A VAE-based intelligent verification method for aircraft guidance and control model, electronic equipment and storage medium

Abstract

A method, electronic device and storage medium for intelligent verification of aircraft guidance and control models based on VAE, belonging to the technical field of intelligent verification of simulation models. In order to improve the accuracy and efficiency of simulation model verification, the present invention includes obtaining multi-sample reference data sets and multi-sample simulation data sets through repeated experiments based on aircraft guidance and control systems and aircraft guidance and control simulation models and performing normalization processing; constructing a VAE model and initializing the model structure and parameters; using the normalized reference data set as input, training and optimizing the VAE model; extracting features from the normalized reference data set and simulation data set using the optimized VAE model to obtain reference data features and simulation data features; measuring the difference between reference data features and simulation data features from the two aspects of numerical value size and probability distribution, and then obtaining the credibility of the aircraft guidance and control model. The present invention improves the efficiency of model verification.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent verification of simulation models, and specifically relates to an intelligent verification method, electronic equipment and storage medium for an aircraft guidance and control model based on VAE. Background Art

[0002] As an important means to assist in the design, development, evaluation and optimization of complex systems, simulation technology has been widely used in many fields such as military, manufacturing, biology, medicine and transportation. Simulation model verification is an important means to analyze and evaluate the credibility of simulation models, which is conducive to verifying simulation results, discovering simulation problems and reducing simulation risks. With the widespread application of simulation technology and the rapid development of related technologies, we have entered a new era dominated by complex simulation, which has brought many new challenges to simulation model verification.

[0003] The simulation process of complex simulation models such as aircraft guidance and control simulation models is long and uncertain, resulting in large sample sizes and long time series in their output data. Existing simulation evaluation technologies cannot meet such evaluation requirements well. Simulation result verification is the most direct and effective model verification method. Its essence is to evaluate the consistency between the simulation model and the simulated object based on the principle of output result difference. At present, for aircraft guidance and control simulation models, the most commonly used dynamic simulation result verification method in engineering is to directly measure the difference between simulation data and reference data based on the original data, such as the verification method based on the Grey Relational Analysis method (GRA) and the Theil Inequality Coefficient method (TIC). In addition, comparing the feature differences between simulation data and reference data is also a common method in dynamic simulation result verification. Most of the existing feature extraction methods used in the verification process of aircraft guidance and control simulation models artificially select or design features according to different tasks and data, such as selecting position, shape, power spectrum density, etc. as the features of dynamic data.

[0004] The dynamic output verification method based on raw data can only find the differences on the surface of the data, without touching the intrinsic characteristics of the data itself and quantifying and comparing the uncertainty of the data. In addition, it has high complexity when applied to verify large-scale data. The existing feature-based verification methods often use manual feature extraction, which not only relies heavily on experience and luck, but is also very time-consuming and laborious when the amount of data is large. In summary, the existing verification methods have problems such as low processing efficiency, difficulty in selecting data features, and high workload for evaluators when used to process large-scale dynamic output data. Therefore, it is necessary to study a simulation result verification method based on intelligent feature extraction for the large-scale dynamic output of the aircraft guidance and control simulation model, measure the consistency between the simulation data and the reference data from the two aspects of data essential characteristics and uncertainty, and improve the accuracy and efficiency of simulation model verification. Summary of the invention

[0005] The problem to be solved by the present invention is to improve the accuracy and efficiency of simulation model verification, and propose an aircraft guidance and control model intelligent verification method, electronic equipment and storage medium based on VAE.

[0006] To achieve the above object, the present invention is implemented through the following technical solutions:

[0007] A VAE-based intelligent verification method for an aircraft guidance and control model comprises the following steps:

[0008] S1. Repeat the test to obtain a multi-sample reference data set and a multi-sample simulation data set based on the aircraft guidance control system and the aircraft guidance control simulation model, and normalize the data to obtain a normalized reference data set and a normalized simulation data set for use;

[0009] S2. Build a VAE model and initialize the model structure and parameters for later use;

[0010] S3. Using the normalized reference data set obtained in step S1 to train the VAE model constructed in step S2, to obtain an optimized VAE model;

[0011] S4. The normalized reference data set and the normalized simulation data set obtained in step S1 are respectively input into the optimized VAE model for feature extraction to obtain a reference data feature matrix and a simulation data feature matrix;

[0012] S5. Using the reference data feature matrix and the simulation data feature matrix obtained in step S4, measure the difference between the reference data features and the simulation data features in terms of numerical value and probability distribution, and then obtain the credibility of the aircraft guidance and control model.

