An airplane simulation model confidence optimization method

Abstract

The present application relates to a kind of airplane simulation model confidence optimization method, for airplane simulation model confidence is not high and airplane online data transmission delay problem, proposed the method for improving simulation model confidence and solving data delay problem in combination with simulation system, airplane transmission data.The characteristic deficiency phenomenon of simulation output data in real data exists in offline network correction, online network can be combined with delayed transmission real data to the secondary correction of offline correction result, so the method involved in the present application combines two parts of offline correction network and online correction network.Random forest algorithm and MDA method are used to extract the features of offline network and online network.The simulation data error is preliminarily corrected by bidirectional long short-term memory neural network (BiLSTM) network and time attention mechanism in offline network;And the error of offline correction result and real data is secondarily corrected by online algorithm FTRL.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of aircraft simulation digital twin, and particularly relates to an aircraft simulation model confidence optimization method. BACKGROUND

[0002] With the rapid development of modern aviation technology, aircraft simulation models have been widely used in flight performance analysis, pilot training, aircraft simulation digital twin and aircraft design. However, the existing aircraft simulation models still have many problems in precision, complexity, parameter uncertainty and adaptability.

[0003] The current simulation model still has deficiencies in the precision of simulating flight states and dynamic characteristics, resulting in a large deviation between the simulation results and the actual flight data, which affects its reliability. High-precision simulation models are usually complex in structure and large in calculation, which is difficult to meet the demand of real-time simulation and rapid response, and increases the development and maintenance cost. A large number of parameters in the model are difficult to accurately measure or estimate, resulting in large parameter uncertainty, which further affects the confidence of the simulation results. The adaptability of the existing simulation model under different flight states and environmental conditions is poor, especially in adverse weather conditions, the prediction accuracy significantly decreases. In order to solve these problems, an aircraft simulation model confidence optimization method is proposed, which combines simulation model input data, simulation model output data and real data, adopts deep learning network BiLSTM and time attention mechanism to preliminarily correct errors. Then, combined with real-time real data and FTRL algorithm, the errors are corrected online, so that the optimized results are closer to the real data, thereby achieving the purpose of improving the confidence of the simulation model.

[0004] Therefore, the application is provided. SUMMARY

[0005] The application solves the technical problem of overcoming the shortcomings of the prior art and providing an aircraft simulation model confidence optimization method to solve the problems proposed in the background.

[0006] To solve the above technical problems, the basic idea of the technical solution of the application is:

[0007] An aircraft simulation model confidence optimization method comprises:

[0008] S1: Based on the simulation input feature evaluation of the simulation model and the MDA, the importance of the simulation input feature to the to-be-corrected simulation output feature is evaluated, and the high-importance simulation input feature is retained.

[0009] Taking advantage of the invariant parameters of the simulation model, the baseline accuracy of the simulation model MDA is set to 0. The simulation input features are shuffled sequentially, and the difference between the accuracy of the simulation output to be corrected and the accuracy of the shuffled simulation output to be corrected is calculated. The absolute value of the difference is taken and normalized. The normalized result is used as the importance score, and the simulation input features with high importance scores are saved as the high importance input feature set.

[0010] The calculation formula for each simulated input feature after shuffling is as follows:

[0011]

[0012] j is the j-th shuffled simulation model input feature sequence. Let y be the sample after random swapping of the j-th feature, n be the number of input samples, and y be the number of input samples. i To avoid disrupting the values ​​output by the previous simulation model, The output of the simulation model is to shuffle the j-th feature.

[0013] S2: Simulation output feature evaluation based on random forest MDA. The simulation output features are used to calculate the importance of the simulation output features to be corrected and retain the high-importance output features.

[0014] The importance of the simulation output features to be corrected is calculated. A random forest model is trained using the output features of the complete simulation model, and the baseline accuracy of the model is calculated. For each output feature, the value of the feature in the dataset is shuffled, and the accuracy of the model on the shuffled dataset is calculated. The absolute value of the difference between the baseline accuracy and the shuffled accuracy is taken, and the result after normalization of the absolute value is used as the importance score. Based on the score, the set of high-importance output features is evaluated.

