Flight control system virtual test verification method based on generative adversarial network

Abstract

The invention belongs to a flight control system virtual test, and relates to a flight control system virtual test verification method based on a generative adversarial network, which comprises the following steps of: 1, acquiring flight control computer data from a flight control computer of an aircraft; 2, preprocessing the flight data of the flight control computer; 3, building a generative adversarial network model; 4, training the generative adversarial network model; 5, generating an enhanced data sample, and generating new data on the basis of the trained generative adversarial network model; according to the method, data enhancement is carried out on small sample flight control computer data, the sample size is enlarged, and the small sample problem is solved.

Description

Technical Field

[0001] The present invention belongs to a flight control system virtual test, and relates to a flight control system virtual test verification method based on a generative adversarial network. Background Art

[0002] In the aviation field, the flight control computer plays a vital role, which is equivalent to the "central nervous system" of the aircraft. In order to ensure that the simulation model of the flight control computer can truly reflect the flight behavior of the aircraft, the seamless connection between virtual test data and actual flight data is the core link to verify the effectiveness of the simulation model. The accuracy of the simulation data of the flight control computer is directly related to the accuracy of the design, analysis and verification of the flight control system, and its importance is self-evident.

[0003] However, in real engineering practice, flight tests are not only expensive and risky, but also difficult to fully simulate actual flight conditions, which results in the fact that actual flight control computer data often has the characteristics of small samples. This limitation seriously restricts the effective application of flight control computer data in simulation model verification, making the verification of simulation models face many challenges.

[0004] At present, there are various methods for verifying the consistency of simulation data, among which the most commonly used and critical ones include static data and dynamic data verification methods. Static data verification can be carried out through graphical methods, statistical analysis methods, hypothesis testing methods, etc., while dynamic data verification can be carried out by similarity comparison methods and feature comparison methods. These methods have been successfully applied in the verification of various simulation test models.

[0005] Nevertheless, due to the cost constraints of actual aircraft tests, the small sample characteristics of actual data are in sharp contrast to the large sample size of virtual test results. This difference may cause flight test information to be submerged in massive simulation data, thereby increasing the risk of decision-making. To meet this challenge, researchers have proposed a series of solutions, such as Bayesian estimation, Bootstrap method, master-slave integration test, etc. Although these methods help solve the problem to a certain extent, they may increase the complexity of the calculation, rely on additional assumptions, or introduce additional uncertainties, which may further increase the risk of decision-making.

[0006] The limitation of existing technologies is that they may not be effective enough when dealing with small sample data. This is because these methods usually rely on large sample theories or assumptions, which may not be applicable in small sample cases. In addition, the statistical characteristics of small sample data may not fully capture the complexity of the real world, which limits the reliability and credibility of the verification results. Therefore, developing a more accurate simulation data consistency verification method that is suitable for small sample characteristics has important practical significance and far-reaching strategic value for improving the verification efficiency and accuracy of flight control computer simulation models. Summary of the invention

[0007] Purpose of the invention: The present invention provides a flight control system virtual test verification method that uses a generative adversarial network to perform data enhancement on small sample flight control computer data, expand the sample size, and solve the small sample problem.

[0008] The technical solution of the present invention:

[0009] A virtual test verification method for a flight control system based on a generative adversarial network includes the following steps:

[0010] Step 1: Collect flight control computer data from the flight control computer of the aircraft;

[0011] Step 2: Preprocess the flight data of the flight control computer;

[0012] Step 3: Build a generative adversarial network model;

[0013] Step 4: Train the generative adversarial network model;

[0014] Step 5: Generate enhanced data samples and generate new data based on the trained generative adversarial network model;

[0015] Step 6: Static and dynamic data consistency verification.

[0016] Furthermore, step 2 specifically cleans the flight data of the flight control computer, including processing data missing, duplication, and singular values.

