A wind turbine blade damage identification method and system based on improved convolutional neural network

Abstract

The present invention discloses a method and system for identifying wind turbine blade damage based on an improved convolutional neural network. The method involves acquiring wind turbine blade images and constructing a training set and a test set. The training set is used to train a wind turbine blade damage identification model constructed based on the improved convolutional neural network, and the test set is used to test the model. Wind turbine blade damage detection is then performed based on the trained wind turbine blade damage identification model. The constructed wind turbine blade damage identification model based on the improved convolutional neural network can improve the accuracy of wind turbine blade damage identification.

Description

Technical Field

[0001] The present invention relates to the technical field of wind turbine operation and maintenance, and in particular to a wind turbine blade damage identification method and system based on an improved convolutional neural network. Background Art

[0002] Wind turbine blades, as one of the most critical components of wind turbines, directly impact their power generation efficiency. Wind turbines operate around the clock, making them vulnerable to strong winds, sandstorms, lightning, rain, and snow. Therefore, regular inspections are crucial. In the past, wind blade inspections were mostly manual, with workers typically observing blade damage through a telescope. This method is time-consuming and labor-intensive, as the image of the blades observed through the telescope is often blurry and often affected by weather and light, making the inspection less than ideal. Timely maintenance and care for wind turbine blades is crucial. Even minor wear can be prevented by prompt repairs. However, if repairs are not carried out promptly, blade wear can gradually develop into cracks, leading to blade breakage and loss, resulting in significant economic losses and potentially safety incidents. Therefore, developing timely and effective methods for identifying wind blade damage has significant practical engineering significance and application value.

[0003] Common methods for wind blade damage identification include fiber optic sensing, thermal imaging, laser speckle interferometry, and acoustic sequencing. These methods suffer from limitations such as high cost and cumbersome operation. Some methods can only identify a single type of damage defect, hindering their practical application in engineering. The application of deep learning methods to wind blade damage identification has improved the efficiency of wind blade damage detection, effectively overcoming the high cost and cumbersome operation limitations of traditional methods and enabling the identification of multiple types of damage defects. With the in-depth research of numerous Chinese and international researchers, convolutional neural network architectures have been continuously refined, and the efficiency of wind blade damage identification has gradually increased. However, the convolution operation in the network inevitably loses image detail information. Blade damage characteristics vary greatly in length and depth, and wear varies in size and shape. This loss of detail often results in significant errors in identification, leaving room for improvement in recognition accuracy. Summary of the Invention

[0004] To this end, the present invention provides a fan blade damage identification method and system based on an improved convolutional neural network to solve the problem of low recognition accuracy in existing fan blade damage identification methods.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] According to a first aspect of an embodiment of the present invention, a method for identifying wind turbine blade damage based on an improved convolutional neural network is proposed. The method includes:

[0007] Obtain wind turbine blade images and construct training and test sets;

[0008] Using the training set to train the constructed wind turbine blade damage recognition model based on the improved convolutional neural network, and using the test set to test the model;

[0009] Wind blade damage detection is performed based on the trained wind blade damage recognition model.

[0010] Furthermore, the method further includes preprocessing the image, specifically:

[0011] The damaged leaf images were selected and the undamaged images were used as negative samples to classify and label the damage. The damaged area of ​​each image was manually captured as a ROI region, and the dataset was enhanced and grayscaled.

[0012] Furthermore, the wind turbine blade damage identification model based on the improved convolutional neural network specifically includes:

[0013] The first convolution layer is used to increase the dimension of the feature map, and then the BN layer is introduced to perform data normalization to improve the convergence speed of the model. Then, three groups of identical synthetic structures are sequentially passed through.

[0014] An EAC Block module, a DPCI_SE Block module, and a dimensionality reduction layer are sequentially connected to form a structure. The enhanced asymmetric convolution module EAC Block with residual connection is used to strengthen the content extracted at the convolution kernel backbone to enhance the extraction of feature information by the convolution operation. Then it enters the dual-pooled splicing input attention module DPCI_SEBlock to obtain more global and local channel information and improve the ability to obtain spatial location information of the damage. After that, it enters the dimensionality reduction layer to reduce the dimension of the feature map and reduce the calculation of the internal parameters of the model.

[0015] After the feature information passes through three groups of synthetic structures, it is finally input into the FC fully connected layer and the SoftMax function for combined classification output.

[0016] Furthermore, the EAC Block module is improved based on the asymmetric convolutional layer, including three parallel branches. A convolution layer with a convolution kernel size of 1 is added to each branch, and then the GELU activation function is passed to further improve the nonlinearity of the module and enhance the expressive ability of the module.

[0017] Furthermore, the DPCI_SE Block module is improved based on the attention module SE block. A dual pooling layer including an average pooling layer and a maximum pooling layer is added before the SEBlock module. The feature map is input into the parallel structure of the average pooling layer and the maximum pooling layer, and the ReLU activation function in the SE Block is replaced by the GELU activation function.

[0018] For the input feature map, the average pooling layer can retain the global feature information, and the maximum pooling layer can retain the local feature information. Finally, the feature information of the two parts is aggregated through the splicing operation to obtain richer feature information, ensuring that the subsequent attention module can more accurately extract the spatial position information of the image.

[0019] Furthermore, the training set is used to train the constructed wind turbine blade damage recognition model based on the improved convolutional neural network, which specifically includes:

[0020] The model is trained using the cross entropy improved by label smoothing as the loss function to improve the overfitting phenomenon of the model.

[0021] Furthermore, the training set is used to train the constructed wind turbine blade damage recognition model based on the improved convolutional neural network, which specifically includes:

[0022] Use the Adam gradient descent optimizer with good performance for optimization, set the initial learning rate, and formulate a learning rate decay strategy with a fixed step size.

[0023] According to a second aspect of an embodiment of the present invention, a wind turbine blade damage identification system based on an improved convolutional neural network is proposed, the system comprising:

[0024] The dataset construction module is used to obtain wind turbine blade images and construct training and test sets;

[0025] A model training module, configured to train the constructed wind turbine blade damage identification model based on the improved convolutional neural network using the training set, and to test the model using the test set;

[0026] The damage detection module is used to detect fan blade damage based on the trained fan blade damage recognition model.

[0027] According to a third aspect of an embodiment of the present invention, a computer storage medium is proposed, wherein the computer storage medium contains one or more program instructions, and the one or more program instructions are used to be executed by a wind blade damage identification system based on an improved convolutional neural network to perform any of the methods described above.

[0028] The present invention has the following advantages:

[0029] This invention proposes a fan blade damage identification method and system based on an improved convolutional neural network. The method obtains fan blade images and constructs training and test sets. The training set is used to train a fan blade damage identification model based on the improved convolutional neural network, and the test set is used to test the model. Fan blade damage detection is then performed based on the trained fan blade damage identification model. The constructed fan blade damage identification model based on the improved convolutional neural network can improve the accuracy of fan blade damage identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other implementation drawings based on the provided drawings without inventive effort.

[0031] Figure 1 A schematic flow chart of a method for identifying wind turbine blade damage based on an improved convolutional neural network provided in Example 1 of the present invention;

[0032] Figure 2 A schematic diagram of a specific implementation flow of a wind turbine blade damage identification method based on an improved convolutional neural network provided in Example 1 of the present invention;

[0033] Figure 3 A structural diagram of a fan blade damage identification model based on an improved convolutional neural network in a fan blade damage identification method based on an improved convolutional neural network provided in Example 1 of the present invention;

[0034] Figure 4 This is a structural diagram of the EAC Block module in a wind turbine blade damage identification method based on an improved convolutional neural network provided in Example 1 of the present invention;

[0035] Figure 5 This is a structural diagram of the DPCI_SE Block module in a fan blade damage identification method based on an improved convolutional neural network provided in Example 1 of the present invention;

[0036] Figure 6 The accuracy and loss of the recognition model in the wind turbine blade damage recognition method based on an improved convolutional neural network provided in Example 1 of the present invention;

[0037] Figure 7 A lightweight network structure diagram for a transfer learning comparison experiment in a wind turbine blade damage identification method based on an improved convolutional neural network provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0038] The following describes the implementation of the present invention using specific embodiments. Those skilled in the art will readily understand the other advantages and benefits of the present invention from the disclosure herein. Obviously, the embodiments described are only a portion of the present invention, not all of it. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are intended to fall within the scope of protection of the present invention.

[0039] Example 1

[0040] like Figure 1 As shown, this embodiment proposes a wind turbine blade damage identification method based on an improved convolutional neural network, the method comprising:

[0041] S100, obtaining wind turbine blade images and constructing a training set and a test set;

[0042] S200, using the training set to train the constructed wind turbine blade damage recognition model based on the improved convolutional neural network, and using the test set to test the model;

[0043] S300: Perform fan blade damage detection based on the trained fan blade damage recognition model.

[0044] like Figure 2 As shown, the specific embodiment process is as follows:

[0045] 1. Image Collection

[0046] Extract wind turbine blade images from some frames of drone aerial video sequences.

[0047] 2. Image preprocessing

[0048] 3. Manually select damaged leaf images and use undamaged images as negative samples to classify and annotate the damage. A region of interest (ROI) is manually captured around the damaged area in each image. Data augmentation is then performed on the dataset, adding random noise, mirroring, and rotating the damaged images clockwise. Grayscaling the dataset preserves gradients while reducing computational effort and accelerating training.

[0049] 4. Construction of wind turbine blade damage identification model based on improved convolutional neural network

[0050] The overall structure of the fan blade damage recognition model based on the improved convolutional neural network is as follows: Figure 3 shown.

[0051] The model's first convolution layer increases the dimensionality of the feature map, followed by a batch normalization layer for data normalization, accelerating model convergence. The model then enters the enhanced asymmetric convolution module (EAC Block) with residual connections. This improves on the asymmetric convolution structure by enhancing the content extracted at the kernel's core, thereby enhancing the convolution operation's ability to extract feature information. The model then enters the dual-pooled input attention module (DPCI_SE Block), an improvement on the SE Block that captures more global and local channel information. The dimensionality reduction layer then enters. This is essentially a convolution layer with a kernel size of 1. Its purpose is to reduce the dimensionality of the feature map and minimize the computational effort required for internal model parameters. An EAC Block, a DPCI_SEBlock, and a dimensionality reduction layer form a structure. After passing through three such structures, the feature information is finally fed into the fully connected (FC) layer and the SoftMax function for combined classification output.

[0052] 5. Model training

[0053] The preprocessed wind blade image set is trained based on the improved convolutional neural network model, and several optimal trained models are exported.

[0054] 6. Model Deployment

[0055] The trained optimal model is deployed to the computer (or mobile device) where wind blade damage personnel conduct damage detection.

[0056] 7. Damage identification

[0057] The software required for wind blade damage identification is built on the computer (or mobile device) used by the staff to perform damage detection, and wind blade damage detection is performed.

[0058] The wind turbine blade damage recognition model based on the improved convolutional neural network proposed in this embodiment introduces an enhanced asymmetric convolution module based on the improved asymmetric convolution, which enhances the feature extraction capability in the convolution process. The enhanced asymmetric convolution module EAC Block (Enhanced Asymmetric Convolution Block) is constructed based on the asymmetric convolution layer, with N set to 3. Its internal structure is as follows: Figure 4 As shown in the figure, the EAC Block adds a convolution layer with a kernel size of 1 to each branch, and then passes it through the GELU activation function to further improve the nonlinearity of the module and enhance its expressive power.

[0059] The double pooling input attention module DPCI_SE Block (Double Pooling Concatenated InputSE Block) is improved based on SEblock. The core idea of ​​SE block is to separate channels and obtain the learning weight of each channel feature map according to the loss degree, so that the effective feature map has a large learning weight, and the invalid or poor feature map has a small learning weight, which can enable the model to obtain better training results. The specific operations are divided into Squeeze stage, Excitation stage, and Reweight stage. The basic idea of ​​the Squeeze stage is to compress the feature input along the spatial dimension, use a global average pooling to improve the receptive field, and match the output dimension with the number of input feature channels. Let the feature map of each channel be u c , the global information of each channel feature map is obtained as Z c , the formula is as follows. Where H represents the height of the feature map and W represents the width of the feature map.

[0060]

[0061] The basic idea of ​​the Excitation stage is based on the correlation of feature channels. Each feature channel generates a weight, which is used to represent the importance of the feature channel. The input is the global information z obtained in the Excitation stage, which is then passed through a fully connected layer, a ReLU activation function, and an excitation operation in the fully connected layer. Finally, the output result is limited to the interval [0,1] by sigmoid, and the output learning weight s is output. The formula is as follows. W1 and W2 are the weight matrices of the two fully connected layers respectively. r is the decay factor, which is 16 as shown in the literature.

[0062] s=F ex (z,W)=σ sigmoid (W2σ ReLU (W1z))

[0063] Finally, the reweighting phase takes the weights obtained in the excitation phase and multiplies them channel by channel with the previous features. This enhances important features and weakens unimportant ones, making the desired features more specific. The formula is as follows.

[0064]

[0065] The dual-pooled splicing input attention module DPCI_SE Block proposed in this scheme adds a dual-pooling layer before SE Block, flows the feature map U into the parallel structure of the average pooling layer and the maximum pooling layer, and replaces the ReLU activation function in SE Block with the GELU activation function. The specific structure is as follows Figure 5 As shown in the figure, for the input feature map, the average pooling layer can retain global feature information, while the maximum pooling layer can retain local feature information. Finally, the two parts of feature information are aggregated through the splicing operation to obtain richer feature information U*, as shown in the following formula. This ensures that the subsequent attention module can more accurately extract the spatial location information of the image.

[0066] U * =[AvgPool(U),MaxPool(U)]

[0067] Experimental example

[0068] 1. Model training:

[0069] The graphics card used in this experiment is an NVIDIA GeForce RTX 3060 laptop GPU with 6GB of video memory. The CPU is an 11th-generation Intel Core i7-11800H 2.30GHz. The deep learning framework used is Pytorch version 1.7.1, and GPU training and testing are performed.

[0070] The drone images were categorized by damage type: normal, cracked, and surface debonding. 600 images were normal, 1,008 images were cracked, and 1,040 images were surface debonding. The preprocessed wind turbine images were divided into training and test sets in a 9:1 ratio. Before training the network, the dataset was resized to 56×56 pixels and data normalization was performed to ensure the image data conformed to a standard normal distribution, accelerating model convergence.

[0071] After smoothing the labels, the loss function becomes more robust against overfitting. Essentially, the loss function encourages the model's output to approximate the corresponding sample labels. In traditional one-hot encoded network learning, the model is encouraged to predict the probability of the target class close to 1 and the probability of the non-target class close to 0. The traditional one-hot encoded label vector is yi, as shown in the following formula.

[0072]

[0073] Among them, y represents the target label and i represents the output value. However, this is likely to cause overfitting. If the output is very close to the sample label, but the training sample is wrong, it will cause great damage to the model. Because during the training process, a problematic sample is forcibly learned and its probability is made very high, which will have a great impact on the estimation of the posterior probability. And sometimes there is no correlation between classes. If the output probabilities are greatly different, this will lead to a certain degree of overfitting. Therefore, the distance between the two extremes can be appropriately reduced and adjusted to alleviate the impact of the wrong label. In this way, the new label vector To replace the traditional one-hot encoded label vector, the formula is as follows.

[0074]

[0075] Here, ε represents a small positive real hyperparameter (typically 0.1), and M represents the total number of multi-class classifications. Thus, the extreme values ​​are 1-ε and ε / M, rather than 1 and 0. Label smoothing is equivalent to adding noise to the true distribution, preventing the model from being overconfident about the correct label. This reduces the difference between the predicted output values ​​of positive and negative samples, thus preventing overfitting and mitigating the impact of incorrect labels, thereby improving the model's generalization ability.

[0076] The cross entropy after label smoothing is used as the loss function, and the Adam gradient descent optimizer with good performance is used for optimization. The decay rate β1 = 0.9, β2 = 0.99, batch_size is set to 32, and a fixed step learning rate decay strategy is adopted. The initial learning rate is set to 0.001. After every 10 epochs, the learning rate is reduced to 0.6 times the current value. A total of 30 epochs are trained. After each training epoch, the accuracy and loss of the current model are recorded, as shown in the following example: Figure 6 shown.

[0077] Depend on Figure 6 The convergence of the model's accuracy and loss can be seen with increasing training rounds. It is clearly observed that during the first 10 rounds of training, the model's accuracy and loss both rise and fall rapidly. Between rounds 11 and 20, the accuracy fluctuates slightly, and the loss plateaus briefly. During the 20th round of training, after the learning rate undergoes a second decay, the model's accuracy exhibits a slow and steady upward trend, ultimately converging to 99.138%. The loss also continues a slow downward trend, finally converging to 9.4234e-3. Based on the final model convergence, the proposed model is effective in improving wind turbine blade damage identification capabilities.

[0078] 2. Transfer learning comparison experiment:

[0079] Since the proposed wind turbine blade damage recognition model based on the improved convolutional neural network has fewer parameters and is a lightweight convolutional model, other common lightweight models such as MobileNet_V1, MobileNet_V2, 1.0-SqueezeNext-23, ShuffleNet_V1, and ShuffleNet_V2 are selected for comparison. The overall structure of their respective main models is as follows: Figure 7 shown.

[0080] The experiment designed the above five pre-trained models for transfer learning on a wind blade damage dataset. Each model was trained for 30 epochs, and the final accuracy and training time were recorded. Because ShuffleNet_V1 achieves a balanced accuracy and parameter count when the number of groups g = 3, g was set to 3 for the experiment. The main evaluation metrics used in the experiment include accuracy, recall, precision, and F1 index.

[0081] (1) The accuracy rate is calculated by dividing the number of correctly identified samples by the total number of samples, as shown below:

[0082]

[0083] Although accuracy can determine the overall accuracy rate, it is not a good indicator for measuring results when the sample is unbalanced. Therefore, other model evaluation metrics are needed for joint experimental evaluation.

[0084] (2) Confusion Matrix. For the fan blade surface damage classification and identification experiment, the dataset contains three labels, including cracks, surface peeling, and normal. Therefore, in the confusion matrix for fan blade surface damage, each column represents a type of damage identified, and each row represents the prediction result of each type. The sum of each row of data is the number of real samples of this type. Therefore, the basic form of the fan blade surface damage confusion matrix is ​​shown in Table 1:

[0085] Table 1 Confusion matrix of wind blade surface damage identification structure

[0086]

[0087] Here, a, b, and c represent the number of samples identified as positive for the three labels: crack, surface delamination, and normal. The network model's final classification and recognition results can be divided into six parts: positive crack damage is represented by a, and false crack damage is represented by b'+c'. Positive surface delamination is represented by b', and false surface delamination is represented by a'+c". Positive normal is represented by c, and false normal is represented by a"+b".

[0088] (3) Recall rate. Recall rate refers to the original sample and means the probability of being identified as a positive sample among the samples that are actually positive. Taking the recall rate of crack labels as an example, its calculation is shown in the following formula.

[0089]

[0090] (4) Precision. Precision refers to the recognition result. It means the probability of a sample being actually positive among all samples identified as positive. Taking the accuracy of crack labeling as an example, its calculation is shown in the following formula.

[0091]

[0092] From a classification perspective, recall rate and precision rate are mutually exclusive. If the recognition ability of the model remains unchanged, improving the recall rate will relatively reduce the precision rate, and vice versa.

[0093] (5) F1 index. The F1 index is an evaluation index that combines the first two and is used to comprehensively reflect the overall index. The calculation method is shown in the following formula:

[0094]

[0095] Similarly, the precision, recall, and F1 index of the classification and recognition of damaged and normal images of wind turbine blade surfaces can be calculated. Finally, the overall evaluation index of the model is obtained by macro-averaging, as shown in the following formula:

[0096]

[0097]

[0098]

[0099] The results of the transfer learning comparison experiment are shown in Table 2.

[0100] Table 2 Comparison results of transfer learning experiments

[0101]

[0102]

[0103] As can be seen from Table 2, for the wind blade damage dataset, the model proposed in this embodiment outperforms the MobileNet_V1, MobileNet_V2, SqueezeNext-23, and ShuffleNet_V1 models in terms of accuracy, recall, precision, and F1 index. The MobileNet_V1 and MobileNet_V2 models have more parameters than the proposed model. Therefore, the proposed model can, to a certain extent, replace both for the task of wind blade damage identification. Although the first four metrics of the SqueezeNext-23 model are inferior to those of the proposed model, it has fewer parameters and a particularly high training time. Considering model size and training time, the SqueezeNext-23 model is also a worthy option. ShuffleNet_V2 has slightly higher accuracy than the proposed model, and it also has fewer parameters and takes less training time, but its recall is slightly lower than that of the proposed model. In subsequent research, the proposed model's parameter count and computational complexity can be reduced by applying improved methods such as depthwise separable convolution and channel reordering. In actual task scenarios, we would prefer to identify all images of wind turbine blades with damage features, and would rather spend less manpower to do so. Therefore, the high recall rate of the proposed model is an advantage.

[0104] Example 2

[0105] Corresponding to the above-mentioned embodiment 1, this embodiment proposes a wind turbine blade damage identification system based on an improved convolutional neural network, the system comprising:

[0106] The dataset construction module is used to obtain wind turbine blade images and construct training and test sets;

[0107] A model training module, configured to train the constructed wind turbine blade damage identification model based on the improved convolutional neural network using the training set, and to test the model using the test set;

[0108] The damage detection module is used to detect fan blade damage based on the trained fan blade damage recognition model.

[0109] The functions performed by each component in the wind blade damage identification system based on an improved convolutional neural network provided by an embodiment of the present invention have been introduced in detail in the above embodiment 1, so they will not be described in detail here.

[0110] Example 3

[0111] Corresponding to the above embodiment, this embodiment proposes a computer storage medium, which contains one or more program instructions, and the one or more program instructions are used to be executed by a wind blade damage identification system based on an improved convolutional neural network to perform the method of Example 1.

[0112] Although the present invention has been described in detail above using general descriptions and specific embodiments, it will be apparent to those skilled in the art that modifications and improvements may be made thereto. Therefore, such modifications and improvements, without departing from the spirit of the present invention, are intended to be within the scope of protection claimed herein.

Claims

1. A wind turbine blade damage identification method based on an improved convolutional neural network, characterized in that: The method comprises: Obtain wind turbine blade images and construct training and test sets; Using the training set to train the constructed wind turbine blade damage recognition model based on the improved convolutional neural network, and using the test set to test the model; Perform fan blade damage detection based on the trained fan blade damage recognition model; The wind turbine blade damage identification model based on the improved convolutional neural network includes: The first convolution layer is used to increase the dimension of the feature map, and then the BN layer is introduced to perform data normalization to improve the convergence speed of the model. Then, three groups of identical synthetic structures are sequentially passed through. An EAC Block module, a DPCI_SE Block module, and a dimensionality reduction layer are sequentially connected to form a structure. The enhanced asymmetric convolution module EAC Block with residual connection is used to strengthen the content extracted at the convolution kernel backbone to enhance the extraction of feature information by the convolution operation. Then it enters the dual-pooled splicing input attention module DPCI_SE Block to obtain more global and local channel information and improve the ability to obtain spatial location information of the damage. After that, it enters the dimensionality reduction layer to reduce the dimension of the feature map and reduce the calculation of the internal parameters of the model. After the feature information passes through three groups of synthesis structures, it is finally input into the FC fully connected layer and the SoftMax function for combined classification output; The EAC Block module is based on the improvement of the asymmetric convolutional layer. It includes three parallel branches. A convolution layer with a convolution kernel size of 1 is added to each branch, and then a GELU activation function is applied to further improve the nonlinearity of the module and enhance the module's expressive power. The DPCI_SE Block module is improved based on the attention module SE block. A double pooling layer including an average pooling layer and a maximum pooling layer is added before the SE Block module. The feature map is input into the parallel structure of the average pooling layer and the maximum pooling layer, and the ReLU activation function in the SE Block is replaced by the GELU activation function. For the input feature map, the average pooling layer can retain the global feature information, and the maximum pooling layer can retain the local feature information. Finally, the feature information of the two parts is aggregated through the splicing operation to obtain richer feature information, ensuring that the subsequent attention module can more accurately extract the spatial position information of the image.

2. The wind turbine blade damage identification method based on an improved convolutional neural network according to claim 1 is characterized in that: The method further includes preprocessing the image, specifically: The damaged leaf images were selected and the undamaged images were used as negative samples to classify and label the damage. The damaged area of ​​each image was manually captured as a ROI region, and the dataset was enhanced and grayscaled.

3. The wind turbine blade damage identification method based on an improved convolutional neural network according to claim 1 is characterized in that: The training set is used to train the fan blade damage recognition model based on the improved convolutional neural network, which specifically includes: The model is trained using the cross entropy improved by label smoothing as the loss function to improve the overfitting phenomenon of the model.

4. The wind turbine blade damage identification method based on an improved convolutional neural network according to claim 1 is characterized in that: The training set is used to train the fan blade damage recognition model based on the improved convolutional neural network, which specifically includes: Use the Adam gradient descent optimizer with good performance for optimization, set the initial learning rate, and formulate a learning rate decay strategy with a fixed step size.

5. A wind turbine blade damage identification system based on an improved convolutional neural network, characterized in that: The system comprises: The dataset construction module is used to obtain wind turbine blade images and construct training and test sets; A model training module, configured to train the constructed wind turbine blade damage identification model based on the improved convolutional neural network using the training set, and to test the model using the test set; A damage detection module is used to detect fan blade damage based on the trained fan blade damage recognition model; The wind turbine blade damage identification model based on the improved convolutional neural network includes: The first convolution layer is used to increase the dimension of the feature map, and then the BN layer is introduced to perform data normalization to improve the convergence speed of the model. Then, three groups of identical synthetic structures are sequentially passed through. An EAC Block module, a DPCI_SE Block module, and a dimensionality reduction layer are sequentially connected to form a structure. The enhanced asymmetric convolution module EAC Block with residual connection is used to strengthen the content extracted at the convolution kernel backbone to enhance the extraction of feature information by the convolution operation. Then it enters the dual-pooled splicing input attention module DPCI_SE Block to obtain more global and local channel information and improve the ability to obtain spatial location information of the damage. After that, it enters the dimensionality reduction layer to reduce the dimension of the feature map and reduce the calculation of the internal parameters of the model. After the feature information passes through three groups of synthesis structures, it is finally input into the FC fully connected layer and the SoftMax function for combined classification output; The EAC Block module is based on the improvement of the asymmetric convolutional layer. It includes three parallel branches. A convolution layer with a convolution kernel size of 1 is added to each branch, and then a GELU activation function is applied to further improve the nonlinearity of the module and enhance the module's expressive power. The DPCI_SE Block module is improved based on the attention module SE block. A double pooling layer including an average pooling layer and a maximum pooling layer is added before the SE Block module. The feature map is input into the parallel structure of the average pooling layer and the maximum pooling layer, and the ReLU activation function in the SE Block is replaced by the GELU activation function. For the input feature map, the average pooling layer can retain the global feature information, and the maximum pooling layer can retain the local feature information. Finally, the feature information of the two parts is aggregated through the splicing operation to obtain richer feature information, ensuring that the subsequent attention module can more accurately extract the spatial position information of the image.

6. A computer storage medium, characterized in that The computer storage medium contains one or more program instructions, and the one or more program instructions are used to be used by a wind turbine blade damage identification system based on an improved convolutional neural network to execute the method according to any one of claims 1 to 4.

Images