility to adapt to complex local phenomena (such as terrain effect, thunderstorm), enhancing the generalization ability of the model. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the specific embodiments or prior art of the present application, the following will briefly introduce the drawings needed in the specific embodiments or prior art description. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual proportions.

[0051] Figure 1 is a flow chart of an aircraft turbulence prediction method based on decoupled physical constraints provided by the present application;

[0052] Figure 2 is a structural schematic diagram of an aircraft turbulence prediction system based on decoupled physical constraints provided by the present application;

[0053] Figure 3 is an internal structure diagram of an electronic device provided by the present application. DETAILED DESCRIPTION

[0054] The embodiments of the technical solutions of the present application will be described in detail below with reference to the drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present application, therefore, only as an example, and cannot limit the protection scope of the present application.

[0055] It should be noted that, unless otherwise specified, the technical terms or scientific terms used in this application should be understood as the usual meaning understood by the skilled person in the field to which the present application belongs.

[0056] The application provides a method, system and device for predicting aircraft turbulence based on decoupled physical constraints. The method integrates multi-temporal weather data and aircraft EDR (flight data recorder) data, and not only uses a single numerical value, but also converts the data into probabilistic features containing mean and standard deviation. This processing method also takes into account the inherent uncertainty of the data in the model, making the prediction result more robust. In addition, for sparse or missing EDR data, atmospheric physics prior knowledge is used for filling instead of simple mathematical interpolation. This ensures that the input data complies with the basic physical laws from the beginning, laying a solid physical foundation for the subsequent model learning.

[0057] Secondly, the autocorrelation attention mechanism of the Autoformer is upgraded, and a decoupled physical constraint is introduced. This constraint consists of two parts: a fixed global physical decay matrix and a trainable local fine-tuning matrix. The global matrix is based on the exponential decay physical law of atmospheric vorticity, ensuring that the model always follows the physical common sense that "the closer the distance, the greater the influence". The local fine-tuning matrix allows the model to adaptively capture physical deviations caused by local terrain or sudden weather events. This decoupled design not only solves the problem of gradient instability in long sequence model training, but also perfectly combines the rigor of physics and the flexibility of data-driven models, significantly improving the accuracy and physical reasonableness of the prediction. Finally, the model output is not only the predicted value, but also the uncertainty (standard deviation) and the interpretable weight matrix, providing transparent and comprehensive basis for flight decision-making.

[0058] Embodiment 1: Please refer to Figure 1 , Figure 1 A method for predicting aircraft turbulence based on decoupled physical constraints is provided for the embodiment 1 of the application, which specifically comprises the following steps:

[0059] S101 align the pre-collected multi-temporal weather forecast data and aircraft EDR data on the same spatio-temporal grid, and convert them into probabilistic features containing mean and standard deviation;

[0060] S102 for sparse aircraft EDR data, introduce an atmospheric physics model correction term based on the relationship between wind shear and turbulent energy dissipation rate to fill in, and obtain a probabilistic input tensor;

[0061] S103 input the probabilistic input tensor into the autocorrelation attention model with physical constraints pre-established to predict the probability of aircraft turbulence, and use the Gaussian negative log-likelihood as the loss function to train the model;

[0062] S104 extracts a contribution matrix from the attention weight of the trained model, quantifies the contribution proportion of each prediction time point to the prediction result, and outputs the probabilistic prediction result of the aircraft turbulence and the corresponding contribution matrix as the final prediction result.

[0063] In the above embodiment, step S101 specifically comprises:

[0064] Convert the timestamp of the multi-time series weather forecast data into UTC time.

[0065] Take the recording time of the aircraft EDR data as the reference, and match the weather forecast data with a time difference less than a preset threshold for each EDR recording point;

[0066] The matched weather forecast data is interpolated using spherical linear interpolation to interpolate the spatial resolution of the weather forecast data to the latitude and longitude position of the EDR recording point, and the aligned original data is obtained.