[0013] Furthermore, in step S1, S is setr and S s Represent the aircraft guidance control system and the aircraft guidance control simulation model respectively. Let S r and S s Run N times under given input conditions r times and N s Each time, the output data of length T is obtained at the same sampling interval, and the multi-sample reference data set Y is obtained. r and multi-sample simulation dataset Y s as follows:

[0014]

[0015] in, represents the reference time series obtained from the nth single operation of the aircraft guidance and control system; represents the simulation time series obtained from the nth single run of the aircraft guidance and control simulation model;

[0016] The data in the collected multi-sample reference data set and the multi-sample simulation data set are normalized to obtain normalized reference data and normalized simulation data, and the expression is:

[0017]

[0018] in, Y r The nth r elements, n r =1,2,....,N r T, Y s The nth s elements, n s =1,2,....,N s T, and They are and The corresponding normalized result, {y n} indicates that Y r and Y s The set of all elements in contains (N r +N s )·T elements, min{y n} represents the set {y n}; max{y n} represents the set {y n} is the largest element in .

[0019] Furthermore, the specific implementation method of step S2 includes the following steps:

[0020] S2.1. Build the VAE model, including the encoder and decoder p θ (Y|Z) two parts;

[0021] S2.2. The encoder receives the input data Y, passes through a hidden layer consisting of a fully connected layer FC1 and a ReLU activation function, and then outputs h1. Then, it passes through two parallel fully connected layers FC2 and FC3, outputting the mean μ and variance σ respectively, and then passes through a reparameterization layer to obtain the potential representation Z, which is expressed as:

[0022] Z=μ+ε×σ

[0023] Among them, ε is obtained by multiple sampling of the standard Gaussian distribution N(0,1);

[0024] S2.3. The decoder receives the latent representation Z, passes through a hidden layer consisting of a fully connected layer FC4 and a ReLU activation function, and then outputs h2, which is then passed through a fully connected layer FC5 and a Sigmoid activation function, and finally outputs the reconstructed data Y′;

[0025] The potential representation is a vector in a latent space mapped from an input data sample by an encoder, and the dimension of the latent space is set according to the complexity of the data. For low-dimensional data, the dimension of the latent space is set to 2-20 dimensions; for high-dimensional data, the dimension of the latent space is set to 20-100 dimensions;

[0026] The hidden layer is used to capture important features of the input data. Its dimension and number of layers are determined according to the length and complexity of the time series. If the input data dimension is less than 1000 dimensions, the dimension of the hidden layer is set to 2 to 10 times the input dimension; if the input data dimension is greater than 1000 dimensions, the dimension of the hidden layer is set to 1 / 2, 1 / 4 or 1 / 10 of the input dimension; if the input data dimension is greater than 1000 dimensions and has complex dynamic characteristics, 1-2 hidden layers are added to the encoder and decoder respectively.

[0027] Furthermore, the normalized reference data set obtained in step S1 is input into the VAE model constructed in step S2 for model training. The specific implementation method is to first map the input data to the latent space through the encoder, then reconstruct the data through the decoder, and then calculate the training loss. If the termination condition is not met, the parameters of the VAE model are updated through back propagation.

[0028] Setting the loss function of the VAE model The expression is:

[0029]

[0030] in, and θ are the parameters of the encoder and decoder respectively; Y is the input reference data sample set; Z is the potential representation corresponding to Y; is the original input data Y (i) The expected value of the mean square error between and the reconstructed data Y′; is the approximate posterior distribution of the latent representation Z learned by the encoder With the prior distribution p θ The Kullback-Leibler (KL) divergence between (Z) is expressed as:

[0031]

[0032] Where D is the dimension of the latent space; μ d and are the mean and variance of the d-th dimension potential representation, respectively;

[0033] Repeatedly train the VAE model to optimize the model's parameters and structure. During the training process, monitor the changes in the loss value and analyze the latent space distribution parameters after training, adjust the model's structure and hyperparameters. The model's structure includes the number of hidden layers, adjusts the hidden layer dimension and the latent space dimension, and the hyperparameters include the learning rate and batch size to obtain an optimized VAE model.

[0034] Furthermore, the normalized simulation data set and reference data set obtained in step S1 are input into the optimized VAE model respectively, and the two mean matrices obtained are used as the simulation data feature matrices F s and the reference data feature matrix F r , the expression is:

[0035]

[0036] in, For nth s The mean vector corresponding to the simulated data samples, n s =1,2,...N s ; For nth r The mean vector corresponding to the reference data samples, n r =1,2,...N r ; D is and The vector dimension of .

[0037] Furthermore, the specific implementation method of step S5 includes the following steps:

[0038] S5.1. For the F obtained in step S4 sand F r Take the average by column to get the characteristic mean vector of the simulation data and the reference data feature mean vector

[0039] S5.2. Normalize the simulated data feature mean vector and the reference data feature mean vector obtained in step S5.1, and then measure the feature difference using the Euclidean distance, expressed as:

[0040]

[0041]

[0042] in, and F ms and F mr The data of the dth dimension in U, d = 1, 2, ..., D; F and L F Respectively represent F ms and F mr The maximum and minimum values ​​of all elements in D eu Represents the Euclidean distance measurement result;

[0043] S5.3. Parameter estimation is performed on the distribution corresponding to the simulation data feature matrix and the reference data feature matrix obtained in step S4. Considering that the multi-sample features follow the multivariate normal distribution, F s The distribution is expressed as P~N D (μ P ,Σ P ), F r The distribution of is represented by Q~N D (μ Q ,Σ Q ),set up but