[0015] Features with higher importance scores are saved as the high-importance output feature set, calculated using the following formula:

[0016]

[0017] In the formula, the absolute value of MDA'(j) is taken as the importance assessment score, and T is the number of trees in the random forest (X). i y i ) as a sample, For X i The sample after random swapping in the j-th dimension, D t For the out-of-bag sample set of random tree t, Let R(X) be the sample set formed after the j-th dimension is exchanged. i ) is the sample X i The predicted output.

[0018] S3: integrate the high-importance input features, high-importance output features, and to-be-corrected simulation output features into a simulation data reinforcement feature set, take the simulation data reinforcement feature set as the input of the bidirectional long short-term offline correction network of the time attention mechanism, and take the difference between the real data and the to-be-corrected simulation output data as the label for training and testing;

[0019] The simulation data reinforcement feature set is divided into a training set and a test set with a ratio of 0.8:0.2, the difference between the real data and the simulation output data is taken as the label, and the error of offline correction is added to the simulation model output as the result of offline correction.

[0020] S4: train using the BiLSTM network and the time attention mechanism algorithm, save the trained BiLSTM network, preliminarily correct the error using the trained offline network, and take the corrected error and the to-be-corrected simulation result as the final result of offline correction;

[0021] The BiLSTM network and the time attention mechanism algorithm are used to correct the error between the real value and the to-be-corrected simulation output value, so that the result after adding the simulation output to the corrected error is close to the value at the real time of the aircraft. BiLSTM is composed of two independent LSTMs, and LSTM unit is composed of input gate, forget gate, output gate and cell state. The main formula of LSTM unit calculation is as follows:

[0022] Input gate: i t =σ(W xi x t +W hi h t-1 +W ci c t-1 +b i );

[0023] Forget gate: f t =σ(W xf x t +W hf h t-1 +W cf c t-1 +b f );

[0024] Cell state:

[0025] Cell state update:

[0026] Output gate: o t =σ(W xo x t +W ho h t-1 +W co ct +b o );

[0027] Hidden state update: h t =o t ·tanh(c t );

[0028] In the formula, x t h is the input for time step t. t-1 It is the hidden state of the previous time step, c t-1 It is the cell state at the previous time step, i t For the input gate, f t For the Gate of Oblivion In cellular state, c t Update the state, where W and b are weight parameters.

[0029] In BiLSTM, the input data at each time step is processed by two independent LSTM processors, one from front to back and the other from back to front. The outputs of the forward and backward LSTM layers are concatenated to form a combined output vector. The temporal attention mechanism calculates the attention score for each time step on the output of the bidirectional LSTM and generates a context vector. This context vector is combined with the output of the bidirectional LSTM and passed to the fully connected layer to generate the final corrected output. The attention calculation steps are as follows:

[0030] The forward LSTM processes the sequence from beginning to end, and the backward LSTM processes the sequence from end to beginning. The final BiLSTM output is a concatenation of the forward and backward states, as shown in the following formula:

[0031]

[0032] Calculate the attention score: The attention score at each time step is calculated using a linear transformation, as shown in the following formula:

[0033] e t =W att ·H t ;

[0034] Normalized attention score: The score is normalized to weights using the softmax function, as shown in the following formula:

[0035]

[0036] Calculate the context vector: The BiLSTM output is weighted and summed using attention weights to obtain the context vector, as shown in the following formula:

[0037]

[0038] The context vector c is output by a fully connected layer, and the formula is as follows:

[0039] output=W fc ·c+b fc .

[0040] Dropout is set in each LSTM hidden layer, and the activation value of a certain neuron is stopped working with a certain probability p, so that the model has better generalization. The BiLSTM network uses the MSE loss function and the Adam optimizer. In the forward propagation process of the BiLSTM network, the input data passes through the BiLSTM and the time attention mechanism to generate the value The BiLSTM network uses the MSE loss function to calculate the correction error and the loss between the real error y, and the formula is as follows:

[0041]

[0042] The loss L is calculated by back propagation to the gradient of the model parameter The Adam optimizer is used to update the model parameter θ according to the gradient , and the calculation formula is as follows:

[0043] m t =β1m t-1 +(1-β1)g t ;

[0044]

[0045] In the formula, L is the loss value, which represents the loss value between the correction error value and the real error value, N is the sample number, which is the total number of samples in the data set, y i is the real error value of the i-th sample, is the model correction error value of the i-th sample, m t is the first-order momentum estimate of time step t, which represents the exponentially weighted moving average of the gradient, v t is the second-order momentum estimate of time step t, which represents the exponentially weighted moving average of the gradient square, β1 is the first-order momentum decay factor, β2 is the second-order momentum decay factor, g t is the gradient of time step t, which represents the gradient of the loss function to the parameter, θ t is the model parameter of time step t, θ t+1 is the model parameter of time step t+1, α is the learning rate, which controls the step size of parameter update, and ∈ is a small constant to prevent the denominator from being zero.