[0017] Furthermore, after the flight data is cleaned, it is necessary to perform noise reduction on the cleaned flight data. The noise reduction process is: sliding a window on the cleaned flight control computer data X to obtain multiple window data. The multiple window data represent the flight control computer status within a certain period of time. The data set after the sliding window processing is where f i w ={x s+i×d ,x s+i×d-1 ,…,x i×d}∈R s×N,i∈{0,1,2,…,M-1},s+(M-1)×d≤T,f i w Represents the window data at the i-th moment; s represents the width of a single window; the sliding window step is d, which determines the sensitivity of the algorithm. The larger d is, the lower the sensitivity is; the smaller d is, the greater the sensitivity is; M is the total number of samples after sliding window processing, and this value is related to the total time length T, the window width s, and the sliding window step d; take a sliding window every d time points on the cleaned flight control computer data X, that is, every d time points, the number of samples is increased by 1 until the sliding window slides to the end, and the total number of samples obtained is recorded as, and data denoising is achieved through sliding window processing.

[0018] Further, in step 3, building a generative adversarial network model includes the following steps:

[0019] Step 3.1 Define the generator: The task of the generator is to convert random noise vectors into data points. The generator consists of a series of neural network layers, including fully connected layers and convolutional layers. The output of the generator has the same dimension as the target dataset.

[0020] Step 3.2 Define the discriminator: The task of the discriminator is to distinguish the fake data generated by the generator from the real data in the real data set. The discriminator is a two-classification model that uses convolutional layers, pooling layers, and fully connected layers.

[0021] Step 3.3 Initialize weights: Initialize weights for the generator and discriminator networks;

[0022] Step 3.4 Compile the model: Compile the model for the generator and discriminator, that is, set the appropriate loss function and optimizer; for the discriminator, the loss function is binary cross entropy, and for the generator, the loss function is defined by maximizing the probability of the discriminator output.

[0023] Furthermore, step 4 trains the generative adversarial network model, including a generator training stage and a discriminator training stage.

[0024] Furthermore, the generator training phase is as follows: the weights of the discriminator are fixed and only the generator is trained; the goal of the generator is to generate data that can deceive the discriminator, and the input of the generator is random Gaussian noise The output is F(N Gaussian ), which is then used as the discriminator input. σ represents random Gaussian noise N Gaussian The standard deviation of μ represents the random Gaussian noise N Gaussian The mean of; a represents a random variable;

[0025] Training the discriminator: fix the weights of the generator and only train the discriminator. Use the window data f after sliding window processing in step 2 i wand the generator generated G(N Gaussian ) as the discriminator input, allowing the discriminator to learn to distinguish the two data.

[0026] Furthermore, step 5: generate enhanced data samples, and generate new data based on the trained generative adversarial network model. Specifically, first, load the trained generator, which has learned how to simulate the distribution characteristics of real flight data; then, the generator starts to generate data according to the set sequence length; for each new data point, the generator will consider the previous data point and generate samples based on the learned probability distribution; this process will be repeated until a complete data sequence is generated. The generator can generate multiple complete data sequences in parallel, thereby quickly expanding a large number of new data samples; then, the newly generated data is verified and adjusted to ensure that the newly generated data meets specific application requirements and data quality standards.

[0027] Furthermore, step 6 specifically combines the flight control computer data collected in step 1 and the newly generated data in step 5 to form combined data, divides the combined data into two categories: static data and dynamic data, and performs consistency verification on the static data and the dynamic data.

[0028] Further, static data consistency verification: When the actual flight control computer data is enhanced, the data sample size is expanded and it is possible to satisfy the normal distribution. The F test method is used, that is, the variables of the combined data and the virtual test data to be verified are P and Q respectively, and their distribution functions are set to F(p) and F(q). The following formula is used to determine whether it is established: H0: F(p) = F(q).

[0029] further,

[0030] Dynamic data consistency verification: Verification is carried out through the grey correlation analysis method. The specific steps are as follows:

[0031] Data segmentation: If the dynamic data is long, the data sequence is divided into several segments according to the characteristics of the data and the needs of model verification. Each segment consists of a certain number of data points.

[0032] Calculate the grey relational degree: For each segmented data, use the grey relational degree model to calculate the grey relational degree γ of each data point k ,The grey relational degree is determined by comparing the similarity between the ,combined data and the virtual test data to be verified;

[0033] Determine weight: Assign weight ω to each grey relational degree based on expert experience k ;

[0034] Solving the comprehensive grey correlation degree: Using the weighted average method, the grey correlation degrees of all segments are combined with weights to obtain the comprehensive grey correlation degree γ, which is expressed as Where N is the number of segments.