[0044] S5.4. Calculate the JS divergence D between the two distributions P and Q JS , the expression is:

[0045]

[0046] Among them, KL(P||M) and KL(Q||M) are the KL divergences of P and Q with respect to M respectively;

[0047] If two multivariate normal distributions A~N are given k (μ A ,Σ A ) and B~N k (μ B ,Σ B), then the KL divergence between the two distributions is calculated as follows:

[0048]

[0049] Where k is the dimension of the distribution; tr(·) is the trace of the matrix, and det is the determinant of the matrix;

[0050] S5.5. D obtained in step S5.2 eu After conversion, the similarity C1 between the simulation data features and the reference data features in terms of numerical value is obtained, which is expressed as:

[0051]

[0052] Among them, C1∈(0,1], and the closer C1 is to 1, the more similar the two sets of features are in terms of numerical value. α is the scaling factor. The larger the α value, the more C1 affects D. eu The more sensitive the change;

[0053] The setting of scaling factor α and feature dimension D and distance percentage threshold D epth When the simulated data features and the reference data features are considered to be significantly different, the corresponding feature similarity C1 is reduced to 0.6, and the expression is:

[0054]

[0055] Then the scaling factor α is set as follows:

[0056]

[0057] S5.6. Convert the JS divergence obtained in step S5.4 to obtain the similarity C2 between the simulation data features and the reference data features in terms of probability distribution, expressed as:

[0058]

[0059] Among them, C2∈[0,1], and the closer C2 is to 1, the more similar the two sets of features are in terms of probability distribution;

[0060] S5.7. Calculate the reliability of simulation results based on feature differences C(Y s ,Y r ), the expression is:

[0061] C(Y s ,Y r )=ω1C1+ω2C2

[0062] Among them, ω1 and ω2 are the weights of feature similarities C1 and C2, respectively, indicating the contribution of the two feature similarity measurement results to the credibility of the simulation results, and satisfying 0≤ω1,ω2≤1; ω1+ω2=1

[0063] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of a VAE-based aircraft guidance and control model intelligent verification method when executing the computer program.

[0064] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements a VAE-based intelligent verification method for an aircraft guidance and control model.

[0065] Beneficial effects of the present invention:

[0066] The VAE-based intelligent verification method for an aircraft guidance and control model described in the present invention performs simulation result verification on the large-scale dynamic output of an aircraft guidance and control simulation model, aiming to compare the consistency of multi-sample simulation data with multi-sample reference data. Since the verification data has the characteristics of large samples and long time series, the present invention converts the data consistency analysis problem into the consistency analysis problem of the corresponding features through the method of intelligent feature extraction. In summary, in order to solve the simulation result verification problem of the above-mentioned aircraft guidance and control simulation model, a simulation result verification framework based on intelligent feature extraction is proposed for single-variable dynamic output, such as Figure 2 The variational autoencoder (VAE) model is used to verify the data for intelligent feature extraction, and the credibility of the simulation results is further obtained by measuring the feature distance.

[0067] The VAE-based intelligent verification method for an aircraft guidance and control model described in the present invention can basically always provide reasonable measurement results and draw correct verification conclusions. In addition, the verification method of the present invention greatly shortens the verification time, improves the efficiency of model verification, and has significant advantages. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a flow chart of a VAE-based intelligent verification method for an aircraft guidance and control model according to the present invention;

[0069] Figure 2 This is a framework diagram of intelligent verification of simulation results of the present invention;

[0070] Figure 3 This is a structural diagram of the VAE model of the present invention;

[0071] Figure 4 Schematic diagram of the network layer structure in the VAE model of the present invention;

[0072] Figure 5 This is the feature extraction result curve based on VAE of the present invention;

[0073] Figure 6 This is a diagram of the verification results based on VAE of the present invention;

[0074] Figure 7 This is a verification result diagram of the comparative example based on BF. DETAILED DESCRIPTION

[0075] In order to make the purpose, technical solution and advantages of the present invention more clear, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention, that is, the specific embodiments described are only part of the embodiments of the present invention, rather than all of the specific embodiments. The components of the specific embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations, and the present invention can also have other embodiments.

[0076] Therefore, the following detailed description of the specific embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents the selected specific embodiments of the present invention. Based on the specific embodiments of the present invention, all other specific embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0077] In order to further understand the content, features and effects of the present invention, the following specific implementation methods are given as examples, and the attached Figure 1 -Attached Figure 7 The detailed instructions are as follows:

[0078] Embodiment 1:

[0079] A VAE-based intelligent verification method for an aircraft guidance and control model comprises the following steps:

[0080] S1. Repeat the test to obtain a multi-sample reference data set and a multi-sample simulation data set based on the aircraft guidance control system and the aircraft guidance control simulation model, and normalize the data to obtain a normalized reference data set and a normalized simulation data set for use;

[0081] Furthermore, in step S1, S is set r and S s Represent the aircraft guidance control system and the aircraft guidance control simulation model respectively. Let S r and S s Run N times under given input conditions rtimes and N s Each time, the output data of length T is obtained at the same sampling interval, and the multi-sample reference data set Y is obtained. r and multi-sample simulation dataset Y s as follows:

[0082]

[0083] in, represents the reference time series obtained from the nth single operation of the aircraft guidance and control system; Represents the simulation time series obtained from the nth single run of the aircraft guidance and control simulation model.