[0046] S5: Evaluate the real features that are important to the corrected simulation output features in the real data set, and retain the high importance real feature set;

[0047] S6: Online network training and correction based on FTRL algorithm, reconstitute time series of high importance real feature set and offline output final result, take time series as online network FTRL training input, difference between real feature data corresponding to to-be-corrected feature and offline output final result as online training label, error of online training plus offline final result as network optimization final result. High importance real feature set evaluated is input to algorithm FTRL for correction, difference between real feature data and offline output final result as online training label, set time sliding window with size of 10 to perform online training and correction, correction error of online training plus offline correction result as entire network optimization final result;

[0048] The FTRL algorithm adopts different update modes for weights of online input features, and the weight of the i-th feature is w i The gradient vector parameter is g i The weight update steps are as follows:

[0049] The cumulative gradient update formula is as follows:

[0050]

[0051] The cumulative gradient square sum n i The update formula is as follows:

[0052]

[0053] For the weight w i The update formula is as follows:

[0054]

[0055] In the formula, alpha and beta are hyperparameters, lambda 1>0 and lambda 2>0 are L1 and L2 regularization coefficients respectively. z i is the cumulative gradient of feature i, g i is the gradient of feature i in the previous step, n i is the cumulative gradient square sum of feature i, w i is the model weight corresponding to feature i.

[0056] After the above technical solution is adopted, the present application has the following beneficial effects compared with the prior art, of course, any product implementing the present application does not necessarily need to achieve all the advantages described below:

[0057] 1. The difference between the actual value and the simulation model output value is initially optimized by using an offline BiLSTM network. Then, the error between the actual data and the offline network optimization result is further optimized by combining real online data. The optimized error is added to the final result of the offline network optimization as the final optimized output, which improves the confidence of the simulation system and makes the simulation output closer to the real data.

[0058] 2. To solve the problem of data transmission delay in the "cloud box" of future aircraft development, and to ensure the authenticity and timeliness of simulation data.

[0059] 3. The importance assessment, combined with the characteristics of the simulation model and simulation input data, significantly improves the accuracy of feature selection.

[0060] 4. By employing BiLSTM and temporal attention mechanisms, the model's ability to process time series data is enhanced, effectively capturing temporal correlations in the data.

[0061] 5. By using Dropout and Adam optimization algorithms, the training effect and stability of the network are improved, and the risk of overfitting is reduced.

[0062] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0063] The accompanying drawings described below are merely some embodiments. Those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0064] In the picture:

[0065] Figure 1 This is a schematic diagram of the method involved.

[0066] It should be noted that these accompanying drawings and textual descriptions are not intended to limit the scope of the invention in any way, but rather to illustrate the concept of the invention to those skilled in the art by referring to specific embodiments. Detailed Implementation

[0067] The invention will now be described in further detail with reference to the accompanying drawings.

[0068] Please see Figure 1 As shown, this implementation provides a method for optimizing the confidence of an aircraft simulation model in the correction of the pitch angle output, including:

[0069] The system acquires input data (such as flight control parameters, environmental conditions, etc.) and output data (such as flight attitude, speed, position, etc.) from the aircraft simulation model; it also acquires corresponding output data from actual flight records, and performs data cleaning and normalization on the collected simulation data and real flight data.

