Aircraft wing skin corrosion depth prediction method based on GAN and improved LightGBM algorithm

Abstract

The invention provides an aircraft wing skin corrosion depth prediction method based on a GAN and an improved LightGBM algorithm. The aircraft wing skin corrosion depth prediction method comprises the following steps: S1, data enhancement: extracting feature information between columns by using four fully connected layers; s2, improving a condition generator and a discriminator, and updating parameters; and S3, the LightGBM algorithm is improved, and the corrosion depth of the aircraft wing skin is predicted. According to the aircraft wing skin corrosion depth prediction method, firstly, the generative adversarial network is used for data enhancement, then the gradient regression model is used for sample regression, and small sample data can be accurately predicted. Meanwhile, the gradient regression model LightGBM algorithm is improved, the method is more suitable for predicting the corrosion depth of the aircraft wing skin, and the prediction precision of predicting the corrosion depth of the aircraft wing skin can be greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of aircraft wing skin corrosion prediction, and in particular to a method for predicting aircraft wing skin corrosion depth based on GAN and an improved LightGBM algorithm. Background Art

[0002] During aircraft service, wing skins are exposed to high atmospheric humidity for extended periods, making metal components susceptible to corrosion. This can lead to dangerous events such as component failure, reduced load-bearing area, and reduced performance, threatening flight safety and increasing aircraft maintenance costs. Therefore, predicting the corrosion behavior and corrosion loss of metals under specific time and conditions has considerable application value in the aviation industry.

[0003] Machine learning models have a natural advantage in predicting metal corrosion behavior in complex atmospheric environments due to their powerful nonlinear fitting and data learning capabilities. That is, some environmental parameters that are more obvious to corrosion behavior, as well as outdoor exposure time, are used as model inputs, and the amount of metal corrosion loss under the above conditions is used as output. The machine learning model is designed as a regression problem under supervised learning. However, unlike the application research of machine learning in other fields, when predicting metal corrosion loss, the metal corrosion dataset used to train the model needs to be obtained through a large number of outdoor exposure experiments. That is, the dataset Acquisition is costly and difficult, and the model's transferability is poor. Therefore, a lack of data has become a significant factor limiting model prediction accuracy. Data augmentation technology can effectively address this data shortage. Data augmentation, also known as data expansion, relies on using specific algorithms or strategies to make existing data generate the equivalent value of more data without substantially increasing the amount of data. Data augmentation can effectively improve the accuracy of machine learning models within data-constrained environments. Specifically, designing a machine learning regression model based on data augmentation strategies can significantly improve the accuracy of the original machine learning algorithm in predicting corrosion loss of wing skins in atmospheric environments during service. Summary of the Invention

[0004] In order to address the deficiencies of the above-mentioned prior art, the purpose of the present invention is to provide a method for predicting the corrosion depth of aircraft wing skin based on GAN and an improved LightGBM algorithm. The method first uses a generative adversarial network for data enhancement and then uses a gradient regression model for sample regression.

[0005] Specifically, on the one hand, the present invention provides a method for predicting the corrosion depth of aircraft wing skin based on GAN and improved LightGBM algorithm, which includes the following steps: S1. Data preprocessing: Collect aircraft skin atmospheric corrosion environment data, normalize the continuous data, and map the value range to the [-1, 1] interval; S2. Construct a generative adversarial network to enhance the aircraft skin atmospheric corrosion environment data and obtain an enhanced dataset. S3. Improve the LightGBM algorithm and use the enhanced dataset to predict the corrosion depth of aircraft wing skin. Specifically: S31. Search weight center point : Use the K-Means clustering algorithm to divide the output of LightGBM into n subsets, and find the cluster center of each subset respectively , the weight center point of column i The calculation formula is: ; S32. Calculate threshold , the threshold value of the output of column i The calculation formula is: ; in, is the average value of the weight of column i; S33. Generate weight vector and weight matrix of predicted values: For a set of predicted values , weight matrix The calculation formula for the weight vector of the i-th row and j-th column is: ; in, is the weight matrix The weight vector of row i and column j of is the predicted value of the data in row i and column j, and are all model hyperparameters, , , is the total number of rows in the prediction result data table, The total number of columns in the prediction result data table; S34. Calculate the predicted corrosion depth of aircraft wing skin , The calculation method is as follows: ; Among them, Y is the predicted value output by the initial LightGBM, and T is the matrix transpose symbol.

