Polar wind turbine airfoil ice shape data enhancement method based on improved generative adversarial network

Abstract

This invention provides a polar wind-powered airfoil ice shape data augmentation method based on an improved generative adversarial network, comprising the following steps: S1, obtaining a raw high-fidelity icing dataset of the polar environment using icing numerical simulation software, performing solid mask conversion on the airfoil icing images in the raw high-fidelity icing dataset, and normalizing the airfoil icing images and icing condition parameters to construct a training sample set; S2, constructing a CIN-SAM-GAN high-quality ice shape data augmentation model, and performing adversarial training on the CIN-SAM-GAN high-quality ice shape data augmentation model using the training sample set, outputting a trained high-quality ice shape data augmentation model; S3, inputting a massive amount of polar icing conditions into the trained high-quality ice shape data augmentation model to construct a large-scale icing augmentation dataset. This invention can generate complex polar icing samples with high physical fidelity at extremely low cost and high efficiency, significantly improving the generalization ability of downstream ice shape prediction models.

Description

Technical Field

[0001] This invention relates to the field of airfoil icing technology, and more particularly to a method for enhancing polar wind turbine airfoil icing data based on an improved generative adversarial network. Background Technology

[0002] Wind turbine blades operating in polar and high-altitude cold regions are highly susceptible to icing. Airfoil icing alters the original aerodynamic shape of the blades, leading to decreased lift, increased drag, and in severe cases, structural fatigue and operational safety accidents. Data-driven intelligent prediction methods have received widespread attention in recent years for rapidly predicting blade icing patterns. These methods rely on a large number of high-quality ice shape samples as training data, establishing a mapping relationship between icing conditions and ice shapes through deep learning models. However, obtaining sufficient high-quality ice shape samples in practical engineering faces numerous challenges, making data augmentation techniques a key means to alleviate the data scarcity problem.

[0003] Currently, obtaining airfoil icing shape data mainly relies on two technical approaches. The first is icing wind tunnel experiments, which simulate real icing conditions to obtain airfoil icing image data. Due to the high experimental costs and limitations of this method, some conditions that lead to severe icing cannot be achieved, such as extremely high wind speeds and very low ambient temperatures. The second approach is traditional numerical simulation methods, using icing software based on computational fluid dynamics to generate high-fidelity ice shapes by solving aerodynamic and icing thermodynamic equations. This method offers high accuracy, but a single simulation takes tens of minutes to several hours, consuming enormous computational resources and making it difficult to cover the massive combinations of operating conditions in polar environments.

[0004] The aforementioned existing technologies still have significant shortcomings in practical applications. First, traditional numerical simulation methods are computationally expensive and cannot meet the data production needs under massive operating conditions, resulting in a severe shortage of high-quality ice shape data under extreme conditions. Second, under polar conditions, blade leading edges are prone to forming complex ice shapes with sharp ice corners. Existing generative adversarial networks, when dealing with such microscopic geometric abrupt changes, are limited by feature extraction capabilities and normalization mechanisms, resulting in widespread high-frequency noise and artifacts in the generated results. Ice corner boundaries are over-smoothed, exhibiting a rounded and blunt distortion, and the physical fidelity of the samples is difficult to meet engineering requirements. Third, due to the extremely limited number of real ice shape samples, directly using the original dataset to train downstream prediction models is prone to overfitting, resulting in poor model generalization ability and an inability to achieve high-precision ice shape prediction in practical engineering. Therefore, how to efficiently generate high-fidelity polar complex icing data based on a small number of real samples has become a core bottleneck restricting the development of this field. Summary of the Invention

[0005] In view of this, the purpose of this invention is to propose a polar wind airfoil ice shape data augmentation method based on an improved generative adversarial network, so as to solve the technical problems of scarcity of high-quality ice shape data under polar conditions, severe distortion of traditional generative models in handling complex ice corners, and easy overfitting of downstream prediction models in the existing technology.