[0035] Beneficial effects of the invention: The invention proposes that the invention provides a flight control system virtual test verification method which uses a generative adversarial network to perform data enhancement on small sample flight control computer data, expand the sample size, and solve the small sample problem. It is more accurate and suitable for the consistency verification of simulation data with small sample characteristics, and has important practical significance and far-reaching strategic value for improving the verification efficiency and accuracy of flight control computer simulation models. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flow chart of the present invention;

[0037] Figure 2 This is the principle diagram of sliding window noise reduction processing;

[0038] Figure 3 Schematic diagram of the generative adversarial network model structure. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] The present invention provides a virtual test verification method for a flight control system based on a generative adversarial network.

[0041] The present invention implements the virtual test verification of the flight control system based on the generative adversarial network through the following steps:

[0042] Step 1: Data Collection

[0043] The flight control computer data is collected from the flight control computer of a certain type of aircraft.

[0044] Flight control computer data include: flight altitude, flight speed, pitch angle, roll angle, Mach number, and yaw angle.

[0045] Step 2: Data Preprocessing

[0046] The flight control computer data is cleaned, that is, data missing, duplication, singular values, etc. are processed. After cleaning, noise reduction is performed. The noise reduction process is: sliding a sliding window on the cleaned flight control computer data X to obtain multiple window data. The multiple window data represent the flight control computer status within a fragment of time. The data set after sliding window processing is where f i w ={x s+i×d ,x s+i×d-1 ,…,x i×d}∈R s×N ,i∈{0,1,2,…,M-1},s+(M-1)×d≤T,f i w Represents the window data at the i-th moment; s represents the width of a single window; the sliding window step is d, which determines the sensitivity of the algorithm. The larger d is, the lower the sensitivity is; the smaller d is, the greater the sensitivity is. M is the total number of samples after sliding window processing. This value is related to the total time length T, the window width s, and the sliding window step d. Take a sliding window every d time points on the cleaned flight control computer data X, that is, every d time points, the number of samples is increased by 1 until the sliding window slides to the end, and the total number of samples obtained is recorded as M. Data denoising is achieved through sliding window processing, such as Figure 2 shown.

[0047] In step 2, linear interpolation is used to fill missing value data; duplicate value data is removed according to timestamp information; singular value data is directly deleted.

[0048] Step 3: Generative Adversarial Network Model Construction

[0049] Generative Adversarial Networks (GANs) is a deep learning model. It consists of two parts: the generator and the discriminator. These two parts compete with each other during the training process. The goal of the generator is to generate data that is as realistic as possible, while the goal of the discriminator is to distinguish between real data and fake data generated by the generator as accurately as possible. Its specific sub-steps include:

[0050] Step 3.1 Define the generator: The task of the generator is to convert the random noise vector into data points. It usually consists of a series of neural network layers, which can be fully connected layers, convolutional layers (for data generation). The output of the generator should have the same dimension as the target dataset.

[0051] Step 3.2 Define the discriminator: The task of the discriminator is to distinguish between the fake data generated by the generator and the real data in the real dataset. It is usually a two-classification model, using convolutional layers, pooling layers, and fully connected layers.

[0052] Step 3.3 Initialize weights: Initialize weights for the generator and discriminator networks. Weight initialization is critical to the training and performance of the model.

[0053] Step 3.4 Compile the model: Compile the model for the generator and discriminator, set the appropriate loss function and optimizer. For the discriminator, the loss function is usually binary cross entropy because it is a binary classification problem. For the generator, the loss function is usually defined by maximizing the probability of the discriminator output.

[0054] The generator uses a recurrent neural network (RNN) and the discriminator uses a convolutional neural network (CNN).

[0055] The structure of the built generative adversarial network model is as follows: Figure 3 shown.

[0056] Step 4: Generative adversarial network model training; The generative adversarial network model training process includes two stages:

[0057] Train the generator: fix the weights of the discriminator and only train the generator. The goal of the generator is to generate data that can deceive the discriminator. The input of the generator is random Gaussian noise The output is G(N Gaussian ), which is then used as the discriminator input. σ represents random Gaussian noise N Gaussian The standard deviation of μ represents the random Gaussian noise N Gaussian The mean of ; a represents a random variable.

[0058] Training the discriminator: fix the weights of the generator and only train the discriminator. Use the window data f after sliding window processing in step 2 i w and the generator generated G(N Gaussian ) as the discriminator input, and let the discriminator learn to distinguish them.