[0006] Preferably, in step S2, the generative adversarial network includes a conditional generator and a discriminator, the conditional generator includes four fully connected layers, and the discriminator includes two fully connected layers. The scalar of continuous data is generated by the hyperbolic tangent function tanh through the fully connected layers of the generator, the parameters of the generator and the discriminator are initialized, and the training hyperparameters are set; in the training iteration, the parameters of the discriminator and the generator are alternately updated, and the feature information between the columns of the aircraft skin sample data is extracted as enhanced output data and a training set is constructed; Preferably, in step S2, during the training iteration, the specific steps of updating the discriminator parameters are: Sampling and generation: Sample a batch of random noise vectors from the noise distribution, input them into the generator to generate a batch of virtual data samples, and at the same time, obtain a batch of real data samples; Calculate loss: Input real data and fake data into the discriminator respectively, and use the binary cross entropy loss function to calculate the discriminator's loss for real data and fake data. The discriminator's goal is to maximize the ability to correctly classify real data and fake data, that is, to make the discriminator output close to 1 for real data and close to 0 for fake data; Backpropagation update: Based on the calculated loss, the backpropagation algorithm is used to calculate the gradient of the discriminator parameters with respect to the loss function, and the discriminator parameters are updated according to the specified learning rate.

[0007] Preferably, in step S2, during the training iteration, the specific steps of updating the generator parameters are: Re-sampling and regeneration: resample a batch of random noise vectors from the noise distribution and generate a batch of new virtual data samples through the generator; Calculate the loss: The generated virtual data is input into the discriminator. The goal of the generator is to make the discriminator misjudge the generated virtual data as real data, that is, to make the discriminator output close to 1 for the generated virtual data. Based on this goal, the loss of the discriminator for the generated virtual data is calculated; Backpropagation update: Based on the calculated loss, the gradient of the generator parameters with respect to the loss function is calculated through the backpropagation algorithm, and the parameters of the generator are updated according to the specified learning rate.

[0008] Preferably, the first fully connected layer and the second fully connected layer of the conditional generator use the ReLU activation function, the third fully connected layer and the fourth fully connected layer use the LeakyReLU activation function, and the fourth fully connected layer performs Dropout processing.

[0009] Preferably, the four fully connected layers of the conditional generator are processed using batch normalization operations.

[0010] Preferably, the first fully connected layer and the second fully connected layer of the discriminator use the ReLU activation function, and the number of neurons is 128; the number of neurons in the third fully connected layer and the fourth fully connected layer of the conditional generator is 512, and the Dropout processing of the fourth fully connected layer of the conditional generator inactivates the neurons with a probability of 0.3.

[0011] Preferably, the improved condition generator is described by the following formula: ; in, is the input layer, are the first, second, third, and fourth layers of the generator model, respectively; z is a random noise vector; ReLU is the ReLU activation function; BN is the batch normalization operation; FC is the fully connected layer; drop is the dropout regularization technique; LeakyReLUn is the LeakyReLU activation function, n is the retention probability of neurons in the neural network; gumbel_softmax is the discrete distribution reparameterization operation; is the specific numerical scalar of the i-th data in a row; is the one-hot vector of the i-th data in a row.

[0012] Preferably, the improved discriminator is described by the following formula: ; Among them, r is the sample data input to the discriminator, is the output of the discriminator.

