A method for expanding eVTOL runway database

Abstract

This application provides a method for expanding an eVTOL runway database, belonging to the technical field of runway data. Specifically, it includes the following steps: designing an activation function S_ReLU and introducing a generative adversarial model; constructing a generative network model G and designing a generator loss function V. G =V G′ +V T The proposed method involves constructing a discriminant network model D, designing a discriminator loss function to train the network model, and training the generator and discriminant network models separately and alternately, fixing one network model while optimizing the other. The desired samples are then generated by randomly selecting noise samples z from the noise distribution and inputting them into the trained generator network model G. This processing scheme improves the runway recognition rate and confidence level.

Description

Technical Field

[0001] This application relates to the field of runway data, and in particular to a method for expanding an eVTOL runway database. Background Technology

[0002] eVTOL (eVehicle-to-Air Transport) aircraft first gained attention due to the concept of the Airbus. Utilizing distributed electric propulsion technology, eVTOLs are characterized by their lightweight, flexibility, energy efficiency, environmental friendliness, and low cost. They are not only highly valued in the military field, but the civilian sector is also accelerating the development of related industries. During the entire eVTOL flight process, the takeoff and landing phases are the most demanding, complex, and prone to flight problems for pilots. eVTOL intelligent takeoff and landing technology trains runway recognition algorithms using a runway database, identifies runways, detects runway edges, and performs intelligent takeoffs and landings based on the detected runway feature lines, helping to alleviate pilot stress and improve takeoff and landing safety. However, airport runway databases suffer from small sample sizes and difficulties in sample acquisition, affecting the training of runway recognition algorithms. Using a sample generation algorithm to generate reliable runway images expands the runway database, improves the training of runway recognition algorithms, and increases the recognition rate.

[0003] Sample generation algorithms primarily utilize oversampling to process datasets and generate new sample data. The simplest oversampling is simple random oversampling, which randomly replicates small samples without adding any new information, potentially leading to overfitting and wasting training time. K-nearest neighbor-based oversampling methods segment iterative data using the K-nearest neighbor algorithm, reducing noise and improving generation quality. The SMOTE algorithm, based on random oversampling, generates new data between small samples and their K-nearest neighbors using a sampling formula, effectively increasing sample diversity. Generative adversarial networks (GANs), as an emerging approach, can learn data distributions, generate small samples, and produce high-quality sample data.

[0004] Generative Adversarial Networks (GANs) consist of a generator network and a discriminator network, aiming to generate data highly similar to real data. The generator G receives a random variable Z and outputs generated data G(Z), striving to make the distribution of generated data G(Z) as similar as possible to that of real data X. The discriminator D tries to distinguish between real data X and generated data G(Z), taking both real data X and generated data G(Z) as input and outputting a probability value representing the probability that generated data G(Z) is real data X. Simultaneously, the output of the discriminator D is fed back to the generator G, aiding in its training. During training, the generator G and discriminator D engage in a game of mutual competition, continuously strengthening their generation and discrimination abilities. Ideally, after training, the discriminator D's output becomes 1 / 2, unable to distinguish the type of input data. Deep Convolutional Generative Adversarial Networks (DCGANs) add a deep convolutional network structure to GANs. The generator G uses deconvolution operations to upsample the input of each layer to generate images. The original DCGAN model is unstable during training, prone to gradient vanishing and gradient exploding problems, and the quality of generated samples needs to be improved. Summary of the Invention

[0005] In view of this, this application provides a method for expanding the eVTOL runway database, which solves the problems in the prior art, expands the eVTOL runway database, improves the training of the eVTOL runway recognition algorithm, and enhances the recognition rate and confidence.

[0006] The eVTOL runway database expansion method provided in this application adopts the following technical solution:

[0007] A method for expanding an eVTOL runway database includes the following steps:

[0008] S100, design the activation function S_ReLU function, and introduce a generative adversarial model;

[0009] S200, construct the generator network model G, and design the generator loss function V. G =V G′ +V T ;

[0010] Among them, V G To generate the loss function of the network model, V G′ V is the loss function of the original generative network model. T For a simplified total variation loss function;

[0011] S300, construct the discriminant network model D, and design the discriminant loss function.