[0070] S1: Based on the simulation model and MDA simulation input feature evaluation, evaluate the importance of simulation model input features to the corrected simulation output pitch angle, and retain high importance simulation input features (co-pilot side rudder force sensor signal, center of gravity estimation, left inboard elevator actuator position, right outboard elevator position after voting, right aileron position after voting, crosswind calculation, co-pilot side rudder force sensor signal, center of gravity position at takeoff, left outboard elevator actuator position, flaps, total fuel consumption in flight, turbulence, co-pilot side rudder force sensor signal, engine thrust rating, left inboard elevator position after voting, cockpit-flap configuration position, total fuel quantity, static air temperature, co-pilot side rudder force sensor signal, right inboard elevator actuator position, flap position (corrected), wind speed, engine throttle lever angle (corrected position), center of gravity position, elevator feel position (right RVDT) signal, right outboard elevator actuator position (LVDT) signal, left aileron position after voting, wind direction);

[0071] Based on the characteristics of the simulation model parameters, the simulation model MDA baseline accuracy value is set to 0, the simulation input features are randomly disturbed, the difference between the simulation output pitch angle and the disturbed pitch angle is calculated, the absolute value of the difference is taken, the absolute value is normalized, and the normalized result is taken as the importance score.

[0072] The formula for disturbing each simulation input feature is as follows:

[0073]

[0074] j is the jth disturbed simulation model input feature sequence, is the sample after random exchange of the jth feature, n is the number of input samples, y i is the value of the simulation model output before disturbance, is the output of the simulation model after disturbing the jth feature.

[0075] S2: Based on the random forest MDA simulation output feature evaluation, calculate the importance of simulation output features to the corrected simulation output pitch angle through simulation model output features, and retain high importance output features (rate of change of altitude data, fuel-oil heat exchanger pressure difference, right inertial reference unit hybrid GPS altitude, left inertial reference unit hybrid GPS altitude, calculated airspeed from integrated standby flight display (ISFD));

[0076] The importance of the output features of the calculation simulation is evaluated on the pitch angle of the simulation output to be corrected; a random forest model is trained using the output features of the complete simulation model, and the baseline accuracy of the model is calculated. For each output feature, the values of the feature in the data set are shuffled, and the accuracy of the model on the shuffled data set is calculated. The absolute value of the baseline accuracy minus the accuracy after shuffling is taken, and the normalized result is used as the importance score. According to the score, a set of high-importance output features is evaluated.

[0077] The features with high importance scores are saved as the set of high-importance output features, and the calculation formula is:

[0078]

[0079] In the formula, the absolute value of MDA'(j) is taken as the importance evaluation score, T is the number of random forest trees (X i , y i ) is the sample, is the j-th dimensionally shuffled sample of X i , D t is the out-of-bag sample set of random tree t, is the sample set R(X i ) formed after the j-th dimension is exchanged, and i is the predicted output of sample X i .

[0080] S3: The high-importance input features, high-importance output features, and simulation output pitch angle to be corrected are integrated into a simulation data reinforcement feature set, which is used as the input of the bidirectional long short-term offline correction network of the time attention mechanism. The difference between the real pitch angle data and the simulation output pitch angle data to be corrected is used as the label for training and testing.

[0081] The simulation data reinforcement feature set is divided into a training set and a test set with a ratio of 0.8:0.2. The difference between the real pitch angle and the simulation output pitch angle is used as the label, and the error of the offline correction is added to the simulation model output pitch angle to obtain the offline corrected pitch angle.

[0082] S4: The BiLSTM network and the time attention mechanism algorithm are used for training and preliminary correction of errors. The simulation data reinforcement feature set is used as the training data input, and the difference between the real pitch angle and the simulation output pitch angle to be corrected is used as the label.

[0083] The BiLSTM network and the time attention mechanism algorithm are used to correct the error between the real pitch angle and the simulation output pitch angle, so that the result after adding the simulation output to the corrected error is close to the real-time pitch angle value of the aircraft. BiLSTM is composed of two independent LSTMs, and LSTM unit is composed of input gate, forget gate, output gate and cell state. The main formula for calculating the LSTM unit is as follows:

[0084] Input gate: i t = σ(W xi x t +W hi h t-1 +W ci c t-1 +b i ) ;

[0085] Forget gate: f t = σ(W xf x t +W hf h t-1 +W cf c t-1 +b f ) ;

[0086] Cell state:

[0087] Cell state update:

[0088] Output gate: o t = σ(W xo x t +W ho h t-1 +W co c t +b o ) ;

[0089] Hidden state update: h t = o t · tanh(c t ) ;

[0090] where x t is the input at time step t, h t-1 is the hidden state at the previous time step, c t-1 is the cell state at the previous time step, i t is the input gate, f t is the forget gate, is the cell state, c t is the updated state, W, b are weight parameters.