[0084] The data in the collected multi-sample reference data set and the multi-sample simulation data set are normalized to obtain normalized reference data and normalized simulation data, and the expression is:

[0085]

[0086] in, Y r The nth r elements, n r =1,2,....,N r T, Y s The nth s elements, n s =1,2,....,N s T, and They are and The corresponding normalized result, {y n} indicates that Y r and Y s The set of all elements in contains (N r +N s )·T elements, min{y n} represents the set {y n}; max{y n} represents the set {y n};

[0087] S2. Build a VAE model and initialize the model structure and parameters for later use;

[0088] Furthermore, the specific implementation method of step S2 includes the following steps:

[0089] S2.1. Build the VAE model, including the encoder and decoder p θ (Y|Z) two parts;

[0090] S2.2. The encoder receives the input data Y, passes through a hidden layer consisting of a fully connected layer FC1 and a ReLU activation function, and then outputs h1. Then, it passes through two parallel fully connected layers FC2 and FC3, outputting the mean μ and variance σ respectively, and then passes through a reparameterization layer to obtain the potential representation Z, which is expressed as:

[0091] Z=μ+ε×σ

[0092] Among them, ε is obtained by multiple sampling of the standard Gaussian distribution N(0,1);

[0093] S2.3. The decoder receives the latent representation Z, passes through a hidden layer consisting of a fully connected layer FC4 and a ReLU activation function, and then outputs h2, which is then passed through a fully connected layer FC5 and a Sigmoid activation function, and finally outputs the reconstructed data Y′;

[0094] The potential representation is a vector in a latent space mapped from an input data sample by an encoder, and the dimension of the latent space is set according to the complexity of the data. For low-dimensional data, the dimension of the latent space is set to 2-20 dimensions; for high-dimensional data, the dimension of the latent space is set to 20-100 dimensions;

[0095] The hidden layer is used to capture important features of the input data. Its dimension and number of layers are determined according to the length and complexity of the time series. If the input data dimension is less than 1000 dimensions, the dimension of the hidden layer is set to 2 to 10 times the input dimension; if the input data dimension is greater than 1000 dimensions, the dimension of the hidden layer is set to 1 / 2, 1 / 4 or 1 / 10 of the input dimension; if the input data dimension is greater than 1000 dimensions and has complex dynamic characteristics, 1-2 hidden layers are added to the encoder and decoder respectively;

[0096] S3. Using the normalized reference data set obtained in step S1 to train the VAE model constructed in step S2, to obtain an optimized VAE model;

[0097] Considering that the reference data is the output data of the real system, it is the "expected value" corresponding to the output data of the simulation model. In the simulation model verification problem, the reference data is used as a benchmark to compare the difference between the simulation data and the reference data. Therefore, when training the deep learning model, the reference data is used as the training data, hoping that the model can effectively extract the features of the reference data through training and optimization. Then the optimized model is used to extract the features of all verification data, so that when the model parameters and structure are the same, the difference between the simulation data and the reference data can be characterized by the difference in the corresponding features of the two.

[0098] The training of VAE aims to learn effective latent representations while maintaining the continuity of the latent space, which is achieved by minimizing the reconstruction error and the regularization term of the latent space. Therefore, the loss function of VAE contains two parts: reconstruction loss (ReconstructionLoss) and KL divergence (Kullback-Leibler Divergence). Among them, the reconstruction loss is used to measure the difference between the reconstructed data and the original data; the KL divergence is used to measure the difference between the latent space distribution of the encoder output and the prior distribution (usually assumed to be a standard normal distribution). Since the mean square error (MSE) measures the average square difference between the model output and the true data, it can intuitively reflect the distance between the predicted value and the true value. It is suitable for the case where the input and output of the model are continuous numerical data. MSE is used to calculate the reconstruction loss;

[0099] Furthermore, the normalized reference data set obtained in step S1 is input into the VAE model constructed in step S2 for model training. The specific implementation method is to first map the input data to the latent space through the encoder, then reconstruct the data through the decoder, and then calculate the training loss. If the termination condition is not met, the parameters of the VAE model are updated through back propagation.

[0100] Setting the loss function of the VAE model The expression is:

[0101]

[0102] in, and θ are the parameters of the encoder and decoder respectively; Y is the input reference data sample set; Z is the potential representation corresponding to Y; is the original input data Y (i) The expected value of the MSE between and the reconstructed data Y′; is the approximate posterior distribution of the latent representation Z learned by the encoder With the prior distribution p θ The Kullback-Leibler (KL) divergence between (Z) is expressed as:

[0103]

[0104] Where D is the dimension of the latent space; μ d and are the mean and variance of the d-th dimension potential representation, respectively;

[0105] Repeatedly train the VAE model to optimize the model's parameters and structure. During the training process, monitor the changes in the loss value and analyze the latent space distribution parameters after training, adjust the model's structure and hyperparameters. The model's structure includes the number of hidden layers, adjusts the hidden layer dimension and the latent space dimension, and the hyperparameters include the learning rate and batch size to obtain an optimized VAE model.