[0012] Among them, V D To determine the loss function of the network model, V D′ The loss function of the original discriminant network model. This is a gradient penalty term;

[0013] S400: Train the network model by training the generative network model and the discriminative network model separately and alternately, fixing one network model while optimizing the other.

[0014] S500 generates the required samples by randomly selecting noise samples z from the noise distribution and inputting them into the generative network model G trained by S400 to obtain the required samples.

[0015] Optionally, the S_ReLU function in S100 is:

[0016] When x > 0, S_ReLU = ξx;

[0017] When x≤0, S_ReLU=η(e x -1).

[0018] Optionally, the S200 generative network model structure is as follows: the input layer takes noise Z as input, which passes through a fully connected layer, several deconvolutional layers, a normalization layer, and an S_ReLU activation layer to generate feature data. The last layer uses a tanh activation layer, and the final output layer outputs the generated data G(Z).

[0019] Optionally, a simplified total variation loss function in the generator loss function of S200.

[0020] Where i and j are the pixel coordinates of the generated data;

[0021] The loss function V of the original generative network model G′ Generator loss function

[0022] Optionally, the discriminative network model structure of S300 is as follows: the input layer takes generated data G(Z) and real data X as input, and extracts data features through several convolutional layers, normalization layers and S_ReLU activation layers. The output layer outputs the probability that its input is real data through a fully connected layer.

[0023] Optionally, the gradient penalty term in the discriminator loss function of S300.

[0024] Where ε is the coefficient term, λ∈(0,1);

[0025] The loss function V of the original discriminant network modelD′ Discriminator loss function

[0026] Optionally, training the S400 network model includes:

[0027] S410, Randomly select noise samples {z} from the noise distribution. (1) ,z (2) ,...,z (m)};

[0028] S420, Select training data {x} (1) ,x (2) ,...,x (m)};

[0029] S430, train the discriminant network model D, maximize the discriminant loss function, and calculate the discriminant loss function with respect to the discriminant network model parameters θ. d gradient t d Then use θ d_new =θ d +t d Update parameter θ d ,in

[0030] S440, train the generator network model G, minimize the generator loss function, and calculate the generator loss function with respect to the generator network model parameters θ. g gradient t g Then use θ g_new =θ g -t g Update parameter θ g ,in

[0031] In summary, this application includes the following beneficial technical effects:

[0032] This application expands the eVTOL runway database through sample generation, improving the training of the eVTOL runway recognition algorithm and increasing the recognition rate and confidence. Furthermore, the sample generation algorithm designed in this invention solves the problems of gradient vanishing and gradient exploding during the training process, generating samples with less noise, clearer images, and more detailed images, further improving the runway recognition rate and confidence. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a flowchart illustrating the process of this application;

[0035] Figure 2 This is a diagram of the generated network structure in a specific embodiment of this application;

[0036] Figure 3 This is a diagram of the discrimination network structure in a specific embodiment of this application;

[0037] Figure 4 This is a rendering of the generated runway according to a specific implementation of this application. Detailed Implementation

[0038] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0039] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0040] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this application, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0041] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0042] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0043] This application provides a method for expanding an eVTOL runway database.

[0044] like Figure 1 As shown, an eVTOL runway database expansion method is characterized by the following steps:

[0045] S100, design the activation function S_ReLU function, and introduce a generative adversarial model;

[0046] S200, construct the generator network model G, and design the generator loss function V. G =V G′ +V T ;

[0047] Among them, V G To generate the loss function of the network model, V G′ V is the loss function of the original generative network model. T For a simplified total variation loss function;

[0048] S300, construct the discriminant network model D, and design the discriminant loss function.

[0049] Among them, V D To determine the loss function of the network model, V D′ The loss function of the original discriminant network model. This is a gradient penalty term;

[0050] S400: Train the network model by training the generative network model and the discriminative network model separately and alternately, fixing one network model while optimizing the other.