[0091] In BiLSTM, the input data of each time step is processed by two independent LSTM processors, one from front to back and the other from back to front, the output of the forward LSTM layer and the backward LSTM layer is connected to form a comprehensive output vector, the time attention mechanism calculates the attention score of each time step on the output of the bidirectional LSTM, and generates a context vector, which is combined with the output of the bidirectional LSTM and passed to the fully connected layer to generate the final correction output, the attention calculation steps are as follows:

[0092] The forward LSTM processes the sequence from beginning to end, and the backward LSTM processes the sequence from end to beginning. The final BiLSTM output is the connection of the forward and backward hidden states, and the formula is as follows:

[0093]

[0094] Calculate attention score: Calculate the attention score of each time step by linear transformation, the formula is as follows:

[0095] e t =W att ·H t ;

[0096] Normalize attention score: Use the softmax function to normalize the score to weight, the formula is as follows:

[0097]

[0098] Calculate context vector: Weighted sum of BiLSTM output with attention weight to get context vector, the formula is as follows:

[0099]

[0100] Context vector c is modified by a fully connected layer to get the final output result, the formula is as follows:

[0101] output=W fc ·c+b fc .

[0102] Set Dropout in each LSTM hidden layer, let the activation value of a certain neuron stop working with a certain probability p, so that the model has better generalization ability. BiLSTM network uses MSE loss function and Adam optimizer. In the forward propagation process of BiLSTM network, the input data passes through BiLSTM and time attention mechanism to generate value BiLSTM network uses MSE loss function to calculate the loss between correction error and real error y in the training process, the formula is as follows:

[0103]

[0104] Compute the gradient of the loss L with respect to the model parameters Using the Adam optimizer, update the model parameters θ according to the gradient The update rule for the model parameters θ is given by:

[0105] m t = β1m t-1 + (1 - β1)g t ;

[0106]

[0107] where L is the loss value, N is the number of samples, yi is the true error value for the i-th sample, and mi is the model corrected error value for the i-th sample. i t is the first-order momentum estimate at time step t, which represents the exponentially weighted moving average of the gradients, and v t is the second-order momentum estimate at time step t, which represents the exponentially weighted moving average of the squared gradients, β1 is the first-order momentum decay factor, β2 is the second-order momentum decay factor, g t is the gradient at time step t, which represents the gradient of the loss function with respect to the parameters, and θ t is the model parameters at time step t, θ t+1 is the model parameters at time step t+1, α is the learning rate, which controls the step size of the parameter updates, and ∈ is a small constant to prevent division by zero.

[0108] S5: Evaluate the importance of the features in the real dataset for the corrected simulated output pitch angle, and retain the high-importance real feature set (rate of change of altitude data, fuel-oil heat exchanger pressure difference, right inertial reference unit hybrid GPS altitude, left inertial reference unit hybrid GPS altitude, calculated airspeed from integrated standby flight display (ISFD));

[0109] S6: Online network training and correction based on the FTRL algorithm, reconstitute the time series of the high-importance real feature set and the offline output pitch angle, input the time series into the online network FTRL training, the difference between the real pitch angle and the final offline network corrected pitch angle as the online training label, the error of online training plus the final result of the offline network as the network optimization final pitch angle;

[0110] The high-importance real feature set evaluated is input to the online algorithm FTRL for correction, a time sliding window of size 10 is set for online training and correction, and the correction error of online training plus the offline network pitch angle as the entire network optimization pitch angle result; ​​

[0111] The FTRL algorithm adopts different update methods for the weights of online input features, and the weight of the i-th dimension feature is w i The gradient vector parameter is g i The weight update step is as follows:

[0112] The cumulative gradient update formula is as follows:

[0113] The cumulative gradient update formula is as follows:

[0114]

[0115] The cumulative gradient square sum n i The update formula is as follows:

[0116]

[0117] The weight w i The update formula is as follows:

[0118]

[0119] In the formula, alpha and beta are hyperparameters, lambda1>0 and lambda2>0 are L1 and L2 regularization coefficients respectively. i The cumulative gradient of feature i is z i The gradient of feature i in the previous step is g i The cumulative gradient square sum of feature i is n i The model weight corresponding to feature i is w.