[0013] Preferably, the ReLU activation function expression is f(x)=max(0,x), and the LeakyReLU activation function expression is: ; in, is a positive integer less than 1.

[0014] Compared with the prior art, the present invention has the following beneficial effects: The present invention provides a method for predicting the corrosion depth of aircraft wing skin based on GAN and an improved LightGBM algorithm. The method first uses an improved generative adversarial network for data enhancement, and then uses a gradient regression model to perform sample regression to predict the corrosion depth of aircraft wing skin. The prediction accuracy can also be guaranteed in the case of small samples.

[0015] The generative adversarial network of the present invention constructs a four-layer fully connected network. Compared with the traditional two-layer structure, it can more deeply explore the potential feature information between data columns and effectively capture the complex spatiotemporal dependencies and nonlinear interactions between variables in atmospheric data, such as the mutual influence between indicators such as temperature, humidity, and pollutant concentration, thereby improving the extraction accuracy and completeness of data features.

[0016] The present invention improves the gradient regression model LightGBM algorithm, which is more suitable for the prediction of aircraft wing skin corrosion depth and can greatly improve the prediction accuracy of aircraft wing skin corrosion depth prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is the overall flow chart of the present invention; Figure 2 It is a schematic diagram of the workflow of the present invention; Figure 3 This is a schematic diagram of the LightGBM algorithm model in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0018] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

[0019] The present invention provides a method for predicting the corrosion depth of aircraft wing skin based on GAN and improved LightGBM algorithm. Figure 1 and Figure 2 As shown, it includes the following steps: S1. Data preprocessing: Aircraft skin atmospheric corrosion environment data are collected, the continuous data are normalized, and the value range is mapped to the [-1, 1] interval to obtain the aircraft skin atmospheric corrosion environment dataset. Each dataset contains five input attributes: average temperature, average relative humidity, annual precipitation, sunshine hours, and average air pressure.

[0020] S2. Construct a generative adversarial network to enhance the aircraft skin atmospheric corrosion environment data and obtain an enhanced dataset. The generative adversarial network includes a conditional generator and a discriminator. The conditional generator includes four fully connected layers, and the discriminator includes two fully connected layers. The hyperbolic tangent function tanh is used to generate scalars of continuous data through the fully connected layers of the generator; the model vector of discrete data is generated by the reparameterization method of discrete random variables, the parameters of the generator and discriminator are initialized, and the training hyperparameters are set; in the training iteration, the parameters of the discriminator and generator are updated alternately, the feature information between columns is extracted as enhanced output data, and a training set is constructed.

[0021] The discriminator parameter update process is: Sampling and generation: Sample a batch of random noise vectors from the noise distribution and input them into the generator to generate a batch of virtual data samples. At the same time, obtain a batch of real table data samples.

[0022] Calculating loss: Real data and fake data are fed into the discriminator, and the binary cross-entropy loss function is used to calculate the discriminator's loss for real data and fake data. The discriminator's goal is to maximize its ability to correctly classify real data and fake data, that is, to make the discriminator output close to 1 for real data and close to 0 for fake data.

[0023] Back propagation update: Based on the calculated loss, the back propagation algorithm is used to calculate the gradient of the discriminator parameters with respect to the loss function, and the discriminator parameters are updated according to a certain learning rate.

[0024] The generator parameter update process is: Re-sampling and generation: Resample a batch of random noise vectors from the noise distribution and generate a batch of new virtual data samples through the generator.

[0025] Calculate the loss: The generated virtual data is input into the discriminator. The goal of the generator is to make the discriminator misjudge the generated virtual data as real data, that is, to make the discriminator output close to 1 for the generated virtual data. Based on this, calculate the loss of the discriminator for the generated virtual data.

[0026] Backpropagation update: Based on the loss, the gradient of the generator parameters with respect to the loss function is calculated through the backpropagation algorithm, and the parameters of the generator are updated according to a certain learning rate.