[0067] The aligned original data is represented by a probabilistic feature, and a probabilistic feature vector is obtained at each spatio-temporal grid point; wherein the probabilistic feature vector includes an EDR mean component, an EDR standard deviation component, a weather mean component and a weather standard deviation component.

[0068] Further, the probabilistic feature representation is realized by: calculating the mean and standard deviation of the aircraft EDR data in a preset sliding time window to obtain the EDR mean component and the EDR standard deviation component; calculating the mean and standard deviation of the weather variable at each prediction time point in the same sliding time window to obtain the weather mean component and the weather standard deviation component; and splicing the above four components to form a probabilistic feature vector.

[0069] Wherein, the weather forecast data: for each prediction time point Its weather characteristics (such as wind speed, wind shear) are also input in the form of mean And standard deviation .

[0070] In the above step S102, for sparse aircraft EDR data, an atmospheric physics model correction term based on the relationship between wind shear and turbulent energy dissipation rate is introduced to fill in, and the probabilistic input tensor includes:

[0071] Interpolate the probabilistic feature vector of the known point (such as using Gaussian process regression or Kriging interpolation) to obtain the interpolation mean and interpolation standard deviation.

[0072] Based on the correction term of the atmospheric physics model, the probabilistic feature vector of the unknown point with missing EDR value is filled in. The correction term is constructed based on the physical relationship between wind shear and turbulent energy dissipation rate.

[0073] Specifically, a Richardson number is calculated according to the wind shear and the temperature gradient, and the Richardson number is mapped to an estimate of the turbulent energy dissipation rate.

[0074] The estimate of the turbulent energy dissipation rate and the interpolated mean are superimposed, and the interpolated standard deviation is retained, to form a probabilistic feature vector of the unknown point, and finally a complete probabilistic input tensor is formed: .

[0075] Where B is the batch size; L is the sequence length. is the enhanced feature dimension. The input vector of each time step t contains EDR historical data: mean and standard deviation .

[0076] The above step S103 inputs the probabilistic input tensor into a pre-established self-correlation attention model fused with physical constraints to perform aircraft jolt probabilistic prediction, and uses a Gaussian negative log-likelihood as a loss function to train the model.

[0077] Wherein the self-correlation attention model fused with physical constraints includes an encoder, an attention layer and a decoder.

[0078] The encoder converts the probabilistic input tensor into one or more of a query matrix, a key matrix and a value matrix through linear mapping for self-correlation mapping.

[0079] Optionally, the decoder includes: a first linear layer that maps the physical constraint self-correlation feature sequence to a hidden state;

[0080] a second linear layer that maps the hidden state to an EDR mean prediction vector and an EDR standard deviation prediction vector.

[0081] The attention layer calculates a cross-correlation function based on the query matrix and the key matrix, constructs an attention score matrix, and introduces a decoupled physical constraint, which includes a global physical attenuation matrix calculated according to a time difference and a spherical distance, and a trainable local fine-tuning matrix for capturing local meteorological bias.

[0082] The decoder takes the physical constraint self-correlation feature sequence as input and outputs probabilistic prediction results of the EDR mean and standard deviation at multiple future time steps.

[0083] In the above embodiment, the cross-correlation function of the query matrix and the key matrix includes:

[0084] performing a fast Fourier transform on each column of the query matrix;

[0085] performing a fast Fourier transform on each column of the key matrix and taking the conjugate;

[0086] multiplying the transformed query matrix with the elements of the key matrix, inverse fast Fourier transform, to obtain the cross-correlation function.

[0087] In one embodiment, the encoder converts the probabilistic input tensor into a query matrix Q, a key matrix K, and a value matrix V through a linear mapping.

[0088] In one embodiment, the attention layer is the core innovation of the model, which introduces decoupled physical constraints containing global physical decay matrix and local fine-tuning matrix The calculation process includes:

[0089] Based on the query matrix Q and the key matrix K, the cross-correlation function is calculated through the fast Fourier transform (FFT) and its inverse transform (IFFT) to obtain the initial attention score matrix.