[0106] S4. The normalized reference data set and the normalized simulation data set obtained in step S1 are respectively input into the optimized VAE model for feature extraction to obtain a reference data feature matrix and a simulation data feature matrix;

[0107] Each data sample input to the encoder corresponds to a set of mean vectors and variance vectors in the output latent space. The mean vector can represent the main information of the input data and is usually considered to be a compressed representation of the data sample. It defines the position in the latent space and is directly related to the content and structure of the data. The variance vector mainly represents the uncertainty in the latent representation. It indicates the possible range of variation of the latent representation around a given mean and does not directly reflect the content characteristics of the data.

[0108] In order to achieve the subsequent feature difference measurement, it is hoped to extract the content and structural features of each data sample, while the uncertainty features are mainly for each group of verification data, which are reflected by the feature distribution of multiple samples. Therefore, in order to simplify the calculation, the two mean matrices output by the encoder are taken as the simulation output data F s and the feature F of the reference data r ;

[0109] Furthermore, the normalized simulation data set and reference data set obtained in step S1 are input into the optimized VAE model respectively, and the two mean matrices obtained are used as the simulation data feature matrices F s and the reference data feature matrix F r , the expression is:

[0110]

[0111] in, For nth s The mean vector corresponding to the simulated data samples, n s =1,2,...N s ; For nth r The mean vector corresponding to the reference data samples, n r =1,2,...N r ; D is and The vector dimension of

[0112] S5. Using the reference data feature matrix and the simulation data feature matrix obtained in step S4, measure the difference between the reference data features and the simulation data features in terms of numerical value and probability distribution, and then obtain the credibility of the aircraft guidance and control model;

[0113] The above verification data feature extraction is realized through the unsupervised learning method. Each sample corresponds to a D-dimensional feature vector. The feature vector obtained by the multi-sample verification data is also multi-sample, that is, the simulation data and the reference data each correspond to a set of feature vectors. and On this basis, the consistency measurement problem between simulation data and reference data can be transformed into the similarity measurement problem of two sets of features, and a simulation credibility calculation method based on feature differences is proposed.

[0114] Since each set of features is composed of one-dimensional vectors of multiple samples, it is expressed in the form of a two-dimensional matrix, with two dimensions, namely, feature dimension and sample dimension. Therefore, the differences between the two sets of features are compared from two dimensions. First, from the feature dimension, the differences in the average features of multiple samples are compared; second, from the sample dimension, the differences in the distribution of features of multiple samples are compared;

[0115] Furthermore, the specific implementation method of step S5 includes the following steps:

[0116] S5.1. For the F obtained in step S4 s and F r Take the average by column to get the characteristic mean vector of the simulation data and the reference data feature mean vector

[0117] S5.2. Normalize the simulated data feature mean vector and the reference data feature mean vector obtained in step S5.1, and then measure the feature difference using the Euclidean distance, expressed as:

[0118]

[0119] in, and F ms and F mr The data of the dth dimension in U, d = 1, 2, ..., D; F and L F Respectively represent F ms and F mr The maximum and minimum values ​​of all elements in D eu Represents the Euclidean distance measurement result;

[0120] S5.3. Parameter estimation is performed on the distribution corresponding to the simulation data feature matrix and the reference data feature matrix obtained in step S4. Considering that the multi-sample features follow the multivariate normal distribution, F s The distribution is expressed as P~N D (μ P ,Σ P ), F r The distribution of is represented by Q~N D (μ Q ,Σ Q ),set up but

[0121] S5.4. Calculate the JS divergence D between the two distributions P and Q JS , the expression is:

[0122]

[0123] Among them, KL(P||M) and KL(Q||M) are the KL divergences of P and Q with respect to M respectively;

[0124] If two multivariate normal distributions A~N are given k (μ A ,Σ A ) and B~N k (μ B ,Σ B ), then the KL divergence between the two distributions is calculated as follows:

[0125]

[0126] Where k is the dimension of the distribution; tr(·) is the trace of the matrix, and det is the determinant of the matrix;

[0127] S5.5. D obtained in step S5.2 eu After conversion, the similarity C1 between the simulation data features and the reference data features in terms of numerical value is obtained, which is expressed as:

[0128]

[0129] Among them, C1∈(0,1], and the closer C1 is to 1, the more similar the two sets of features are in terms of numerical value. α is the scaling factor. The larger the α value, the more C1 affects D. eu The more sensitive the change;

[0130] The setting of scaling factor α and feature dimension D and distance percentage threshold D epth When the simulated data features and the reference data features are considered to be significantly different, the corresponding feature similarity C1 is reduced to 0.6, and the expression is:

[0131]

[0132] Then the scaling factor α is set as follows:

[0133]

[0134] S5.6. Convert the JS divergence obtained in step S5.4 to obtain the similarity C2 between the simulation data features and the reference data features in terms of probability distribution, expressed as:

[0135]

[0136] Among them, C2∈[0,1], and the closer C2 is to 1, the more similar the two sets of features are in terms of probability distribution;

[0137] S5.7. Calculate the reliability of simulation results based on feature differences C(Y s ,Y r ), the expression is:

[0138] C(Y s ,Y r )=ω1C1+ω2C2

[0139] Among them, ω1 and ω2 are the weights of feature similarities C1 and C2 respectively, indicating the contribution of the two feature similarity measurement results to the credibility of the simulation results, and satisfying 0≤ω1,ω2≤1; ω1+ω2=1.

[0140] Since the evaluators pay more attention to the characteristics of the verification data than to the uncertainty of the data, ω1 is often set to be greater than ω2, usually ω1 = 0.7, ω2 = 0.3. s ,Y r ) is the credibility of the simulation results, and satisfies C(Y s ,Y r )∈(0,1].C(Y s ,Y r ) is smaller, the simulation data Y s Relative to Y r The worse the consistency, the less credible the simulation model.

[0141] The method of this embodiment is described in detail by taking the simulation model of the aircraft terminal guidance stage as an example as follows:

[0142] The atmospheric density coefficient is an uncertain parameter in the model, which leads to uncertainty in the model output data. The simulation model under certain parameter conditions is used as the reference model. The simulation test includes a reference model and five simulation models. The ballistic inclination angle is used as the dynamic output variable to be verified. The drag disturbance coefficient C in each model is adjusted according to the information in Table 1. DD , while keeping other parameters in the model the same, and repeatedly running the simulation model and the reference model according to the number of samples specified in Table 1 to obtain the ballistic inclination data corresponding to each model. Each model runs for 50 seconds, and the sampling interval of the output data is 0.005 seconds, that is, each sample is a time series with a length of 10000.

[0143] Table 1 Verification data information

[0144]

[0145] The simulation results are verified based on the verification data obtained from the above simulation test. First, the verification data feature extraction is realized through the VAE model. Figure 4 The VAE model is constructed with the network structure of , and the dimensions of the hidden layer and the latent space are set to 500 and 50 respectively according to the length of the validation data time series. After training and optimization, the optimized VAE model is used to extract the features of the validation data. Figure 5 The average eigenvectors of the output data of the six models in the experiment are shown. Since the data distribution is relatively close, a local zoom is performed for easy observation. It can be seen that the smaller the difference in model parameters between the simulation model and the reference model, the closer the average characteristic distribution of the output data is, which is consistent with the actual situation.

[0146] Then, based on the extracted feature data, the credibility of the simulation results is further calculated. After measuring the feature distance by two methods, the distance percentage threshold D epth Set to 5% to complete the conversion of credibility. Set the weights of the two feature similarities ω1 = 0.7, ω2 = 0.3, and perform credibility synthesis to obtain the verification results of each simulation model in the experiment as shown in Table 2.

[0147] Table 2 Verification results based on VAE

[0148]

[0149] The verification results in Table 2 show that the smaller the difference in model parameters between the simulation model and the reference model, the closer the verification result is to 1; as the difference in model parameters increases, the verification result becomes smaller, which shows that the present invention can effectively verify the simulation results for single variable dynamic output with large samples and long time series.

[0150] In order to verify the superiority of the present invention, the proposed method is compared with a simulation dynamic output verification method based on data features (this method realizes simulation result verification by artificially extracting the features of the verification data and performing feature consistency analysis based on an improved Bayesian factor (BF), hereinafter referred to as the BF-based verification method). The two methods are applied to verify the above simulation results, and considering the uncertainty of the verification results, the verification is repeated 10 times respectively. The verification results obtained by the two methods are statistically shown as follows: Figure 6 and Figure 7 shown.

[0151] In the BF-based verification method, a null hypothesis is made that the simulation data is consistent with the reference data. If the verification result exceeds 0.5, the hypothesis is accepted; and the larger the verification result, the higher the consistency between the corresponding simulation data and the reference data. Figure 6 It can be seen that the BF-based verification method can sensitively give reasonable verification results when the model credibility is low, but it is not good at distinguishing models with higher credibility. The VAE-based verification method can basically always provide reasonable measurement results and draw correct verification conclusions. In addition, the average verification time of the BF-based verification method and the VAE-based verification method are 530.71 seconds and 43.86 seconds respectively, indicating that the VAE-based verification method greatly shortens the verification time, improves the efficiency of model verification, and has significant superiority.

[0152] Embodiment 2:

[0153] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of an intelligent verification method for an aircraft guidance and control model based on VAE described in Example 1 are implemented.

[0154] The computer device of the present invention may be a device including a processor and a memory, such as a single chip microcomputer including a central processing unit, etc. Moreover, the processor is used to implement the steps of the above-mentioned intelligent verification method of an aircraft guidance and control model based on VAE when executing the computer program stored in the memory.