[0027] By continuously repeating the above two steps and alternately updating the parameters of the discriminator and generator, the generator gradually generates more realistic tabular data, and the discriminator also improves its discrimination ability accordingly, eventually reaching a balance point, that is, the discriminator cannot accurately distinguish between the generated data and the real data.

[0028] The first and second fully connected layers of the conditional generator use the ReLU activation function, while the third and fourth fully connected layers use the LeakyReLU activation function. Dropout is also applied to the fourth fully connected layer. Batch normalization is used for all four fully connected layers of the conditional generator. The first and second fully connected layers of the discriminator use the ReLU activation function and have 128 neurons. The third and fourth fully connected layers of the conditional generator have 512 neurons, and Dropout in the fourth fully connected layer inactivates neurons with a probability of 0.3.

[0029] In a specific embodiment, the ReLU activation function expression is f(x)=max(0,x), and the LeakyReLU activation function expression is: ; in, is a positive integer less than 1.

[0030] Specifically, the generator uses four fully-connected layers to extract column-to-column feature information. The first two fully-connected layers use the Relu activation function. To improve the generator's output accuracy, the third and fourth fully-connected layers use the LeakyReLU activation function and increase the number of neurons to 512. To prevent overfitting, the fourth layer performs dropout. Dropout is a regularization technique used in neural network training. It inactivates a subset of neurons according to Gaussian probability, reducing complex co-adaptation between neurons and forcing each neuron in the network to learn useful features independently rather than relying on a specific combination of the remaining neurons. All four fully-connected layers are processed using batch normalization (BN). Scalars are generated using the hyperbolic tangent function, tanh, which constrains neuron outputs to the range -1 to 1, providing better gradients. Model vectors are generated using a reparameterized Gumbel-Softmax algorithm for discrete random variables. GumbelSoftmax simulates the effect of directly drawing samples from a discrete distribution by adding samples from the Gumble distribution to the log-probability of the discrete variable and then applying the Softmax function. This is very suitable for generating discrete data. Finally, the generation layer of the one-hot vector of the discrete value column in the original framework is removed. The improved conditional generator is described by the following formula: ; in, is the input layer, are the first, second, third, and fourth layers of the generator model, respectively; z is a random noise vector; ReLU is the ReLU activation function; BN is the batch normalization operation; FC is the fully connected layer; drop is the dropout regularization technique; LeakyReLUn is the LeakyReLU activation function, n is the retention probability of neurons in the neural network; gumbel_softmax is the discrete distribution reparameterization operation; is the specific numerical scalar of the i-th data in a row; is the one-hot vector of the i-th data in a row.

[0031] In this dataset, the generator complexity is higher than that of the traditional CTGAN framework. Furthermore, the data size is even smaller in small sample scenarios, making the original PacGAN framework redundant and requiring further improvement. Specifically, the Pac value is changed from 10 to 2 to accelerate the convergence of CTGAN. Similar to the improved generator, the discriminator also removes the layer that generates one-hot vectors for discrete-valued columns in the original framework. Furthermore, compared to the original discriminator structure, the improved discriminator still consists of a two-layer fully connected neural network, but the activation function in the second layer is replaced by ReLU instead of LeakyReLU, reducing the number of neurons from 256 to 128.

[0032] The improved discriminator is described as follows: ; Among them, r is the sample data input to the discriminator, is the output of the discriminator.