[0120] The difference between the real pitch angle and the simulation model output pitch angle value is preliminarily optimized by the offline network BiLSTM, and then the error between the real online data and the offline network optimization result is secondarily optimized. The optimized error is added to the final result of the offline network optimization to improve the confidence of the simulation system, so that the simulation output is closer to the real data. The "cloud box" data transmission delay problem in future aircraft research and development is solved, and the authenticity and timeliness of the simulation data are improved. The importance evaluation combined with the simulation model and the simulation input data characteristics significantly improves the feature selection accuracy. The BiLSTM and time attention mechanism are used to enhance the processing capability of the model for time series data, effectively capturing the time sequence correlation in the data. Through Dropout and Adam optimization algorithm, the training effect and stability of the network are improved, and the risk of overfitting is reduced.

[0121] The present application is not limited to the above-mentioned embodiments, and any person should know that the structural changes made under the inspiration of the present application fall within the protection scope of the present application. The technical, shape and structure parts not described in detail in the present application are well-known technologies.

Claims

1. An aircraft mockup confidence optimization method, characterized in that, Comprise: S1: Based on the simulation model and MDA simulation input feature evaluation, evaluate the importance of simulation model input features to the simulation output features to be corrected, and retain high importance simulation input features; S2: Based on the simulation output feature evaluation of the random forest MDA, the simulation output feature is calculated, and the importance of the simulation output feature to the simulation output feature to be corrected is calculated, and the high importance output feature is retained; S3: The high importance input feature, the high importance output feature and the simulation data reinforcement feature set to be corrected are integrated into the simulation data reinforcement feature set, which is used as the input of the time attention mechanism of the bidirectional long short term offline correction network, and the difference between the real data and the simulation output data to be corrected is used as the label for training and testing; S4: Training is carried out by using BiLSTM network and time attention mechanism algorithm, and the trained BiLSTM network is saved, the error of the offline network is corrected, and the corrected error is added to the simulation result to be corrected as the final result of offline correction; S5: Evaluate the real feature data in the real data set which is important to the simulation output feature to be corrected, and retain the high importance real feature set; S6: Online network training and correction based on FTRL algorithm, the high importance real feature set and offline output final result are reconstituted into time series, the time series is used as the input of online network FTRL training, the difference between the real feature data corresponding to the feature to be corrected and the offline output final result is used as the online training label, and the error of online training is added to the offline final result as the final result of network optimization.

2. The method of claim 1, wherein, In step S1, the simulation model parameter is unchanged, the simulation model MDA baseline accuracy value is set to 0, the simulation input features are disturbed in turn, the accuracy of the simulation output to be corrected is calculated, and the difference between the accuracy of the simulation output to be corrected after disturbance is taken, the absolute value of the difference is taken, the normalized result is taken as the importance score, and the simulation input features with high importance score are saved as the high importance input feature set.

3. The method of claim 2, wherein, The shuffling calculation formula for each simulation input feature is as follows: ; j is the jth shuffled simulation model input feature sequence, is the jth feature shuffled sample, n is the number of input samples, y i is the value of the simulation model output before shuffling, is the output of the simulation model after shuffling the jth feature.

4. The method of claim 1, wherein, In step S2, the importance of the simulation output feature to the simulation output feature to be corrected is calculated by using the simulation model output feature, the random forest model is trained by using the complete simulation model output feature, and the baseline accuracy of the model is calculated. For each output feature, disturb the value of the feature in the data set, calculate the accuracy of the model on the disturbed data set, take the absolute value of the baseline accuracy minus the accuracy after disturbance, normalize the result after taking the absolute value, and take the result as the importance score. According to the score, the high importance output feature set is evaluated.

5. The method of claim 4, wherein, The features with higher importance scores are saved as a high-importance output feature set, and the calculation formula is: ; where the importance score is taken as the absolute value T is the number of random forest trees is the sample, is the sample after random permutation of the jth dimension, is the out-of-bag sample set for random tree t, is the sample set formed after permutation of the jth dimension R(Xi) is the prediction output of the sample X i .