[0006] The technical means employed in this invention are as follows:

[0007] A method for augmenting polar wind-driven airfoil ice shape data based on an improved generative adversarial network includes the following steps: S1. Obtain the original high-fidelity icing dataset of the polar environment through icing numerical simulation software, perform solid mask conversion on the airfoil icing images in the original high-fidelity icing dataset, and normalize the airfoil icing images and icing condition parameters to construct a training sample set. S2. Construct a high-quality ice shape data augmentation model using CIN-SAM-GAN, and conduct adversarial training on the CIN-SAM-GAN high-quality ice shape data augmentation model using the training sample set, and output the high-quality ice shape data augmentation model after training. S3. Combine Latin hypercube sampling to generate a massive amount of polar icing conditions. Input the massive amount of polar icing conditions into the trained high-quality ice shape data augmentation model to generate corresponding high-quality ice shape samples in batches, and construct a large-scale icing augmentation dataset.

[0008] Further, in S1, the normalization process specifically includes the following steps: adjusting the solid mask image to a size of 256×256 pixels, and linearly normalizing the pixel value range to the range of -1 to 1; the icing condition parameters include ambient temperature, inflow velocity, liquid water content, median volume diameter, and icing duration; calculating the mean and standard deviation of the icing condition parameters in the entire dataset; and using the mean and standard deviation to perform Z-score standardization on the input icing condition parameters to obtain the condition vector.

[0009] Furthermore, the CIN-SAM-GAN high-quality ice-shaped data augmentation model includes a generator based on an improved U-Net architecture and a discriminator based on a PatchGAN architecture. A conditional instance normalization module is embedded in the bottleneck layer and upsampling block of the generator, and a spatial attention module is embedded before the final output layer of the generator decoder.

[0010] Furthermore, in S2, dynamic instance noise with a linearly decaying standard deviation is introduced into the image input to the discriminator during adversarial training, and the generator weights are updated using a multidimensional joint loss function that integrates adversarial loss, L1 reconstruction loss, Dice cross-union loss, and total variational smoothing loss.

[0011] Furthermore, the conditional instance normalization module performs instance normalization processing on the input feature map without affine parameters; it inputs the conditional vector into a multilayer perceptron conditional mapping network containing a ReLU activation function, outputs affine parameters of a specific dimension, and splits the affine parameters into scaling parameters along the channel dimension. and offset parameters ;Scaling parameters and offset parameters After expanding to a spatial dimension that matches the input feature map, the conditional affine transformation formula is applied. The modulated feature map is obtained.

[0012] Furthermore, the spatial attention module performs average pooling and max pooling operations on the input feature map along the channel dimension, and concatenates the results of the two operations along the channel dimension to obtain a 2-channel feature map; the 2-channel feature map is reduced to 1-channel by a convolutional layer with a kernel size of 7×7, and a 2D spatial attention mask with a value range of 0 to 1 is generated using the Sigmoid activation function; the 2D spatial attention mask is then multiplied element-wise with the original input feature map.

[0013] Furthermore, the decoder architecture of the generator specifically includes: The decoder consists of four upsampling blocks, each composed of a transposed convolutional layer with a stride of 2, a conditional instance normalization module, and a ReLU activation function. The output feature map of each upsampling block is spliced ​​with the feature map of the corresponding layer of the encoder via skip connections. The spliced ​​feature map is then passed through a convolutional layer, a batch normalization layer, and a ReLU activation function for channel compression and feature extraction. The last upsampling block of the decoder outputs a feature map of size 32×256×256. This feature map is then passed through a spatial attention module, an output convolutional layer, and a Tanh activation function to output an airfoil ice shape prediction image of size 1×256×256.

[0014] Furthermore, the specific method for adding dynamic instance noise during adversarial training is as follows: before inputting the real icing image and the fake image generated by the generator into the discriminator, Gaussian noise with a standard deviation that decreases linearly with the number of training rounds is added to both; the specific decay formula for the standard deviation of the dynamic noise is:

[0015] in, The current training round number is 100. When the training round number reaches 100, the noise standard deviation decreases to 0.

[0016] Furthermore, the formula for calculating the joint loss function during adversarial training is as follows:

[0017] in, For least-squares-based adversarial loss, For L1 pixel-level reconstruction loss, Dice loss is used to constrain complex icing topologies. This is the total variational smoothing loss used to penalize jagged edges in the generated image.

[0018] Furthermore, a large-scale icing enhancement dataset was used as training data to supervise the training of the deep learning benchmark prediction model, thereby improving the accuracy of the prediction model in predicting ice shape.