[0033] S3. Improve the LightGBM algorithm and predict the corrosion depth of aircraft wing skin. The LightGBM algorithm model is shown in the figure below. Figure 3 This step specifically includes the following sub-steps: S31. Search weight center point : Use the K-Means clustering algorithm to divide the output of LightGBM into n subsets, and find the cluster center of each subset respectively , the weight center point of column i The calculation formula is: ; S32. Calculate threshold , the threshold value of the output of column i The calculation formula is: ; in, is the average value of the weight of column i; S33. Generate weight vector and weight matrix of predicted values: For a set of predicted values , weight matrix The calculation formula for the weight vector of the i-th row and j-th column is: ; in, is the weight matrix The weight vector of row i and column j of is the predicted value of the data in row i and column j, and are all model hyperparameters, , , is the total number of rows in the prediction result data table, The total number of columns in the prediction result data table; S34. Calculate the predicted corrosion depth of aircraft wing skin , The calculation method is as follows: ; Among them, Y is the predicted value output by the initial LightGBM, and T is the matrix transpose symbol. Specific embodiments S1. Data preprocessing: Collect aircraft skin atmospheric corrosion environment data, normalize the continuous data, and map the value range to the [-1,1] interval.

[0035] This example uses the steel atmospheric corrosion database from the China National Materials Corrosion and Protection Data Center. This database contains atmospheric corrosion test results for three types of steel: Q235, 09Cu, and 10Cr, across 331 regions in China. Each dataset contains five input attributes: average temperature, average relative humidity, annual precipitation, sunshine hours, and average air pressure. The output is two attributes: the loss of each type of steel within one year and the loss within four years (the loss data was measured using an ultrasonic corrosion thickness gauge during the experiment). This totals six attributes, or a vector of 6 dimensions.

[0036] The following table contains selected environmental data from 15 experimental stations in the dataset: Table 1 shows the atmospheric environment data of some test stations S2. Construct a generative adversarial network to perform data augmentation on the atmospheric corrosion environment data of aircraft skins to obtain an enhanced dataset. The original training set is put into the improved CTGAN network to expand the data to five times the original training set. After convergence, it is used to make predictions on the test set.

[0037] S3. The enhanced test set is used to predict the corrosion depth of the aircraft wing skin. The prediction results at the above 15 sites are shown in the following table (taking Q235 steel as an example): Table 2. The true label values ​​and model output prediction values ​​of Q235 steel corrosion thickness at each test site in the dataset (unit: μm) The embodiments described above are merely descriptions of preferred implementations of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A method for predicting aircraft wing skin corrosion depth based on GAN and improved LightGBM algorithm, characterized by: It includes the following steps: S1. Data preprocessing: Collect aircraft skin atmospheric corrosion environment data, normalize the continuous data, and map the value range to the [-1, 1] interval; S2. Construct a generative adversarial network to enhance the aircraft skin atmospheric corrosion environment data and obtain an enhanced dataset. S3. Improve the LightGBM algorithm and use the enhanced dataset to predict the corrosion depth of aircraft wing skin. Specifically: S31. Search weight center point : Use the K-Means clustering algorithm to divide the output of LightGBM into n subsets, and find the cluster center of each subset respectively , the weight center point of column i The calculation formula is: ； S32. Calculate threshold , the threshold value of the output of column i The calculation formula is: ； in, is the average value of the weight of column i; S33. Generate weight vector and weight matrix of predicted values: For a set of predicted values , weight matrix The calculation formula for the weight vector of the i-th row and j-th column is: ； in, is the weight matrix The weight vector of row i and column j of is the predicted value of the data in row i and column j, and are all model hyperparameters, , , is the total number of rows in the prediction result data table, The total number of columns in the prediction result data table; S34. Calculate the predicted corrosion depth of aircraft wing skin , The calculation method is as follows: ； Among them, Y is the predicted value output by the initial LightGBM, and T is the matrix transpose symbol.

2. The aircraft wing skin corrosion depth prediction method based on GAN and improved LightGBM algorithm according to claim 1 is characterized by: In step S2, the generative adversarial network includes a conditional generator and a discriminator. The conditional generator includes four fully connected layers, and the discriminator includes two fully connected layers. The hyperbolic tangent function tanh is used to generate a scalar of continuous data through the fully connected layers of the generator, initialize the parameters of the generator and discriminator, and set the training hyperparameters. During the training iterations, the parameters of the discriminator and generator are updated alternately, and the feature information between columns of aircraft skin sample data is extracted as enhanced output data and a training set is constructed.

3. The aircraft wing skin corrosion depth prediction method based on GAN and improved LightGBM algorithm according to claim 2 is characterized by: In step S2, during the training iteration, the specific steps for updating the discriminator parameters are: Sampling and generation: Sample a batch of random noise vectors from the noise distribution, input them into the generator to generate a batch of virtual data samples, and at the same time, obtain a batch of real data samples; Calculate loss: Input real data and fake data into the discriminator respectively, and use the binary cross entropy loss function to calculate the discriminator's loss for real data and fake data. The discriminator's goal is to maximize the ability to correctly classify real data and fake data, that is, to make the discriminator output close to 1 for real data and close to 0 for fake data; Backpropagation update: Based on the calculated loss, the backpropagation algorithm is used to calculate the gradient of the discriminator parameters with respect to the loss function, and the discriminator parameters are updated according to the specified learning rate.

4. The aircraft wing skin corrosion depth prediction method based on GAN and improved LightGBM algorithm according to claim 2 is characterized by: In step S2, during the training iteration, the specific steps for updating the generator parameters are: Re-sampling and regeneration: resample a batch of random noise vectors from the noise distribution and generate a batch of new virtual data samples through the generator; Calculate the loss: The generated virtual data is input into the discriminator. The goal of the generator is to make the discriminator misjudge the generated virtual data as real data, that is, to make the discriminator output close to 1 for the generated virtual data. Based on this goal, the loss of the discriminator for the generated virtual data is calculated; Backpropagation update: Based on the calculated loss, the gradient of the generator parameters with respect to the loss function is calculated through the backpropagation algorithm, and the parameters of the generator are updated according to the specified learning rate.

5. The aircraft wing skin corrosion depth prediction method based on GAN and improved LightGBM algorithm according to claim 1 is characterized by: The first and second fully connected layers of the conditional generator use the ReLU activation function, the third and fourth fully connected layers use the LeakyReLU activation function, and the fourth fully connected layer performs Dropout processing.

6. The aircraft wing skin corrosion depth prediction method based on GAN and improved LightGBM algorithm according to claim 2 is characterized by: The four fully connected layers of the conditional generator are all processed using batch normalization operations.

7. The aircraft wing skin corrosion depth prediction method based on GAN and improved LightGBM algorithm according to claim 1 is characterized by: The first and second fully connected layers of the discriminator use the ReLU activation function, and the number of neurons is 128; the number of neurons in the third and fourth fully connected layers of the conditional generator is 512, and the Dropout processing of the fourth fully connected layer of the conditional generator inactivates the neurons with a probability of 0.

3.

8. The aircraft wing skin corrosion depth prediction method based on GAN and improved LightGBM algorithm according to claim 2 is characterized by: The improved condition generator is described as follows: ； in, is the input layer, are the first, second, third, and fourth layers of the generator model, respectively; z is a random noise vector; ReLU is the ReLU activation function; BN is the batch normalization operation; FC is the fully connected layer; drop is the dropout regularization technique; LeakyReLUn is the LeakyReLU activation function, n is the retention probability of neurons in the neural network; gumbel_softmax is the discrete distribution reparameterization operation; is the specific numerical scalar of the i-th data in a row; is the one-hot vector of the i-th data in a row.

9. The aircraft wing skin corrosion depth prediction method based on GAN and improved LightGBM algorithm according to claim 2 is characterized by: The improved discriminator is described as follows: ； Among them, r is the sample data input to the discriminator, is the output of the discriminator.

10. The aircraft wing skin corrosion depth prediction method based on GAN and improved LightGBM algorithm according to claim 7 is characterized by: The ReLU activation function expression is f(x)=max(0,x), and the LeakyReLU activation function expression is: ； in, is a positive integer less than 1.

Images