6. The method of claim 1, wherein, In step S3, the simulation data reinforcement feature set is used as the input of the BiLSTM network, the difference between the real data and the simulation output data is used as the label, and the error of offline correction is added to the simulation model output data as the final result of offline correction.

7. The method of claim 1, wherein, In step S4, the BiLSTM network and the time attention mechanism algorithm are used to correct the error between the true value and the simulation output value, so that the result after the simulation output is corrected is close to the true value of the aircraft. The BiLSTM is composed of two independent LSTMs, and the LSTM unit is composed of an input gate, a forget gate, an output gate and a cell state. The main formula for calculating the LSTM unit is as follows: Input gate: ; forget gate: ; Cell state: ; Cell status update: ; Output gate: ; Hidden state update: ; where X t is the input for time step t, is the hidden state of the previous time step, is the cell state of the previous time step, i t is the input gate, f t is the forget gate, is the cell state, c t is the update state, , are the weight and bias parameters of the relevant layer in the LSTM network.

8. The method of claim 7, wherein, In the BiLSTM, the input data of each time step is processed by two independent LSTM processors, one from front to back and the other from back to front. The outputs of the forward LSTM layer and the reverse LSTM layer are connected to form a comprehensive output vector. The time attention mechanism calculates the attention score of each time step on the output of the BiLSTM and generates a context vector. The context vector is combined with the output of the BiLSTM and then passed to the fully connected layer to generate the final correction output. The attention calculation steps are as follows: The forward LSTM processes the sequence from the beginning to the end, and the backward LSTM processes the sequence from the end to the beginning. The final BiLSTM output is the concatenation of the forward and backward hidden states, as follows: ; Compute attention scores: Compute the attention scores for each time step by a linear transformation, formula as follows: ; Normalized attention scores: The scores are normalized to weights using a softmax function, as follows: ; Compute context vector: Weighted sum of BiLSTM outputs with attention weights to get context vector, formula as follows: ; The context vector is passed to a fully connected layer after being combined with the output of the bi-directional LSTM, generating the final correction output, as follows: .

9. The method of claim 8, wherein, Dropout is set in each LSTM hidden layer, which makes the activation value of a neuron stop working with a certain probability p, so that the model has better generalization. The BiLSTM network adopts the MSE loss function and the Adam optimizer. In the forward propagation process of the BiLSTM network, the input data passes through the BiLSTM and the time attention mechanism to generate values ; The BiLSTM network uses an MSE loss function to calculate the correction error during training and the real error y, the formula is as follows: ; The loss L is computed by backpropagation to calculate the gradient of the model parameters Using the Adam optimizer, the gradient Update the model parameters θ, the formula is as follows: ； ； ； ； ； where L is the loss value representing the loss value between the corrected error value and the true error value, N is the number of samples, the total number of samples in the data set, y i is the true error value of the i-th sample, is the model corrected error value of the i-th sample, m t is the first-order momentum estimate of time step t, representing the exponentially weighted moving average of the gradient, v t is the second-order momentum estimate of time step t, representing the exponentially weighted moving average of the gradient square, β1 is the first-order momentum decay factor, β2 is the second-order momentum decay factor, g t is the gradient of time step t, representing the gradient of the loss function with respect to the parameters, θ t is the model parameter of time step t, θ t+1 is the model parameter of time step t+1, α is the learning rate, controlling the step size of parameter update, ∈ is a small constant to prevent the denominator from being zero.

10. The method of claim 1, wherein, In step S6, the set of real features with high importance evaluated in step S5 is input to the algorithm FTRL for correction. The difference between the real feature data and the final result of the offline output is used as the online training label. A time sliding window with a size of 10 is set for online training and correction. The correction error of the online training is added to the offline final result to obtain the final result of the entire network optimization. The FTRL algorithm adopts different update methods for the weights of online input features, and the weight of the i-th dimension feature is w i The gradient vector parameter is g i The weight update step is as follows: The accumulated gradient update formula is as follows: ; cumulative gradient squared sum n i The update formula for n is as follows: ； For the weight w of the feature i The update formula is as follows: ； wherein, and are hyperparameters, are L1, L2 regularization coefficients, respectively, Z i is the cumulative gradient of feature i, g i is the gradient of the previous step feature i, n i is the cumulative sum of squared gradients of feature i, w i is the model weight corresponding to feature i.

Images