[0051] S500 generates the required samples by randomly selecting noise samples z from the noise distribution and inputting them into the generative network model G trained by S400 to obtain the required samples.

[0052] The S_ReLU function in S100 is:

[0053] When x > 0, S_ReLU = ξx;

[0054] When x≤0, S_ReLU=η(e x -1).

[0055] The S200 generative network model structure is as follows: the input layer takes noise Z as input, which passes through a fully connected layer, several deconvolutional layers, a normalization layer, and an S_ReLU activation layer to generate feature data. The last layer uses a tanh activation layer, and the final output layer outputs the generated data G(Z).

[0056] Simplified total variation loss function in the generator loss function of S200

[0057] Where i and j are the pixel coordinates of the generated data;

[0058] The loss function V of the original generative network model G′ Generator loss function

[0059] The discriminant network model structure of S300 is as follows: The input layer takes generated data G(Z) and real data X as input, and extracts data features through several convolutional layers, normalization layers and S_ReLU activation layers. The output layer outputs the probability that its input is real data through a fully connected layer.

[0060] Gradient penalty term in the discriminator loss function of S300

[0061] Where ε is the coefficient term, λ∈(0,1);

[0062] The loss function V of the original discriminant network model D′ Discriminator loss function

[0063] S400 network model training includes:

[0064] S410, Randomly select noise samples {z} from the noise distribution. (1) ,z (2) ,...,z (m)};

[0065] S420, Select training data {x} (1) ,x (2) ,...,x (m)};

[0066] S430, train the discriminant network model D, maximize the discriminant loss function, and calculate the discriminant loss function with respect to the discriminant network model parameters θ. d gradient t d Then use θ d_new =θ d +t d Update parameter θ d ,in

[0067] S440, train the generator network model G, minimize the generator loss function, and calculate the generator loss function with respect to the generator network model parameters θ. g gradient t g Then use θ g_new =θ g -t g Update parameter θ g ,in

[0068] In one embodiment, such as Figure 2 and Figure 3 As shown, an eVTOL runway database expansion method includes the following steps:

[0069] S100, Design the activation function. Design the S_ReLU function as shown in the following formula as the activation function of the discriminant network D.

[0070]

[0071] Where ξ and η are coefficient terms, for eVTOL runway images, as a preferred choice, ξ = 1 and η = 0.63 are selected.

[0072] S200, construct the generator network G, design the generator loss function as shown in the following equation, and construct as follows: Figure 2 The generator network G shown has an input layer that takes noise Z as input. The noise Z passes through a fully connected layer, a deconvolutional layer, a normalization layer, and an S-ReLU activation layer to generate feature data. The last layer uses a tanh activation layer, and the output layer outputs the generated data G(Z).

[0073]

[0074] Among them, V G V represents the loss function of the generative model. G′ V represents the loss function of the original generative model. T This represents the simplified total variation loss function.

[0075] Furthermore, the loss function V of the original generative model G′ It enables generative models to learn the data distribution of real data as much as possible, thereby generating data that conforms to the real data distribution.

[0076]

[0077] Simplified total variation loss function V T It can not only reduce blockiness in the generated data but also reduce data noise, making the generated images clearer and ensuring the detail description of the images.

[0078]

[0079] Where i and j represent the pixel coordinates of the generated data.

[0080] S300, construct the discriminant network D, design the discriminant loss function as shown in the following formula, and construct as follows: Figure 3 The discriminant network D structure shown takes generated data G(Z) and real data X as input layers. After passing through convolutional layers, normalization layers, and S_ReLU activation layers to extract data features, and then through fully connected layers, the output layer outputs the probability that its input is real data.

[0081]

[0082] Among them, V D V represents the loss function of the discriminative model. D′ The loss function of the original discriminant model. ε and λ are the gradient penalty terms, and ε and λ are the coefficient terms. For eVTOL runway images, ε = 0.85 and λ = 0.5 are preferred.