[0059] The training of the generator and discriminator uses batch (Mini-batch) stochastic gradient descent based on the Adam algorithm to update the parameters, with the input Sequence_Length of 50 and Batch_size of 5000 and 400 respectively. The initial learning rate of the generator model is 0.01, the decay rate is 0.95, and the learning rate of the discriminator model is 0.0001.

[0060] Step 5: Generate enhanced data samples, the process of generating new data based on the trained generative adversarial network model: First, load the trained generator, which has learned how to simulate the distribution characteristics of real flight data. Next, the generator starts generating data according to the set sequence length (such as 50 data points). For each new data point, the generator considers the previous data point and generates samples based on the learned probability distribution. This process is repeated until a complete data sequence is generated. The generator can generate multiple such sequences in parallel, thereby quickly expanding a large number of new data samples. Subsequently, these newly generated data will undergo necessary post-processing, including verification and adjustment, to ensure that they meet specific application requirements and data quality standards.

[0061] Step 6: Verify the consistency of static and dynamic data. Combine the flight control computer data collected in step 1 and the newly generated data in step 5 to form combined data. Divide the combined data into two categories: static data and dynamic data.

[0062] For static data: When the actual flight control computer data is enhanced, the data sample size is expanded, and it is possible to satisfy the normal distribution. The F test method is used, that is, the variables of the combined data and the virtual test data to be verified are P and Q respectively, and their distribution functions are set to F(p) and F(q). The following formula is used to determine whether it is established:

[0063] H0:F(p)=F(q).

[0064] For dynamic data: Grey correlation analysis is widely used in the consistency analysis of dynamic data due to its simplicity and effectiveness in processing dynamic data. This method can quantitatively evaluate the consistency between the dynamic performance of the simulation model and the real flight data, thus providing a basis for the verification and improvement of the simulation model. The specific steps are as follows:

[0065] Data segmentation: If the dynamic data is long, the data sequence can be divided into several sections according to the characteristics of the data and the needs of model verification. Each section consists of a certain number of data points.

[0066] Calculate the grey relational degree: For each segmented data, use the grey relational degree model to calculate the grey relational degree γ of each data point k The grey relational degree is determined by comparing the similarity between the reference sequence (combined data) and the comparison sequence (virtual test data to be verified).

[0067] Determine weight: Assign weight ω to each grey relational degree based on expert experience k The weights reflect the importance of different paragraph data in the overall evaluation.

[0068] Solve the comprehensive grey correlation degree: Use the weighted average method to combine the grey correlation degrees of all segments with the weights to obtain a comprehensive grey correlation degree γ. The formula is expressed as Where N is the number of segments.

[0069] The above is only a specific embodiment of the present invention, and the present invention is described in detail. The unexplained part is a conventional technology. However, the protection scope of the present invention is not limited to this. Any changes or substitutions that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. The protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A virtual test verification method for a flight control system based on a generative adversarial network, characterized in that: The following steps are involved: Step 1: Collect flight control computer data from the flight control computer of the aircraft; Step 2: Preprocess the flight data of the flight control computer; Step 3: Build a generative adversarial network model; Step 4: Train the generative adversarial network model; Step 5: Generate enhanced data samples and generate new data based on the trained generative adversarial network model; Step 6: Static and dynamic data consistency verification.

2. The flight control system virtual test verification method based on generative adversarial network according to claim 1 is characterized in that: Step 2 specifically involves cleaning the flight data of the flight control computer, including processing missing data, duplication, and singular value situations.

3. The flight control system virtual test verification method based on generative adversarial network according to claim 2 is characterized in that: After the flight data is cleaned, it is necessary to perform noise reduction on the cleaned flight data. The noise reduction process is: slide the sliding window on the cleaned flight control computer data X to obtain multiple window data. The multiple window data represent the flight control computer status within a certain period of time. The data set after sliding window processing is where f i w ={x s+i×d ,x s+i×d-1 ,…,x i×d }∈R s×N ,i∈{0,1,2,…,M-1},s+(M-1)×d≤T,f i w Represents the window data at the jth moment; s represents the width of a single window; the sliding window step is d, which determines the sensitivity of the algorithm. The larger d is, the lower the sensitivity is; the smaller d is, the greater the sensitivity is; M is the total number of samples after sliding window processing, and this value is related to the total time length T, the window width s, and the sliding window step d; take a sliding window every d time points on the cleaned flight control computer data X, that is, every d time points, the number of samples is increased by 1 until the sliding window slides to the end, and the total number of samples obtained is recorded as, and data denoising is achieved through sliding window processing.

