Dry-type transformer epoxy resin surface crack identification method

Abstract

The invention discloses a dry-type transformer epoxy resin surface crack identification method, and relates to the field of dry-type transformer epoxy resin surface crack detection, and the method comprises the steps: S1, constructing a dry-type transformer epoxy resin surface image data set; s2, constructing a YOLOv8 network model for identifying the surface crack of the epoxy resin of the dry-type transformer; training the YOLOv8 network model based on the dry-type transformer epoxy resin surface image data set; s3, inputting image information acquired by the camera into the YOLOv8 network model, and judging whether the surface of the dry-type transformer is cracked or not and the cracking position through the YOLOv8 network model; according to the method, the dry-type transformer epoxy resin surface cracks are identified by applying the improved YOLOv8 model, the method in the field of dry-type transformer crack identification is widened, and the inspection efficiency and accuracy are improved.

Description

Technical Field

[0001] The present invention relates to the field of dry-type transformer epoxy resin surface crack detection, and specifically to a dry-type transformer epoxy resin surface crack recognition method; further, it relates to a dry-type transformer epoxy resin surface crack recognition method using an improved YOLOV8 network structure. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and may not constitute prior art.

[0003] At present, the detection of cracks on the surface of transformers mainly relies on manual visual inspection. Engineers or technicians use tools such as naked eyes or magnifying glasses to inspect the surface of the transformer and look for visible signs of cracks. This method relies on human experience and professional knowledge, but may be affected by subjective factors and visual fatigue. In addition, during the transformer production process or in certain specific occasions, there are other methods for testing the sealing of various components of the transformer: ultraviolet detection, magnetic particle detection, and ultrasonic detection. These measures can identify defects or cracks on various components of the transformer, but there are also shortcomings.

[0004] 1) Ultraviolet detection: The ultraviolet (UV) detection method uses ultraviolet light to illuminate the surface of the transformer, and then uses a UV camera or special filters to observe the fluorescent reaction of the cracks. Cracks usually show more obvious fluorescence under ultraviolet light, which helps to identify them. This method has certain advantages for detecting small cracks or hidden cracks. However, the UV detection method has high requirements for lighting conditions and ambient light, and may be affected by external interference. In addition, some types of cracks may not produce obvious fluorescence reactions under ultraviolet light, so all cracks may not be accurately detected.

[0005] 2) Magnetic Particle Testing: Magnetic particle testing is a method that uses magnetic powder to detect surface cracks. Magnetic powder is applied to the surface of the transformer and then observed to see if there is any magnetic powder gathering at the crack. This method has certain limitations on the visibility and depth of the cracks. In addition, magnetic particle testing requires the application of a magnetic field and magnetic powder, which may bring certain difficulties to the operation and cleaning of the transformer.

[0006] 3) Install metal oxide arresters: The ultrasonic detection method uses the characteristics of ultrasonic waves propagating inside objects to detect surface and internal cracks. The ultrasonic probe emits ultrasonic pulses and receives signals reflected from the inside of the object. By analyzing the characteristics of the reflected signal, the existence and location of the crack can be detected. The ultrasonic detection method has high requirements for equipment and operators, requiring professional ultrasonic equipment and experienced operators. At the same time, the detection depth and accuracy of cracks will also be affected by the propagation characteristics of ultrasonic waves and materials.

[0007] 4) Other artificial intelligence methods: The artificial intelligence methods currently used for transformer surface defect detection mainly include edge detection methods, such as the patent with patent number: 202310251600.4, which detects surface defects of transformer bushings by detecting the edge of grayscale images. However, in the presence of complex backgrounds or noise, this method may not be able to accurately distinguish cracks from other textures or edges, and this method usually requires adjustment of some parameters, such as thresholds, filter sizes, etc. The selection of these parameters may have a significant impact on the detection results of cracks, and the adjustment of parameters requires certain experience and debugging processes. In addition, edge detection methods are sensitive to illumination changes and noise, which may lead to false detection or missed detection. Illumination changes may cause edges to break or blur, and noise may interfere with edge detection and extraction. Summary of the invention

[0008] The purpose of the present invention is to provide a method for identifying cracks on the epoxy resin surface of a dry-type transformer in view of the problems existing in the prior art.

[0009] The technical solution of the present invention is as follows:

[0010] A method for identifying cracks on epoxy resin surface of a dry-type transformer, comprising:

[0011] Step S1: constructing a dry-type transformer epoxy resin surface image dataset;

[0012] Step S2: constructing a YOLOv8 network model for identifying cracks on the epoxy resin surface of a dry-type transformer; and training the YOLOv8 network model based on a dataset of epoxy resin surface images of a dry-type transformer;

[0013] Step S3: input the image information collected by the camera into the YOLOv8 network model, and use the YOLOv8 network model to determine whether the surface of the dry-type transformer is cracked and the crack position.