[0019] Compared with the prior art, the present invention has the following advantages: This invention requires only a small amount of original real ice shape data to efficiently and cost-effectively generate massive ice shape samples under polar conditions, solving the problem of high computational costs in traditional CFD numerical simulation.

[0020] This invention possesses extremely high physical fidelity and complex feature reconstruction capabilities. Solid Mask representation avoids high-frequency artifacts. The synergistic effect of CIN and SAM modules, coupled with the constraints of the multidimensional joint loss function, enables the generative network to accurately focus on and reconstruct the sharp boundary of the leading edge ice corner, thus solving the distortion phenomenon of the rounded ice corner caused by the basic model.

[0021] Compared to baseline models trained on limited original real datasets, this invention effectively alleviates the overfitting problem that is prone to occur in small sample training by using the augmented dataset expanded by CIN-SAM-GAN, and improves the accuracy of ice shape prediction through prediction model verification, providing reliable data support and a brand-new solution for high-precision ice shape prediction. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is an overall flowchart of the polar wind-powered airfoil ice shape data enhancement method of the present invention.

[0024] Figure 2 This is a comparative schematic diagram of the airfoil ice shape characterization method of the present invention.

[0025] Figure 3 This is a schematic diagram of the data augmentation model structure of the present invention.

[0026] Figure 4 This is a schematic diagram of the conditional instance normalization module structure of the present invention.

[0027] Figure 5 This is a schematic diagram of the spatial attention module structure of the present invention.

[0028] Figure 6 This is a visual comparison of the results of the present invention in the formation of different ice shapes.

[0029] Figure 7 This is a comparative evaluation graph of the prediction model's prediction results on the augmented dataset. Detailed Implementation

[0030] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0031] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0032] like Figure 1 As shown, this invention provides a method for augmenting polar wind airfoil ice shape data based on an improved generative adversarial network. The specific implementation steps are as follows: S1. Characterization and preprocessing of polar icing conditions and ice shape images.

[0033] First, a small amount of raw, high-fidelity icing dataset was obtained using the FENSAP-ICE icing software. For the icing condition data, five physical parameters were extracted, including ambient temperature T, inflow velocity v, liquid water content (LWC), median volume diameter (MVD), and icing duration t. The mean and standard deviation of the parameters in the entire dataset were calculated using the formula... Z-score normalization is applied to the parameter matrix to obtain a normalized conditional vector. For image data, such as... Figure 2 As shown, the real icing image and the original smooth airfoil image are uniformly converted into single-channel grayscale images. Solid Mask is applied for solid masking, that is, the interior of the ice shape and airfoil is filled with black, and the background is white. Then, the image size is adjusted to 256×256 pixels, converted into a tensor, and its pixel value range is linearly normalized to the interval [-1,1] using the standard mean=[0.5], std=[0.5], so as to match the output of the Tanh activation function at the end of the generator and construct a complete training sample set.

[0034] S2. Construct a high-quality ice-shaped data augmentation network architecture using CIN-SAM-GAN.

[0035] like Figure 3 As shown, this invention reconstructs the generator and discriminator of the standard conditional GAN ​​(Generative Adversarial Network). The model includes a generator based on an improved U-Net architecture and a discriminator based on the PatchGAN architecture. The specific network architecture of the generator is as follows: (1) Encoder: It contains 4 consecutive downsampling sequences. The input image is 1×256×256. It passes through a convolutional layer with a stride of 2 and a kernel of 4×4, a BatchNorm2d layer and a ReLU activation function in sequence. The number of channels of the feature map is expanded from 1 to 64, 128, 256 and 512 in sequence, and the size is gradually reduced to 16×16.

[0036] (2) CIN (Conditional Instance Normalization) module: such as Figure 4 As shown, the CIN module is deployed in the bottleneck layer (512 channels) of the generator and in each upsampling block of the decoder. First, instance normalization without affine parameters (InstanceNorm2d, affine=False) is applied to the input feature map. Simultaneously, the normalized 5-dimensional condition vector is input into a multilayer perceptron (MLP) consisting of two hidden layers (64 neurons each), outputting affine parameters with a dimension twice the number of feature channels. These parameters are then chunked into scaling parameters along the channel dimension using a chunking operation. and offset parameters After expanding it to a spatial dimension that matches the feature map, a conditional affine transformation is performed. This enables adaptive modulation of operating condition parameters to global characteristics.

[0037] (3) Decoder and skip connections: Contains 4 upsampling blocks. The feature map is first enlarged by transposed convolution (kernel size 4×4, stride 2, padding 1), modulated by the CIN module, and then concatenated with the features of the corresponding level of the encoder in the channel dimension (dim=1) using torch.cat. The concatenated features are then fused and reduced in channel dimension by 3×3 convolution, batch normalization and ReLU.

[0038] (4) Spatial Attention (SAM) Module: After the last upsampling block of the decoder outputs a feature map of size 32×256×256, the SAM module is connected. For example... Figure 5 As shown, this module performs average pooling and max pooling along the channel dimension, concatenating them into a 2-channel feature map. This is then reduced to a 1-channel feature map via a 7×7 convolutional layer, and a sigmoid activation is applied to the [0,1] interval to form an attention mask. The mask is multiplied back into the original 32-channel feature map, forcing the network to precisely focus on the complex ice corner region at the airfoil's leading edge. Finally, an ice-shaped image of size 1×256×256 is output through output convolution and the Tanh function.

[0039] In the adversarial training phase: CIN-SAM-GAN is trained using the training sample set, and the optimizer is Adam. The following training logic constraints are adopted in this embodiment: (1) Dynamic Instance Noise: To prevent the discriminator from overfitting in the early stages of training, causing the generator gradient to vanish, Gaussian noise with a linearly decreasing standard deviation is added to both the real and generated images before inputting them into the discriminator. The specific decay formula is as follows:

[0040] When the number of training rounds reaches 100, the noise standard deviation decays to 0. This mechanism protects the generator in the early stages to smoothly learn the global topology, and forces it to generate clean and smooth icy boundaries in the later stages.

[0041] (2) Multidimensional Joint Generator Loss Function: The generator adopts a total loss function that integrates multiple constraints, and the calculation formula is as follows:

[0042] in, For LSGAN adversarial loss based on least squares; The L1 pixel-level reconstruction loss is set to a weight of 20.0 as a safety net. The Dice crossover ratio loss is set to a weight of 10.0, which is specifically designed to prevent the complex topological shape of the ice corner from being smoothed out by mapping the output to a probability map of ice (1) and bottom (0). The total variation loss is used with a weight of 0.5. It specifically penalizes jagged edges in generated images by calculating the absolute difference between adjacent pixels in the horizontal and vertical directions of the image.

[0043] S3. Combine LHS (Latin Hypercube Sampling) to generate massive amounts of polar icing conditions, and use the trained generator to expand the batch data.

[0044] To address the diversity of polar climates, the LHS method is used to combine and sample key parameters such as temperature and wind speed, generating a massive amount of unknown polar icing conditions. These generated conditions are then batch-input into a well-trained CIN-SAM-GAN generator to efficiently generate corresponding high-quality ice shape samples. The forward propagation time for a single generation task is less than 0.02 seconds, representing an exponential improvement in efficiency compared to the tens of minutes required for a single condition in traditional CFD software.

[0045] S4. Use the expanded data as training data to conduct supervised training on the baseline prediction model in order to improve the accuracy of the prediction model in predicting ice shape.

[0046] To verify the effectiveness and engineering application value of the augmented data, a large amount of icing augmentation dataset was used as training data. The U-Net standard deep learning benchmark prediction model was trained in a supervised manner through five-fold cross-validation, and the final performance was evaluated on a test set without real samples.

[0047] In one embodiment of the present invention, the FFA-W3-211 airfoil at the tip of the IEA 15MW large wind turbine is selected as the research object. The operating parameters are designed with polar environmental conditions as the background. The FENSAP-ICE icing numerical simulation software is used to generate an original dataset (OD) under the conditions of a fixed chord length of 0.5m and a fixed angle of attack of 4.0°. The dataset contains 1,500 high-fidelity ice shape samples (1,000 simple ice shapes and 500 complex ice shapes with ice corners).

[0048] In this embodiment, the icing condition parameters are set as follows: ambient temperature (T) between -25°C and -10°C, inflow velocity (V) between 65 m / s and 125 m / s, liquid water content (LWC) between 0.2 and 1.2 g / m³, median volume diameter (MVD) between 20 and 60 μm, and icing duration (t) between 300 and 1800 s. The mean and standard deviation of the icing condition parameters in the entire dataset are calculated, and Z-score standardization is performed using a formula to obtain the condition vector.

[0049] In this embodiment, for the image data, both the real icing image and the original smooth airfoil image are uniformly converted into 256×256 single-channel grayscale images, and solid mask processing is applied, that is, the interior of the ice shape and airfoil is filled with black solids, and the background is white. This dense representation provides the network with clear spatial coherence constraints, effectively avoiding the high-frequency artifact problem caused by fine line representation from the underlying mechanism. Subsequently, the pixel value range is linearly normalized to the interval [-1, 1] using a mean of 0.5 and a standard deviation of 0.5, to construct a complete training sample set.

[0050] Icing images and corresponding icing condition parameters were divided into training, validation, and test sets. 750 samples were selected from 1500 samples to form the training set for adversarial training of the airfoil ice shape data augmentation model. 300 samples were used as the validation set to evaluate the quality of the ice shapes generated by the airfoil ice shape data augmentation model. The remaining 450 samples were used as the test set to finally evaluate the effectiveness of the generated data in the downstream prediction model.

[0051] In this embodiment, the CIN-SAM-GAN model was adversarially trained using 750 training sample sets, and ablation experiments were conducted to assess the effectiveness of components in the generator. The intersection-over-union (IoU) ratio was used as the core quantitative indicator. Each model configuration underwent three independent replicate experiments. The baseline model with traditional conditions had an average IoU of only 0.9348; after introducing global conditional injection of CIN, the average IoU increased to 0.9494. Finally, the full model integrating CIN and SAM modules (CIN-SAM-GAN) achieved an optimal average IoU of 0.9670. Figure 6 As shown, the comparison between the real ice shape and the ice shape generated by the CIN-SAM-GAN model shows that the generator not only perfectly reproduces the ice shape outline, but also has extremely high physical fidelity in complex ice corners.

[0052] In this embodiment, the original dataset (OD) is first defined as the baseline, containing only a limited 1050 real samples. Then, using the generator trained according to this invention combined with LHS sampling, 1050, 1350, 1650, and 1950 high-quality generated samples are added respectively, constructing four augmented datasets: Data1 (2100 samples), Data2 (2400 samples), Data3 (2700 samples), and Data4 (3000 samples). Pixel-level accuracy is then evaluated using F1-Score on an independent test set of 450 samples without real samples.

[0053] like Figure 7The prediction evaluation results show that, compared with the baseline model that relies only on limited real data, the prediction model trained with the augmented data of this invention exhibits extremely high prediction accuracy on different ice shapes, with F1 scores reaching 0.9907, 0.9854, and 0.9591, respectively. The prediction boundary is in high agreement with the actual CFD results. This closed-loop verification step demonstrates that the data augmentation method proposed in this invention effectively overcomes the overfitting problem caused by small samples, and has important practical significance for significantly improving the prediction accuracy of ice shape on wind turbine blades.

[0054] In summary, the present invention not only overcomes the bottleneck of capturing local features in the generation of complex ice shapes, but also effectively alleviates the overfitting problem of small sample training through high-quality data augmentation, which has important practical significance for significantly improving the generalization performance of data-driven prediction models.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for augmenting polar wind-driven airfoil ice shape data based on an improved generative adversarial network, characterized in that, Includes the following steps: S1. Obtain the original high-fidelity icing dataset of the polar environment through icing numerical simulation software, perform solid mask conversion on the airfoil icing images in the original high-fidelity icing dataset, and normalize the airfoil icing images and icing condition parameters to construct a training sample set. S2. Construct a high-quality ice shape data augmentation model using CIN-SAM-GAN, and conduct adversarial training on the CIN-SAM-GAN high-quality ice shape data augmentation model using the training sample set, and output the high-quality ice shape data augmentation model after training. S3. Combine Latin hypercube sampling to generate a massive amount of polar icing conditions. Input the massive amount of polar icing conditions into the trained high-quality ice shape data augmentation model to generate corresponding high-quality ice shape samples in batches, and construct a large-scale icing augmentation dataset.

2. The polar wind airfoil ice shape data augmentation method based on improved generative adversarial networks according to claim 1, characterized in that, In S1, the normalization process specifically includes the following steps: adjusting the solid mask image to a size of 256×256 pixels and linearly normalizing the pixel value range to the range of -1 to 1; the icing condition parameters include ambient temperature, inflow velocity, liquid water content, median volume diameter, and icing duration; calculating the mean and standard deviation of the icing condition parameters in the entire dataset; and using the mean and standard deviation to perform Z-score standardization on the input icing condition parameters to obtain the condition vector.

3. The polar wind airfoil ice shape data augmentation method based on improved generative adversarial networks according to claim 1, characterized in that, The CIN-SAM-GAN high-quality ice-shaped data augmentation model includes a generator based on an improved U-Net architecture and a discriminator based on a PatchGAN architecture. Conditional instance normalization modules are embedded in the bottleneck layer and upsampling block of the generator, and a spatial attention module is embedded before the final output layer of the generator decoder.

4. The polar wind airfoil ice shape data augmentation method based on improved generative adversarial networks according to claim 1, characterized in that, In S2, dynamic instance noise with a linearly decaying standard deviation is introduced into the image input to the discriminator during adversarial training, and the generator weights are updated using a multidimensional joint loss function that integrates adversarial loss, L1 reconstruction loss, Dice cross-union loss, and total variational smoothing loss.

5. The polar wind airfoil ice shape data augmentation method based on improved generative adversarial networks according to claim 3, characterized in that, The conditional instance normalization module performs instance normalization processing on the input feature map without affine parameters; it inputs the conditional vector into a multilayer perceptron conditional mapping network containing a ReLU activation function, outputs affine parameters of a specific dimension, and splits the affine parameters into scaling parameters along the channel dimension. and offset parameters ;Scaling parameters and offset parameters After expanding to a spatial dimension that matches the input feature map, the conditional affine transformation formula is applied. The modulated feature map is obtained.

6. The polar wind airfoil ice shape data augmentation method based on improved generative adversarial networks according to claim 3, characterized in that, The spatial attention module performs average pooling and max pooling operations on the input feature map along the channel dimension, and concatenates the results of the two operations along the channel dimension to obtain a 2-channel feature map. The 2-channel feature map is then reduced to 1 channel through a convolutional layer with a kernel size of 7×7, and a 2D spatial attention mask with a value range of 0 to 1 is generated using the Sigmoid activation function. The two-dimensional spatial attention mask is multiplied element-wise with the original input feature map.

7. The polar wind airfoil ice shape data augmentation method based on improved generative adversarial networks according to claim 3, characterized in that, The decoder architecture of the generator specifically includes: The decoder consists of four upsampling blocks, each composed of a transposed convolutional layer with a stride of 2, a conditional instance normalization module, and a ReLU activation function. The output feature map of each upsampling block is spliced ​​with the feature map of the corresponding layer of the encoder via skip connections. The spliced ​​feature map is then passed through a convolutional layer, a batch normalization layer, and a ReLU activation function for channel compression and feature extraction. The last upsampling block of the decoder outputs a feature map of size 32×256×256. This feature map is then passed through a spatial attention module, an output convolutional layer, and a Tanh activation function to output an airfoil ice-shaped image of size 1×256×256.

8. The polar wind airfoil ice shape data augmentation method based on improved generative adversarial networks according to claim 4, characterized in that, The specific method for adding dynamic instance noise during adversarial training is as follows: before inputting the real icing image and the fake image generated by the generator into the discriminator, Gaussian noise with a standard deviation that decreases linearly with the number of training rounds is added to both; the specific decay formula for the standard deviation of the dynamic noise is: in, The current training round number is 100. When the training round number reaches 100, the noise standard deviation decreases to 0.

9. The polar wind airfoil ice shape data augmentation method based on improved generative adversarial networks according to claim 4, characterized in that, The formula for calculating the joint loss function during adversarial training is as follows: in, For least-squares-based adversarial loss, For L1 pixel-level reconstruction loss, Dice loss is used to constrain complex icing topologies. This is the total variational smoothing loss used to penalize jagged edges in the generated image.

10. The polar wind airfoil ice shape data augmentation method based on improved generative adversarial networks according to claim 1, characterized in that, A large-scale icing enhancement dataset was used as training data to supervise the training of a deep learning benchmark prediction model, thereby improving the accuracy of the prediction model in predicting ice shape.

Images