[0155] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), off-the-shelf programmable gate arrays (FPGAs), or a processor.

[0156] Field-Programmable Gate Array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general processor can be a microprocessor or the processor can also be any conventional processor, etc.

[0157] The memory may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, a phone book, etc.), etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0158] Embodiment 3:

[0159] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the VAE-based intelligent verification method for an aircraft guidance and control model described in Example 1.

[0160] The computer-readable storage medium of the present invention can be any form of storage medium that can be read by a processor of a computer device, including but not limited to non-volatile memory, volatile memory, ferroelectric memory, etc. A computer program is stored on the computer-readable storage medium. When the processor of the computer device reads and executes the computer program stored in the memory, the steps of the above-mentioned VAE-based aircraft guidance and control model intelligent verification method can be implemented.

[0161] The computer program includes computer program code, which may be in source code form, object code form, executable file or some intermediate form, etc. The computer readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer readable medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer readable media do not include electric carrier signals and telecommunication signals.

[0162] It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0163] Although the present application has been described above with reference to specific embodiments, various modifications may be made thereto and parts thereof may be replaced with equivalents without departing from the scope of the present application. In particular, as long as there is no structural conflict, the various features in the specific embodiments disclosed in the present application may be used in combination with each other in any manner, and the fact that these combinations are not exhaustively described in this specification is only for the sake of omitting space and saving resources. Therefore, the present application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A VAE-based intelligent verification method for aircraft guidance and control models, characterized in that: The steps include: S1. Repeat the test to obtain a multi-sample reference data set and a multi-sample simulation data set based on the aircraft guidance control system and the aircraft guidance control simulation model, and normalize the data to obtain a normalized reference data set and a normalized simulation data set for use; S2. Build a VAE model and initialize the model structure and parameters for later use; S3. Using the normalized reference data set obtained in step S1 to train the VAE model constructed in step S2, to obtain an optimized VAE model; S4. The normalized reference data set and the normalized simulation data set obtained in step S1 are respectively input into the optimized VAE model for feature extraction to obtain a reference data feature matrix and a simulation data feature matrix; S5. Using the reference data feature matrix and the simulation data feature matrix obtained in step S4, measure the difference between the reference data features and the simulation data features in terms of numerical value and probability distribution, and then obtain the credibility of the aircraft guidance and control model; The specific implementation method of step S5 includes the following steps: S5.

1. For the F obtained in step S4 s and F r Take the average by column to get the characteristic mean vector of the simulation data and the reference data feature mean vector S5.

2. Normalize the simulated data feature mean vector and the reference data feature mean vector obtained in step S5.1, and then measure the feature difference using the Euclidean distance, expressed as: in, and F ms and F mr The data of the dth dimension in U, d = 1, 2, ..., D; F and L F Respectively represent F ms and F mr The maximum and minimum values ​​of all elements in D eu Represents the Euclidean distance measurement result; S5.

3. Parameter estimation is performed on the distribution corresponding to the simulation data feature matrix and the reference data feature matrix obtained in step S4. Considering that the multi-sample features follow the multivariate normal distribution, F s The distribution is expressed as P~N D (μ P ,Σ P ), F r The distribution of is represented by Q~N D (μ Q ,Σ Q ),set up but S5.

4. Calculate the JS divergence D between the two distributions P and Q JS , the expression is: Among them, KL(P||M) and KL(Q||M) are the KL divergences of P and Q with respect to M respectively; If two multivariate normal distributions A~N are given k (μ A ,Σ A ) and B~N k (μ B ,Σ B ), then the KL divergence between the two distributions is calculated as follows: Where k is the dimension of the distribution; tr(·) is the trace of the matrix, and det is the determinant of the matrix; S5.

5. D obtained in step S5.2 eu After conversion, the similarity C1 between the simulation data features and the reference data features in terms of numerical value is obtained, which is expressed as: Among them, C1∈(0,1], and the closer C1 is to 1, the more similar the two sets of features are in terms of numerical value. α is the scaling factor. The larger the α value, the more C1 affects D. eu The more sensitive the change; The setting of scaling factor α and feature dimension D and distance percentage threshold D epth When the simulated data features and the reference data features are considered to be significantly different, the corresponding feature similarity C1 is reduced to 0.6, and the expression is: Then the scaling factor α is set as follows: S5.

6. Convert the JS divergence obtained in step S5.4 to obtain the similarity C2 between the simulation data features and the reference data features in terms of probability distribution, expressed as: Among them, C2∈[0,1], and the closer C2 is to 1, the more similar the two sets of features are in terms of probability distribution; S5.

7. Calculate the reliability of simulation results based on feature differences C(Y s ,Y r ), the expression is: C(Y s ,Y r )=ω1C1+ω2C2 Among them, ω1 and ω2 are the weights of feature similarities C1 and C2 respectively, indicating the contribution of the two feature similarity measurement results to the credibility of the simulation results, and satisfying 0≤ω1,ω2≤1; ω1+ω2=1.