[0083] The loss function V of the original discriminant model D′ It enables the discriminative model to accurately distinguish between generated data and real data. In adversarial training, the generative model will also learn the data distribution of real data as much as possible to generate data that can deceive the discriminator.

[0084]

[0085] Gradient penalty term It can alleviate the problems of gradient explosion and gradient vanishing that may occur during the training of the discrimination network, and plays an important role in improving the stability of network training.

[0086]

[0087] Where ε is the coefficient term, and y=λx+(1-λ)G(z) represents the linear uniform sampling between the real data sample X and the generated data sample G(Z), λ∈(0,1).

[0088] S400, training the network involves training the generative and discriminative models separately and alternately. The generative network G is fixed while the discriminative network D is trained; conversely, the discriminative network D is fixed while the generative network G is trained, and this process is repeated. The specific steps include:

[0089] S411, randomly select 50 noise samples {z} from the noise distribution. (1) ,z (2) ,...,z (50)};

[0090] S412, select 50 runway images as training data {x (1) ,x (2) ,...,x (50)};

[0091] S413, train the discriminant network D by inputting the training data X and generated samples G(Z) into the discriminant network D. Maximize the discriminant loss function, fix the generator network G, and calculate the discriminant loss function with respect to the discriminant model parameters θ. d gradient t d Then use θ d_new =θ d +t d Update parameter θ d ;

[0092]

[0093] S414, train the generator network G by inputting noisy samples into G to obtain generated samples G(Z). Minimize the generator loss function, fix the discriminator network D, and calculate the generator loss function with respect to the generator model parameters θ. g gradient t g Then use θ g_new =θ g -t g Update parameter θ g ;

[0094]

[0095] S500 generates the required samples by randomly selecting noise samples z from the noise distribution and inputting them into the generator network G trained by S400 to obtain the runway samples. Figure 4 The image shown is a sample image of the generated runway.

[0096] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for expanding an eVTOL runway database, characterized in that, Includes the following steps: S100, design the activation function S_ReLU function, and introduce a generative adversarial model; S200, construct the generator network model G, and design the generator loss function. ; in, To generate the loss function for the network model, The loss function of the original generative network model. For a simplified total variation loss function; S300, construct the discriminant network model D, and design the discriminant loss function. ; in, To determine the loss function of the network model, The loss function of the original discriminant network model. This is a gradient penalty term; S400: Train the network model by training the generative network model and the discriminative network model separately and alternately, fixing one network model while optimizing the other. S500 generates the required samples by randomly selecting noise samples z from the noise distribution and inputting them into the generative network model G trained by S400 to obtain the required samples. The S200 generator network model structure is as follows: The input layer takes noise Z, which passes through a fully connected layer, several deconvolutional layers, a normalization layer, and an S_ReLU activation layer to generate feature data. The final layer uses a tanh activation layer, and the output layer outputs the generated data G(Z). The generator loss function in S200 is a simplified total variation loss function. ; Where i and j are the pixel coordinates of the generated data; The loss function of the original generative network model Generator loss function ; The S300 discriminator network model structure is as follows: The input layer takes generated data G(Z) and real data X as input, passes through several convolutional layers, normalization layers, and S-ReLU activation layers to extract data features, and then passes through fully connected layers. The output layer outputs the probability that its input is real data. The gradient penalty term in the S300 discriminator loss function... ; in For coefficient terms, ; The loss function of the original discriminant network model Discriminator loss function .

2. The eVTOL runway database expansion method according to claim 1, characterized in that, The S_ReLU function in S100 is: when hour, ; when hour, .

3. The eVTOL runway database expansion method according to claim 1, characterized in that, S400 network model training includes: S410, Randomly select noise samples from the noise distribution. ; S420, Select training data ; S430, train the discriminant network model D, maximize the discriminator loss function, and calculate the discriminator loss function with respect to the discriminant network model parameters. gradient Then use Update parameters ,in ; S440, train the generative network model G, minimize the generator loss function, and calculate the generator loss function with respect to the parameters of the generative network model. gradient Then use Update parameters ,in .

Images