4. The flight control system virtual test verification method based on generative adversarial network according to claim 3 is characterized in that: In step 3, building a generative adversarial network model includes the following steps: Step 3.1 Define the generator: The task of the generator is to convert random noise vectors into data points. The generator consists of a series of neural network layers, including fully connected layers and convolutional layers. The output of the generator has the same dimension as the target dataset. Step 3.2 Define the discriminator: The task of the discriminator is to distinguish the fake data generated by the generator from the real data in the real data set. The discriminator is a two-classification model that uses convolutional layers, pooling layers, and fully connected layers. Step 3.3 Initialize weights: Initialize weights for the generator and discriminator networks; Step 3.4 Compile the model: Compile the model for the generator and discriminator, that is, set the appropriate loss function and optimizer; for the discriminator, the loss function is binary cross entropy, and for the generator, the loss function is defined by maximizing the probability of the discriminator output.

5. The flight control system virtual test verification method based on generative adversarial network according to claim 4 is characterized in that: Step 4 trains the generative adversarial network model, including the generator training stage and the discriminator training stage.

6. The flight control system virtual test verification method based on generative adversarial network according to claim 5 is characterized in that: The generator training phase is: fix the weight of the discriminator and only train the generator; the goal of the generator is to generate data that can deceive the discriminator, and the input of the generator is random Gaussian noise The output is G(N Gaussian ), which is then used as the discriminator input. σ represents random Gaussian noise N Gaussian The standard deviation of μ represents the random Gaussian noise N Gaussian The mean of; a represents a random variable; Training the discriminator: fix the weights of the generator and only train the discriminator. Use the window data f after sliding window processing in step 2 i w and the generator generated G(N Gaussian ) as the discriminator input, allowing the discriminator to learn to distinguish the two data.

7. The flight control system virtual test verification method based on generative adversarial network according to claim 6 is characterized in that: Step 5: Generate enhanced data samples, and generate new data based on the trained generative adversarial network model. Specifically: first, load the trained generator, which has learned how to simulate the distribution characteristics of real flight data; then, the generator starts to generate data according to the set sequence length; for each new data point, the generator will consider the previous data point and generate samples based on the learned probability distribution; this process will be repeated until a complete data sequence is generated. The generator can generate multiple complete data sequences in parallel, thereby quickly expanding a large number of new data samples; then, the newly generated data is verified and adjusted to ensure that the newly generated data meets specific application requirements and data quality standards.

8. The flight control system virtual test verification method based on generative adversarial network according to claim 7 is characterized in that: Step 6 specifically combines the flight control computer data collected in step 1 and the newly generated data in step 5 to form combined data, divides the combined data into two categories: static data and dynamic data, and performs consistency verification on the static data and the dynamic data.

9. The flight control system virtual test verification method based on generative adversarial network according to claim 8, characterized in that: Static data consistency verification: When the actual flight control computer data is enhanced, the data sample size is expanded and it is possible to satisfy the normal distribution. The F test method is used, that is, the variables of the combined data and the virtual test data to be verified are P and Q respectively, and their distribution functions are set to F(p) and F(q). The following formula is used to determine whether it is valid: H0: F(p) = F(q).

10. The flight control system virtual test verification method based on generative adversarial network according to claim 8, characterized in that: Dynamic data consistency verification: Verification is carried out through the grey correlation analysis method. The specific steps are as follows: Data segmentation: If the dynamic data is long, the data sequence is divided into several segments according to the characteristics of the data and the needs of model verification. Each segment consists of a certain number of data points. Calculate the grey relational degree: For each segmented data, use the grey relational degree model to calculate the grey relational degree γ of each data point k ,The grey relational degree is determined by comparing the similarity between the ,combined data and the virtual test data to be verified; Determine weight: Assign weight ω to each grey relational degree based on expert experience k ; Solving the comprehensive grey correlation degree: Using the weighted average method, the grey correlation degrees of all segments are combined with weights to obtain the comprehensive grey correlation degree γ, which is expressed as Where N is the number of segments.

Images