2. The VAE-based intelligent verification method for aircraft guidance and control model according to claim 1, characterized in that: In step S1, set S r and S s Represent the aircraft guidance control system and the aircraft guidance control simulation model respectively. Let S r and S s Run N times under given input conditions r times and N s Each time, the output data of length T is obtained at the same sampling interval, and the multi-sample reference data set Y is obtained. r and multi-sample simulation dataset Y s as follows: in, represents the reference time series obtained from the nth single operation of the aircraft guidance and control system; represents the simulation time series obtained from the nth single run of the aircraft guidance and control simulation model; The data in the collected multi-sample reference data set and the multi-sample simulation data set are normalized to obtain normalized reference data and normalized simulation data, and the expression is: in, Y r The nth r elements, n r =1,2,....,N r T, Y s The nth s elements, n s =1,2,....,N s T, and They are and The corresponding normalized result, {y n } indicates that Y r and Y s The set of all elements in contains (N r +N s )·T elements, min{y n } represents the set {y n }; max{y n } represents the set {y n } is the largest element in .

3. The VAE-based intelligent verification method for aircraft guidance and control model according to claim 2, characterized in that: The specific implementation method of step S2 includes the following steps: S2.

1. Build the VAE model, including the encoder and decoder p θ (Y|Z) two parts; S2.

2. The encoder receives the input data Y, passes through a hidden layer consisting of a fully connected layer FC1 and a ReLU activation function, and then outputs h1. Then, it passes through two parallel fully connected layers FC2 and FC3, outputting the mean μ and variance σ respectively, and then passes through a reparameterization layer to obtain the potential representation Z, which is expressed as: Z=μ+ε×σ Among them, ε is obtained by multiple sampling of the standard Gaussian distribution N(0,1); S2.

3. The decoder receives the latent representation Z, passes through a hidden layer consisting of a fully connected layer FC4 and a ReLU activation function, and then outputs h2, which is then passed through a fully connected layer FC5 and a Sigmoid activation function, and finally outputs the reconstructed data Y′; The potential representation is a vector in a latent space mapped from an input data sample by an encoder, and the dimension of the latent space is set according to the complexity of the data. For low-dimensional data, the dimension of the latent space is set to 2-20 dimensions; for high-dimensional data, the dimension of the latent space is set to 20-100 dimensions; The hidden layer is used to capture important features of the input data. Its dimension and number of layers are determined according to the length and complexity of the time series. If the input data dimension is less than 1000 dimensions, the dimension of the hidden layer is set to 2 to 10 times the input dimension; if the input data dimension is greater than 1000 dimensions, the dimension of the hidden layer is set to 1 / 2, 1 / 4 or 1 / 10 of the input dimension; if the input data dimension is greater than 1000 dimensions and has complex dynamic characteristics, 1-2 hidden layers are added to the encoder and decoder respectively.

4. The VAE-based intelligent verification method for aircraft guidance and control model according to claim 3, characterized in that: The normalized reference data set obtained in step S1 is input into the VAE model constructed in step S2 for model training. The specific implementation method is to first map the input data to the latent space through the encoder, then reconstruct the data through the decoder, and then calculate the training loss. If the termination condition is not met, the parameters of the VAE model are updated through back propagation; Setting the loss function of the VAE model The expression is: in, and θ are the parameters of the encoder and decoder respectively; Y is the input reference data sample set; Z is the potential representation corresponding to Y; is the original input data Y (i) The expected value of the mean square error between and the reconstructed data Y′; is the approximate posterior distribution of the latent representation Z learned by the encoder With the prior distribution p θ The Kullback-Leibler (KL) divergence between (Z) is expressed as: Where D is the dimension of the latent space; μ d and are the mean and variance of the d-th dimension potential representation, respectively; Repeatedly train the VAE model to optimize the model's parameters and structure. During the training process, monitor the changes in the loss value and analyze the latent space distribution parameters after training, adjust the model's structure and hyperparameters. The model's structure includes the number of hidden layers, adjusts the hidden layer dimension and the latent space dimension, and the hyperparameters include the learning rate and batch size to obtain an optimized VAE model.

5. The VAE-based intelligent verification method for aircraft guidance and control model according to claim 4, characterized in that: The normalized simulation data set and reference data set obtained in step S1 are input into the optimized VAE model respectively, and the two mean matrices obtained are used as the simulation data feature matrix F s and the reference data feature matrix F r , the expression is: in, For nth s The mean vector corresponding to the simulated data samples, n s =1,2,...N s ; For nth r The mean vector corresponding to the reference data samples, n r =1,2,...N r ; D is and The vector dimension of .

6. An electronic device, characterized in that: It comprises a memory and a processor, the memory stores a computer program, and the processor implements the steps of a VAE-based aircraft guidance and control model intelligent verification method as described in any one of claims 1 to 5 when executing the computer program.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the VAE-based intelligent verification method for an aircraft guidance and control model as described in any one of claims 1 to 5 is implemented.

Images