[0090] The initial score matrix is added to the scaled global physical decay matrix and the local fine-tuning matrix

[0091] The added matrix is normalized by SoftMax to obtain the final attention weight. The value matrix V is weighted and summed using the attention weight to obtain the physically constrained autocorrelation feature sequence.

[0092] The global physical decay matrix is defined based on the time and spatial exponential decay law of atmospheric vortex, and its calculation formula is: ; wherein, is the time difference between time step t i and t j ; is the spherical distance between corresponding geographical positions; is the time decay constant, reflecting the average life cycle of atmospheric vortex in time, which can be initialized according to empirical values or atmospheric physical models; is the spatial decay constant, reflecting the average influence range of atmospheric vortex in space, which is also initialized by physical priori.

[0093] In the above embodiment, the global physical decay matrix is defined based on the time and spatial exponential decay law of atmospheric vortex, which includes:

[0094] Reading the time difference and spherical distance between any two points in the space-time grid;

[0095] ​​The time difference is divided by a preset time attenuation constant, the spherical distance is divided by a preset space attenuation constant, and the sum is taken to the negative exponent to obtain an element of a global physical attenuation matrix.

[0096] In the above embodiments, the local fine-tuning matrix is initialized as a full zero matrix and is updated through a gradient descent algorithm during model training to capture local deviations other than global physical laws.

[0097] The decoder takes the physical constraint autocorrelation feature sequence output by the attention layer as input, and outputs a prediction result of the future O time steps through forward propagation. The result includes the EDR mean and standard deviation .

[0098] Further, the training of the autocorrelation attention model fused with physical constraints in step S103 includes:

[0099] Based on the time and space exponential attenuation law of atmospheric vortices, a global physical attenuation matrix is defined;

[0100] A local fine-tuning matrix is established to capture local meteorological deviations;

[0101] Taking Autoformer as the basic architecture, the autocorrelation attention mechanism is improved by taking the global physical attenuation matrix and the local fine-tuning matrix as decoupled physical constraints to construct an autocorrelation attention model fused with physical constraints.

[0102] In the model training stage, the model parameters are optimized through back propagation and gradient update to obtain the trained autocorrelation attention model fused with physical constraints.

[0103] Further, in the model training stage, optimizing the model parameters through back propagation and gradient update includes: initializing the local fine-tuning matrix as a full zero matrix, in each iteration training, calculating the attention weight through forward propagation, and calculating the Gaussian negative log-likelihood loss by combining the real EDR value, calculating the gradient of the Gaussian negative log-likelihood loss on the full zero matrix through back propagation, and applying gradient update to the elements of the full zero matrix.

[0104] The model is trained using Gaussian negative log-likelihood as the loss function L:

[0105]

[0106] wherein, is the real EDR value. This loss function drives the model to optimize both the prediction accuracy and the uncertainty estimation.

[0107] In step S104 above, extracting the contribution matrix from the attention weights, quantifying the contribution ratio of each forecast time point to the prediction result, and outputting the probabilistic prediction result and contribution matrix of aircraft turbulence includes: grouping and summing the attention weights according to the forecast time point index to obtain the total attention score for each forecast time point; calculating the proportion of the total attention score for each forecast time point in the total attention score of all features to obtain the normalized contribution ratio; filling the normalized contribution ratio into the corresponding position of the contribution matrix to generate the contribution matrix; each element of the contribution matrix represents the relative contribution ratio of the corresponding forecast time point to the prediction result in the prediction time step; and outputting the contribution matrix and the probabilistic prediction result.

[0108] Specifically, extract the contribution matrix. :

[0109]

[0110] In the formula, This is the attention matrix optimized by physical constraints; indices j Includes all related to the first j The index of features related to each forecast time point in the input; Intuitively, this represents the prediction of the first... At the time step, the first j The relative contribution of each forecast time point.