[0014] Furthermore, the step S1 comprises:

[0015] Step S11: collecting an image of the epoxy resin surface of a dry-type transformer;

[0016] Step S12: cropping and enlarging the epoxy resin surface layer image of the dry-type transformer, and enlarging the crack position to form a sample set of the epoxy resin surface layer image of the dry-type transformer;

[0017] Step S13: annotating the defect degree of the dry-type transformer epoxy resin surface image sample set based on the surface crack degree and position to generate a VOC2007 data set;

[0018] Step S14: The VOC2007 dataset with surface cracks is expanded by using a GAN generative adversarial network written in Python to achieve data balance and make the model more accurate;

[0019] Step S15: Based on the expanded data set, the data set is divided into a training set, a test set, and a validation set according to a preset ratio.

[0020] Furthermore, the step S13 includes:

[0021] During the marking process, whether there are cracks on the surface of the dry-type transformer and the location of the cracks are marked in an XML format file.

[0022] Furthermore, the YOLOv8 network model consists of four parts: input layer, backbone network layer, neck network layer and detection head; wherein:

[0023] The input layer is used to receive the original image as input and pass the image information to the subsequent network layers;

[0024] The backbone network layer is used to extract image features and capture semantic information at different levels for subsequent target detection tasks.

[0025] The neck network layer fuses features at different levels to enhance the model’s ability to detect objects of different scales.

[0026] The detection head generates the results of target detection, including the location and category information of the target. The detection head part of the YOLOv8 network model uses prediction layers of different scales to adapt to targets of different sizes; each prediction layer is responsible for generating a set of bounding boxes and predicting the category probability, location and confidence of the target existence of each bounding box.

[0027] Furthermore, the step S2 comprises:

[0028] Step S21: randomly select an image in the training set as the input of the YOLOv8 network, and obtain feature maps of different scales after the backbone network layer extracts the features of the input image;

[0029] Step S22: The neck network layer fuses feature maps of different scales;

[0030] Step S23: predicting the probability and location of cracks in the image in the detection head, and generating a bounding box based on the prediction information;

[0031] Step S24: Iterate steps S21 to S23 until all training set images are traversed, calculate the YOLOv8 loss function in each iteration, use the gradient descent method to update the network parameters, and obtain the initial YOLOv8 model.

[0032] Furthermore, in the iterative process, DIou is used as the positioning loss function of YOLOv8.

[0033] Furthermore, in the iterative process, EIou is used as the YOLOv8 detection box regression loss function.

[0034] Furthermore, the step S2 further includes:

[0035] Step S25: Identify the validation set images according to the initial YOLOv8 model, iteratively count the average precision values ​​of the validation sets of different YOLOv8 models until the statistical mean average precision is stable, and obtain the trained YOLOv8 model. Use this model to identify all samples in the test set, obtain the surface cracking situation and position prediction box of the dry-type transformer in the test set image, and map it to the test set image.

[0036] Furthermore, the receptive field convolutional block attention RFCBAM Conv is used to replace the original network structure in the neck network layer; the convolutional block attention module is the basis of RFCBAM Conv, and the convolutional block attention module consists of an independently running channel attention module and a spatial attention module; Squeeze-and-Excitation Networks is used in RFCBAM Conv to replace the channel attention module in the convolutional block attention module to reduce the amount of calculation; K with a step size of K is used 2 The convolution operation extracts feature information and performs weighted calculations on channel and spatial attention at the same time to further enhance RFCBAM Conv’s attention to the spatial features of the receptive field.

[0037] Furthermore, the step S3 comprises:

[0038] The acquisition point of the zoom industrial camera is selected at the upper end of the grille on both sides of the site. The image information of the camera is transmitted to the industrial computer through the CAN bus, and the YOLOv8 network model is used to judge whether the surface of the dry-type transformer is cracked and the crack position.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. The present invention uses an improved YOLOv8 model to identify cracks on the epoxy resin surface of dry-type transformers, broadens the method of dry-type transformer crack identification, and improves the efficiency and accuracy of inspection.

[0041] 2. Compared with other artificial intelligence image recognition methods, the method used in the present invention can detect multiple different types of targets at the same time, which is very useful for surface crack recognition of dry-type transformers, because there may be cracks of various types and sizes. The model can detect these different types of cracks at one time, improving the comprehensiveness and efficiency of recognition.

[0042] 3. The present invention improves the YOLOv8 model by replacing part of the original network structure with the receptive field convolution block attention, thereby improving the defect that the YOLOv8 model cannot accurately identify small targets and improving the accuracy of the YOLOv8 model. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is the structure diagram of the YOLOv8 network model;

[0044] Figure 2 This is the structure diagram of the receptive field convolution block attention RFCBAM Conv;

[0045] Figure 3 Attention module structure diagram for convolutional block;

[0046] Figure 4 This is a schematic diagram for the installation of a zoom industrial camera;

[0047] Figure 5a This is the change diagram of the loss function of the training set of YOLOv8;

[0048] Figure 5b To improve the training set loss function change diagram of YOLOv8;

[0049] