Wind driven generator surface defect detection method

Abstract

The invention discloses a wind driven generator surface defect detection method, which comprises the following steps: acquiring a wind driven generator blade surface defect image, and preprocessing the image; building a surface defect detection model of the wind driven generator; training a defect detection model by using data in the training set, calculating a loss function and updating network parameters, and obtaining a defect detection model with optimal parameters; and detecting surface defects of the wind driven generator by using the defect detection model with the optimal parameters to obtain defect positions, defect types and defect sizes. According to the method, the detection accuracy of the defects of the small surface of the fan and similar to the background is improved, so that the application requirements in a complex environment are met, the operation and maintenance efficiency and safety of wind power generation are improved, and the sustainable development of the wind power industry is supported.

Description

Technical Field

[0001] The invention relates to image target detection technology, and in particular to a method for detecting surface defects of a wind turbine generator. Background Art

[0002] The continuous increase in wind power installed capacity has made the operation and maintenance of wind turbines a key issue. As a key component, wind turbine blades are prone to surface defects due to long-term exposure to complex external environments. If these defects are not detected and repaired in time, they will lead to a decline in wind turbine performance and even cause serious economic losses. At present, the detection of surface defects of wind turbines mainly relies on manual visual inspection and traditional non-destructive testing technologies, such as infrared thermal imaging and ultrasonic testing. Although these methods are effective, they have the problems of high cost, low efficiency, and limited ability to identify small area and complex defects in actual applications.

[0003] In recent years, with the development of deep learning technology, drone-based visible light defect imaging detection technology has begun to be applied to wind turbine defect detection, showing higher efficiency and lower cost than traditional methods, mainly including Two-Stage and One-Stage methods. Two-Stage methods such as the R-CNN series generate target candidate areas first and then classify and refine them. Although they have high accuracy, the computational cost is relatively high; One-Stage methods such as YOLO and SSD directly predict the target bounding box in the image, which is fast but slightly sacrifices accuracy. In addition, some researchers have introduced new network structures and optimization algorithms, such as SENet, GhostNetv2, etc., to improve the accuracy and computational efficiency of the model.

[0004] However, wind turbine surface defect detection faces many challenges in the field environment, such as the interference of natural conditions such as strong wind and strong ultraviolet rays on image acquisition, as well as the diversity of defects and the small proportion of defects in the image. These factors increase the difficulty of detection and make the existing deep learning models inadequate in feature extraction and small target detection. Summary of the invention

[0005] Purpose of the invention: In view of the above problems, the purpose of the present invention is to provide a method for detecting surface defects of a wind turbine.

[0006] Technical solution: A method for detecting surface defects of a wind turbine according to the present invention comprises the following steps:

[0007] Step 1, obtaining a surface defect image of a wind turbine blade, and preprocessing the image, and dividing the preprocessed image into a training set, a verification set, and a test set in proportion;

[0008] Step 2, building a wind turbine surface defect detection model, which includes a backbone network, a neck network and a head network;

[0009] Step 3: Use the data in the training set to train the defect detection model, calculate the loss function and update the network parameters to obtain the defect detection model with the best parameters;

[0010] Step 4: Use the defect detection model with optimal parameters to detect surface defects of the wind turbine to obtain the defect location, defect category and defect size.

[0011] Further, step 1 comprises:

[0012] Wind turbine blades are photographed under various weather conditions, and blade images including various surface defects are collected. All blade images are uniformly adjusted to the same pixels, and color restoration and contrast enhancement are performed in sequence.

[0013] Further, the backbone network includes a C_GSConv module, a first convolutional layer, a second convolutional layer, a first C3 module, a third convolutional layer, a second C3 module, a fourth convolutional layer, a third C3 module, a fifth convolutional layer, a fourth C3 module and a fast spatial pyramid pooling module, which are arranged in sequence;

[0014] The C_GSConv module includes two parallel channels. The first channel includes the first GSConv layer, the second GSConv layer and the 1×1 convolution layer arranged in sequence. The second channel includes the 1×1 convolution layer. Finally, the feature maps obtained after the two parallel channels are spliced.

[0015] Furthermore, the neck network includes a sixth convolutional layer, a first upsampling layer, a first fully connected layer, a fifth C3 module, a seventh convolutional layer, a second upsampling layer, a second fully connected layer, a sixth C3 module, an eighth convolutional layer, a third fully connected layer, a seventh C3 module, a ninth convolutional layer, a fourth fully connected layer, and an eighth C3 module, which are arranged in sequence.

[0016] Furthermore, in step 3, the loss function is obtained by combining the IoU loss and the distance attention mechanism, and the calculation expression is:

[0017] L WIoU =R WIoU L IoU

[0018] Where, L WIoU represents the WIoU loss function, R WIoU represents the penalty term of WIoU, L IoU Represents the IoU loss function, and the expressions are:

[0019]

[0020]

[0021] Where x and y represent the coordinates of the upper left corner of the anchor box, x gt and gt is the coordinate of the upper left corner of the target box, W g and H g Indicates the width and height of the minimum bounding box; W i and H i They represent the coordinates of the upper left corner of the prediction box, w and h represent the width and height of the prediction box, respectively. gt and h gt Represent the width and height of the ground-truth box respectively.

[0022] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0023] 1. The present invention has higher stability and accuracy in detecting blade surface defects for different types of wind turbines and under different climatic conditions;

[0024] 2. Through the defect detection model proposed in the present invention, small defects can be identified in advance, significantly reducing the frequency and cost of large-scale repairs, improving the overall efficiency and economic benefits of wind power generation, and improving the operating efficiency and safety of equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a flow chart of a method for detecting surface defects of a wind turbine;

[0026] Figure 2 This is the structural diagram of the C_GSConv module;

[0027] Figure 3 This is a comparison chart of the mAP curves of the improved model and the original model;

[0028] Figure 4 Comparison of bounding box loss curves between the improved model and the original model. DETAILED DESCRIPTION

[0029] In order to make the objectives, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0030] The present embodiment describes a method for detecting surface defects of a wind turbine generator, the flow chart of which is as follows: Figure 1 As shown, it at least includes the following steps 1 to 4.

[0031] Step 1, obtaining a surface defect image of a wind turbine blade, and preprocessing the image, and dividing the preprocessed image into a training set, a verification set, and a test set in proportion;

[0032] Specifically, step 1 includes:

[0033] Wind turbine blades are photographed under various weather conditions, and blade images including various surface defects are collected. All blade images are uniformly adjusted to the same pixels, and color restoration and contrast enhancement are performed in sequence.

[0034] In one example, a drone equipped with a high-resolution camera was used to photograph wind turbine blades in various weather conditions, collecting 5,000 images of surface defects including cracks, corrosion, and paint film damage, and all images were uniformly adjusted to 640×640 pixels. Color restoration and contrast enhancement were performed for different lighting and weather conditions to ensure that the model can work stably in a changing environment. Subsequently, the data set was divided into a training set, a validation set, and a test set in a ratio of 80%, 10%, and 10%, thereby ensuring that the model can be fully trained and accurately evaluated.

[0035] Step 2: Build a wind turbine surface defect detection model, which includes a backbone network, a neck network, and a head network.

[0036] Specifically, the backbone network includes a C_GSConv module, a first convolutional layer, a second convolutional layer, a first C3 module, a third convolutional layer, a second C3 module, a fourth convolutional layer, a third C3 module, a fifth convolutional layer, a fourth C3 module and a fast spatial pyramid pooling module, which are arranged in sequence;

[0037] Among them, the first C3 module, the second C3 module, the third C3 module and the fourth C3 module have the same structure;

[0038] The C_GSConv module includes two parallel channels. The first channel includes the first GSConv layer, the second GSConv layer and the 1×1 convolution layer arranged in sequence. The second channel includes the 1×1 convolution layer. Finally, the feature maps obtained after the two parallel channels are spliced.

[0039] like Figure 2As shown in the figure, the C_GSConv module is a convolutional network module for efficient feature extraction and fusion. Its input layer receives images or feature maps from the previous layer for further processing. First, through the first GSConv layer, the input data is quickly and efficiently convolved, which combines the standard convolution and pixel mixing process, which helps to extract and enrich features. Then, the feature map is further processed by the second GSConv layer to further extract finer-grained features. After the cascade of two GSConv layers, the feature extraction capability is enhanced. The next 1×1 convolution layer performs channel mixing, adjusts the channel dimension of the feature map, and more closely integrates the information in the feature map. At the same time, the input data is also bypassed through a separate path through a 1×1 convolution layer to retain the information of the original data, making the feature extraction more diversified and alleviating the gradient vanishing problem. Finally, the feature maps of the two channels are spliced ​​to obtain the output feature map of the C_GSConv module.

[0040] The input image is preprocessed before entering the GSConv layer to meet the input requirements of the model; after a series of convolution layers and activation functions, feature maps of different scales and depths are extracted. These feature maps contain information such as the position, shape, and texture in the image. The computational complexity of GSConv convolution is about half of that of ordinary convolution, but its feature extraction capability is the same as the latter. The feature map is input to the GSConv layer for processing. First, SC is used to perform channel dense convolution operations to retain the information connection between channels. Then, the output of SC is mixed with the output of DSC through the shuffle operation to achieve uniform fusion of channel information; finally, a feature map containing richer and more accurate information is output; feature maps of different scales and depths are spliced ​​or summed through feature fusion modules such as feature pyramid network (FPN) or FPN+path aggregation network (PAN) to generate a more comprehensive and detailed feature map; the fused feature map is input to the detection head for target detection. The detection head predicts the category, position, size and other information of the object based on the feature map, and finally outputs the detection results, such as detection box, category label and confidence.

[0041] The neck network includes a sixth convolutional layer, a first upsampling layer, a first fully connected layer, a fifth C3 module, a seventh convolutional layer, a second upsampling layer, a second fully connected layer, a sixth C3 module, an eighth convolutional layer, a third fully connected layer, a seventh C3 module, a ninth convolutional layer, a fourth fully connected layer, and an eighth C3 module, which are arranged in sequence.

[0042] In the neck network, the features extracted from the backbone network are further fused and processed through multiple C3 modules. Upsampling operations and feature connection operations are used to ensure the effective fusion and propagation of cross-scale features, improve the expressiveness of features, and lay a solid foundation for the prediction of the head network.

[0043] Finally, the multi-scale output head in the head network outputs feature maps at four different scales, and finally detects objects of different sizes. The convolutional layer and activation function are further processed to generate detection results containing bounding box coordinates and classification information, so that the model can maintain high efficiency while improving its accuracy and robustness in multi-scale target detection.

[0044] Step 3: Use the data in the training set to train the defect detection model, calculate the loss function and update the network parameters to obtain the defect detection model with the best parameters.

[0045] Specifically, in step 3, the IoU loss is combined with the distance attention mechanism to obtain the improved WIoU loss function as the loss function, and the calculation expression is:

[0046] L WIoU =R WIoU L IoU

[0047] Where, L WIoU represents the WIoU loss function, R WIoU represents the penalty term of WIoU, L IoU Represents the IoU loss function, and the expressions are:

[0048]

[0049] Where x and y represent the coordinates of the upper left corner of the anchor box, x gt and gt is the coordinate of the upper left corner of the target box, W g and H g Indicates the width and height of the minimum bounding box; W i and H i They represent the coordinates of the upper left corner of the prediction box, w and h represent the width and height of the prediction box, respectively. gt and h gt Represent the width and height of the ground-truth box respectively.

[0050] Step 4: Use the defect detection model with optimal parameters to detect surface defects of the wind turbine to obtain the defect location, defect category and defect size.

[0051] The present invention adopts an improved YOLOv5s_TGW model, combined with a lightweight network structure, a Transformer structure and a coordinated attention mechanism. The integration of these technologies not only significantly improves the processing speed of the model, but also greatly enhances the ability to identify defects of various sizes on the surface of wind turbines. In particular, during the operation of wind turbines, new or changing surface defects will continue to be generated due to changes in external environmental factors such as wind speed, temperature and humidity, as well as the wear of the body itself, which requires the detection system to have a high degree of adaptability and real-time performance. By introducing a coordinated attention mechanism, the present invention effectively improves the recognition accuracy of the model for defects under complex backgrounds and reduces misidentification caused by environmental changes. The lightweight model is to adapt to deployment on different hardware platforms, especially in some occasions with limited computing resources. The lightweight model can maintain efficient operation without sacrificing detection accuracy. The improvement in adaptability is mainly aimed at the stability and accuracy of different types of wind turbines and under different climatic conditions, ensuring that the model can be widely used worldwide. The optimization of the detection of small defects is to discover potential damage earlier and prevent these small defects from developing into big problems, thereby reducing maintenance costs and extending the service life of the wind turbine.

[0052] In addition, considering the operating costs and safety of wind turbines, continuous optimization and updating of detection technology will bring significant economic and safety benefits to the industry. Through real-time monitoring and regular inspections, major safety accidents caused by delayed defect handling can be effectively prevented, and energy production losses caused by downtime for maintenance can also be reduced. In terms of economic benefits, identifying and repairing small defects in advance can significantly reduce the frequency and cost of large-scale maintenance, and improve the overall efficiency and economic benefits of wind power generation. The surface defect detection technology of the wind turbine of the present invention, through its high efficiency and accuracy, and the direction of continuous optimization in the future, can not only improve the operating efficiency and safety of the equipment, but also promote the entire wind power industry to develop in a more efficient, more economical and more environmentally friendly direction. This technological advancement is conducive to promoting the widespread application of renewable energy and optimizing and upgrading the energy structure.

[0053] The defect detection model constructed in the present invention firstly utilizes lightweight convolution technology to enhance the information interaction between channels, and improves the small-size defect detection effect through richer feature information; secondly, by introducing the visual attention network module in the backbone network, the context information is enriched and the feature extraction ability of the convolutional neural network is improved; then, the coordinated attention mechanism is introduced in the neck network to capture the defect location information through the spatial direction; finally, the loss function of the bounding box regression is modified to WIoU, and a suitable gradient gain allocation strategy is formulated.

[0054] To further illustrate the superiority of the defect detection model in detecting blade surface defects, the defect detection model constructed in the present invention is compared with the YOLOv5s model. The average precision mean mAP curve comparison is shown in the figure below: Figure 3 As shown in the figure, in the first 50 iterations, the defect detection model detection mAP continued to rise. After more than 100 iterations, the detection mAP reached about 80%. Subsequently, the model's mAP tended to stabilize and showed the best confidence at this time. The weight file was saved once for each training. After completing 300 rounds of training, the weight file with the smallest loss value was obtained.

[0055] Bounding box loss curve comparison chart as follows Figure 4 As shown, it can be seen that the defect detection model in the present invention has a significant convergence effect. After 20 rounds, it reaches a stable range of convergence, and the bounding box loss is stabilized below 0.04. After 150 rounds, it is stabilized at around 0.03. It can be seen that the improvement in the bounding box regression accuracy is better than the YOLOv5s model.

[0056] Through the above experiments, it is concluded that the improved defect detection model of the present invention has a 4.14% higher detection accuracy than the original YOLOv5s model, significantly improves the detection capability of small defects, reduces the number of parameters by 2.29M, and significantly improves the detection speed to meet real-time detection needs.

Claims

1. A method for detecting surface defects of a wind turbine, characterized in that: The steps include: Step 1, obtaining a surface defect image of a wind turbine blade, and preprocessing the image, and dividing the preprocessed image into a training set, a verification set, and a test set in proportion; Step 2, building a wind turbine surface defect detection model, which includes a backbone network, a neck network and a head network; Step 3: Use the data in the training set to train the defect detection model, calculate the loss function and update the network parameters to obtain the defect detection model with the best parameters; Step 4: Use the defect detection model with optimal parameters to detect surface defects of the wind turbine to obtain the defect location, defect category and defect size.

2. A method for detecting surface defects of a wind turbine according to claim 1, characterized in that: Step 1 includes: Wind turbine blades are photographed under various weather conditions, and blade images including various surface defects are collected. All blade images are uniformly adjusted to the same pixels, and color restoration and contrast enhancement are performed in sequence.

3. A method for detecting surface defects of a wind turbine according to claim 1, characterized in that: The backbone network includes a C_GSConv module, a first convolutional layer, a second convolutional layer, a first C3 module, a third convolutional layer, a second C3 module, a fourth convolutional layer, a third C3 module, a fifth convolutional layer, a fourth C3 module and a fast spatial pyramid pooling module, which are arranged in sequence; The C_GSConv module includes two parallel channels. The first channel includes the first GSConv layer, the second GSConv layer and the 1×1 convolution layer arranged in sequence. The second channel includes the 1×1 convolution layer. Finally, the feature maps obtained after the two parallel channels are spliced.

4. A method for detecting surface defects of a wind turbine according to claim 1, characterized in that: The neck network includes a sixth convolutional layer, a first upsampling layer, a first fully connected layer, a fifth C3 module, a seventh convolutional layer, a second upsampling layer, a second fully connected layer, a sixth C3 module, an eighth convolutional layer, a third fully connected layer, a seventh C3 module, a ninth convolutional layer, a fourth fully connected layer, and an eighth C3 module, which are arranged in sequence.

5. A method for detecting surface defects of a wind turbine according to claim 1, characterized in that: In step 3, the loss function is obtained by combining the IoU loss and the distance attention mechanism, and the calculation expression is: L WIoU =R WIoU L IoU Where, L WIoU represents the WIoU loss function, R WIoU represents the penalty term of WIoU, L IoU Represents the IoU loss function, and the expressions are: Where x and y represent the coordinates of the upper left corner of the anchor box, x gt and gt is the coordinate of the upper left corner of the target box, W g and H g Indicates the width and height of the minimum bounding box; W i and H i They represent the coordinates of the upper left corner of the prediction box, w and h represent the width and height of the prediction box, respectively. gt and h gt Represent the width and height of the ground-truth box respectively.

Images