[0111] The final attention weight matrix Grouping is performed based on the prediction time point index corresponding to the input features. The attention weights on all feature indices k belonging to the same prediction time point j are summed:

[0112]

[0113] In the above equation, λ1 and λ2 are adjustable hyperparameters used to control the weights of the global physics constraint term and the local fine-tuning term in the attention mechanism. The matrix, as an additive term, directly participates in the calculation of attention weights. Dividing this sum by the sum of attention weights across all feature indices yields the normalized contribution ratio. This contribution matrix visually demonstrates the weight of forecast information at different forecast time points (e.g., 3 hours or 12 hours in advance) when predicting the i-th time step, providing interpretability for the prediction results.

[0114] Example 2: Based on the above method, the second embodiment of the present invention proposes a method for predicting aircraft turbulence using a probabilistic autocorrelation attention model based on decoupled physical constraints. This method mainly includes data preprocessing, model building, model training, and prediction output.

[0115] 1. Regarding data preprocessing, this embodiment performs the process through the following three steps:

[0116] Data Collection and Alignment: Collect multi-time-series meteorological data for a specific flight route over the past year, including wind speed, wind shear, and temperature gradient, with a data interval of 1 hour. Simultaneously, collect EDR data for the corresponding flights on this route, also with a data interval of 1 hour. Accurately align the meteorological and EDR data according to timestamps and geographical locations to ensure that the meteorological data at each time step matches the corresponding EDR data.

[0117] Probabilistic feature construction: For historical EDR data, calculate the mean at each time step. and standard deviation For example, for time step t=1, the EDR data for the 10 adjacent time steps before and after this time step are collected, and their mean and standard deviation are calculated. For meteorological data, for each forecast time point... (e.g., 3 hours, 6 hours, 12 hours, 24 hours in advance), calculate the average of its meteorological data. and standard deviation This forms the input tensor;

[0118] The batch size is B=32, the historical sequence length is L=24 (i.e., using data from the past 24 hours), and the enhanced feature dimension is D′=16 (including the mean and standard deviation of EDR, as well as the mean and standard deviation of wind speed, wind shear, and temperature gradient at the four forecast time points).

[0119] Sparse data imputation: For partially missing EDR data, the following formula is used: Fill in the blanks.

[0120] in, Advanced statistical methods such as Gaussian process regression (GPR) or Kriging interpolation are used to interpolate based on known EDR data; F is based on the correction term of atmospheric physical models (such as Richardson number), and the physical estimate of turbulent energy dissipation rate is calculated based on wind shear and temperature gradient to correct the interpolation results.

[0121] 2. In terms of model construction, a global physical attenuation matrix is ​​constructed according to the formula. Calculate, where the time decay constant is... Based on the empirical value of the atmospheric eddy current time lifespan, set at 6 hours, the spatial decay constant is... Based on the empirical value of the spatial influence range of atmospheric eddies, it is set to 100 kilometers, and Δtᵢⱼ is the time difference (in hours) between time steps i and j.

[0122] Dlocijis the spherical distance between corresponding geographical locations (unit: km). The local fine-tuning matrix is initialized as a 24x24 zero matrix.

[0123] The auto-correlation attention is calculated as follows: .

[0124] where Q, K, V are the probabilistic input tensors Q, K, V are the probabilistic input tensors = 128, hyperparameters λ1=0.5, λ2=0.3.

[0125] The calculation of Auto-correlation Q, K is realized by FFT transformation. First, FFT transform is performed on Q and K respectively, and then the conjugate of the FFT result of K is multiplied with the FFT result of Q to obtain: .

[0126] In this embodiment, the Autoformer model is used as the model network, which includes a linear layer (input dimension 2, output dimension 16) and a ReLU activation function. The logarithm of the mean μ and the standard deviation σ of each feature log(σ) is taken as the input, and after linear transformation and activation function, the fused vector representation h t .

[0127] where Taking the logarithm operation can avoid numerical instability and ensure that the standard deviation is positive.

[0128] In the above embodiment, the model decoder includes two independent linear layers, each with an input dimension of 128 and an output dimension of 1, which are used to predict the mean and the standard deviation of the EDR, respectively, to form the output tensor , where the predicted sequence length O=12 (i.e., predicting the jolt situation in the next 12 hours).

[0129] 3. In terms of model training, the following steps are included:

[0130] The local fine-tuning matrix is initialized as a 24x24 zero matrix, and the other parameters of the model are initialized using the Xavier initialization method.

[0131] Forward propagation: In each training iteration, a batch (32 samples) of probabilistic input tensors The input model is input into the probabilistic data fusion module to obtain a fused feature vector, and then the fused feature vector is input into the calculation of attention weight , and finally the prediction result is output through the decoder.

[0132] Loss calculation: according to the predicted mean value , standard deviation and real EDR value output by the model, the loss is calculated by using the Gaussian negative log-likelihood loss function, and the formula is: .

[0133] Back propagation and gradient update: using Adam optimizer, learning rate η = 0.001, back propagation of loss function, calculation of loss gradient to all parameters of the model (including local fine-tuning matrix), and parameter update according to the gradient. The number of training iterations is set to 1000 rounds, and the validation set loss is calculated after each iteration. When the validation set loss does not decrease for 10 consecutive rounds, the training is stopped.

[0134] 4. In terms of prediction output, the preprocessed test data (batch size B = 32, history sequence length L = 24) is input into the trained model, and the model outputs the EDR prediction mean and standard deviation of the future 12 hours, such as “the future 1 hour roll index mean is 4.8, and the standard deviation is 0.3; the future 2 hour roll index mean is 5.1, and the standard deviation is 0.4” and the like. At the same time, the interpretable contribution matrix is generated:

[0135]

[0136] Among them, is the attention matrix optimized by physical constraints; contains the index of all features related to the jth prediction time point in the input; directly represents the relative contribution of the jth prediction time point when predicting the ith time step.

[0137] For example, “when predicting the future 1 hour roll, the 3 hour prediction weight is 65%, the 6 hour prediction weight is 20%, the 12 hour prediction weight is 10%, and the 24 hour prediction weight is 5%”, which provides a reference for flight decision-making.

[0138] Based on the same inventive concept, the application also provides a decoupled physical constraint-based aircraft jolt prediction system for implementing the above-mentioned decoupled physical constraint-based aircraft jolt prediction method. The implementation scheme for solving the problem provided by the system is similar to the implementation scheme described in the above-mentioned embodiment method, so the specific limitations in one or more decoupled physical constraint-based aircraft jolt prediction system embodiments provided below can refer to the limitations of the decoupled physical constraint-based aircraft jolt prediction method described above, and will not be repeated here.

[0139] In one embodiment, the application provides a decoupled physical constraint-based aircraft jolt prediction system. In practical application, the system can be deployed on a server of an air traffic control center, real-time receive meteorological data and aircraft EDR data, input the model for prediction after data preprocessing, and timely output the prediction result to provide guarantee for flight safety. As shown in Figure 2 The system includes an integration module 210, a data filling module 220, a model training module 230, and a result output module 240, wherein:

[0140] The integration module 210 is configured to align the pre-acquired multi-time series weather forecast data and aircraft EDR data on the same space-time grid, and convert the data into probabilistic features containing mean and standard deviation.

[0141] The data filling module 220 is configured to introduce an atmospheric physical model correction term based on the relationship between wind shear and turbulent energy dissipation rate to fill the sparse aircraft EDR data, and obtain a probabilistic input tensor.

[0142] The model training module 230 is configured to input the probabilistic input tensor into a pre-constructed and trained self-correlation attention model with fusion physical constraints.

[0143] The result output module 240 is configured to extract a contribution matrix from the attention weight of the trained model, quantify the contribution proportion of each prediction time point to the prediction result, and output the probabilistic prediction result of the aircraft jolt and the corresponding contribution matrix as the final prediction result.

[0144] Embodiment 4: Based on the technical principle of the above-mentioned decoupled physical constraint-based aircraft jolt prediction system, the application provides a decoupled physical constraint-based probabilistic self-correlation attention model in aircraft jolt prediction system, including a data preprocessing module, a model construction module, a model training module, and a prediction output module.

[0145] Among them, the data preprocessing module adopts Python programming language, and carries out data processing based on Pandas library. First, multi-time series meteorological data and aircraft EDR data are obtained through a data interface, and the data are aligned according to the time stamp and geographical position by using the merge function of Pandas. Then, the mean and standard deviation of each data are calculated using the numpy library to construct probabilistic features. For sparse or missing EDR data, the Gaussian process regression model in the Scikit-learn library is called to realize the function of G function, and F function is written based on atmospheric physics formula to realize data filling.

[0146] The model construction module is built based on the PyTorch deep learning framework. In the autocorrelation attention calculation module, the nn.Linear layer of PyTorch is used to realize the linear transformation of Q, K and V, and the FFT transformation and inverse transformation are realized by torch.fft module to calculate the autocorrelation. The global physical attenuation matrix is calculated and generated in Python according to the preset parameters and formula, and the local fine-tuning matrix is initialized by nn.Parameter and set requires_grad=True to realize training. The probabilistic data fusion module is composed of nn.Linear layer and nn.ReLU activation function, and the uncertainty quantification module is realized by two independent nn.Linear layers.

[0147] The model training module is also based on the PyTorch framework, and uses torch.optim.Adam optimizer to write training loop code according to the training steps in embodiment 1 to realize forward propagation, loss calculation, back propagation and parameter update. In the training process, TensorBoard is used to record the training loss and validation loss, which is convenient for monitoring the model training process.

[0148] The prediction output module converts the prediction results and weight matrix of the model into a text format easy to understand, builds a Web interface through the Flask framework, and outputs the results to the front-end page for pilots or relevant staff to view. At the same time, the results can be stored in the database for subsequent analysis and query.

[0149] In one embodiment, an electronic device, which can be a terminal, is provided, and an internal structure diagram of the electronic device can be as shown in FIG. 1. Figure 3As shown in the figure. The electronic device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. Among them, the processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium, an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The communication interface of the electronic device is used for wired or wireless communication with external terminals. Wireless mode can be achieved through WIFI, mobile cellular network, NFC (near field communication) or other technologies. The computer program is executed by the processor to implement any one of steps S101 to S104. The display screen of the electronic device can be a liquid crystal display screen or an electronic ink display screen. The input device of the electronic device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the electronic device, or an external keyboard, touchpad or mouse, etc.

[0150] Those skilled in the art can understand that Figure 3 The skilled in the art can understand that

[0151] Those skilled in the art can understand that

[0152] The embodiments of the present application can be provided as a method, system or computer program product. Therefore, the present application can be in the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code. Figure 1 Figure 1 The embodiments of the present application can be provided as a method, system or computer program product. Therefore, the present application can be in the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code. The embodiments of the present application can be provided as a method, system or computer program product. Therefore, the present application can be in the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0153] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flow Figure 1 The flow or flows and / or blocks Figure 1 The flow or flows and / or blocks

[0154] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions that execute on the computer or other programmable apparatus provide steps for implementing the flow Figure 1 The flow or flows and / or blocks Figure 1 The flow or flows and / or blocks

[0155] The above merely provides an embodiment of the present application, but is not intended to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the scope of the claims of the present application.

Claims

1. A method for predicting aircraft bump based on decoupled physical constraints, characterized in that, The method comprises: aligning the pre-collected multi-temporal weather forecast data and the aircraft EDR data on the same spatio-temporal grid, and converting into probabilistic features containing mean and standard deviation; for sparse aircraft EDR data, introducing an atmospheric physics model correction term based on the relationship between wind shear and turbulent energy dissipation rate to fill in, and obtaining a probabilistic input tensor; inputting the probabilistic input tensor into a pre-constructed and trained self-correlation attention model with physical constraints; wherein the self-correlation attention model with physical constraints comprises an encoder, an attention layer and a decoder; the encoder converts the probabilistic input tensor into one or more of a query matrix, a key matrix and a value matrix through linear mapping for self-correlation mapping; the attention layer calculates the cross-correlation function based on the query matrix and the key matrix, constructs an attention score matrix, and introduces a decoupled physical constraint, which includes a global physical attenuation matrix calculated according to the time difference and the spherical distance, and a trainable local fine-tuning matrix for capturing local weather bias; the decoder takes the physical constraint self-correlation feature sequence as input, and outputs the probabilistic prediction results of the mean and standard deviation of the EDR at multiple future time steps; extracting a contribution matrix from the attention weight of the trained model, quantifying the contribution proportion of each prediction time point to the prediction result, and outputting the probabilistic prediction result of the aircraft jolt and the corresponding contribution matrix as the final result of the prediction; the for sparse aircraft EDR data, introducing an atmospheric physics model correction term based on the relationship between wind shear and turbulent energy dissipation rate to fill in, and obtaining a probabilistic input tensor comprises: interpolating the probabilistic feature vector of the known points to obtain an interpolated mean and an interpolated standard deviation; filling in the probabilistic feature vector of the unknown points with missing EDR values based on the correction term of the atmospheric physics model; wherein the correction term of the atmospheric physics model includes the physical relationship between wind shear and turbulent energy dissipation rate; calculating the Richardson number according to the wind shear and the temperature gradient, and mapping the Richardson number to an estimated value of the turbulent energy dissipation rate; superimposing the estimated value of the turbulent energy dissipation rate on the interpolated mean, and obtaining the probabilistic input tensor based on the interpolated mean and the interpolated standard deviation of the superimposed turbulent energy dissipation rate estimate.

2. The method of claim 1, wherein, The aligning the pre-collected multi-temporal weather forecast data and the aircraft EDR data on the same spatio-temporal grid, and converting into probabilistic features containing mean and standard deviation comprises: converting the time stamp of the multi-temporal weather forecast data into UTC time; taking the recording time of the aircraft EDR data as the reference, matching the weather forecast data with a time difference less than a preset threshold for each EDR record point; using spherical linear interpolation on the matched weather forecast data, interpolating the spatial resolution of the weather forecast data to the latitude and longitude position of the EDR record point to obtain the aligned original data; and performing probabilistic feature representation on the aligned original data to obtain a probabilistic feature vector at each spatio-temporal grid point. The probability feature vector includes an EDR mean component, an EDR standard deviation component, a meteorological mean component, and a meteorological standard deviation component.

3. The method of claim 2, wherein, The probability feature representation of the aligned original data includes: obtaining the EDR mean component and the EDR standard deviation component by calculating the mean and standard deviation of the aircraft EDR data in a preset sliding time window; obtaining the meteorological mean component and the meteorological standard deviation component by calculating the mean and standard deviation of the meteorological variables at each forecast time point in the same sliding time window; The EDR mean component, the EDR standard deviation component, the meteorological mean component, and the meteorological standard deviation component are spliced to form a probability feature vector.

4. The method of claim 1, wherein, The training of the self-correlation attention model fused with physical constraints includes: Based on the time and space exponential decay law of atmospheric vortices, a global physical decay matrix is defined; A local fine-tuning matrix for capturing local meteorological bias is established; Taking Autoformer as the basic architecture, the global physical decay matrix and the local fine-tuning matrix are used as decoupled physical constraints to improve the self-correlation attention mechanism and build a self-correlation attention model fused with physical constraints; In the model training stage, the model parameters are optimized through back propagation and gradient update to obtain the trained self-correlation attention model fused with physical constraints.

5. The method of claim 4, wherein, The global physical decay matrix is defined based on the time and space exponential decay law of atmospheric vortices, which includes: Read the time difference and spherical distance between any two points in the space-time grid; Divide the time difference by a preset time decay constant and the spherical distance by a preset space decay constant, add them together and take the negative exponential to get the global physical decay matrix element.

6. The method of claim 4, wherein, In the model training stage, the model parameters are optimized through back propagation and gradient update, which includes: initializing the local fine-tuning matrix to a zero matrix, in each iteration, the attention weight is calculated by forward propagation, and the Gaussian negative log-likelihood loss is calculated by combining the real EDR value, the gradient of the Gaussian negative log-likelihood loss to the zero matrix is calculated by back propagation, and the gradient update is applied to the elements of the zero matrix.

7. The method of claim 6, wherein, The contribution matrix is extracted from the attention weight of the trained model, the contribution proportion of each forecast time point to the prediction result is quantified, and the probability prediction result of the aircraft buffet and the corresponding contribution matrix are output as the final prediction result, which includes: grouping and summing the attention weight according to the forecast time point index to get the total attention score of each forecast time point; calculate the proportion of the total attention score of each forecast time point in the total attention score of all features to get the normalized contribution proportion; fill the normalized contribution proportion into the corresponding position of the contribution matrix to generate the contribution matrix; each element of the contribution matrix represents the relative contribution proportion of the corresponding forecast time point to the prediction result in the prediction time step; the contribution matrix and the probability prediction result are output.

8. A decoupled physical constraint based aircraft bump prediction system, comprising: The system includes: The integration module is used to align the pre-acquired multi-time series meteorological forecast data and aircraft EDR data on the same space-time grid, and convert them into probability features containing mean and standard deviation; The data filling module is configured to introduce a correction term based on a relationship between wind shear and turbulent energy dissipation rate into sparse aircraft EDR data to obtain a probabilistic input tensor. The data filling module includes: interpolating the probabilistic feature vector of the known point to obtain an interpolation mean and an interpolation standard deviation; filling the probabilistic feature vector of the unknown point with missing EDR values based on a correction term of an atmospheric physical model, wherein the correction term of the atmospheric physical model includes a physical relationship between wind shear and turbulent energy dissipation rate; calculating a Richardson number based on wind shear and temperature gradient, and mapping the Richardson number to a turbulent energy dissipation rate estimate; and superimposing the turbulent energy dissipation rate estimate on the interpolation mean to obtain a probabilistic input tensor based on the interpolation mean and the interpolation standard deviation of the superimposed turbulent energy dissipation rate estimate; The model training module is configured to input the probabilistic input tensor into a pre-constructed and trained self-correlation attention model with physical constraints; wherein the self-correlation attention model with physical constraints includes an encoder, an attention layer, and a decoder. The encoder converts the probabilistic input tensor into one or more of a query matrix, a key matrix, and a value matrix through linear mapping for self-correlation mapping. The attention layer calculates a cross-correlation function based on the query matrix and the key matrix, constructs an attention score matrix, and introduces a decoupled physical constraint, wherein the decoupled physical constraint includes a global physical attenuation matrix calculated based on a time difference and a spherical distance, and a trainable local fine-tuning matrix for capturing local meteorological bias. The decoder takes the physical constraint self-correlation feature sequence as input and outputs a probabilistic prediction result including EDR mean and standard deviation at multiple future time steps. The result output module is configured to extract a contribution matrix from the attention weight of the trained model, quantify the contribution proportion of each prediction time point to the prediction result, and output the probabilistic prediction result of the aircraft turbulence and the corresponding contribution matrix as the final prediction result.

9. An electronic device, comprising: The electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to execute the method of any one of claims 1-7.

Images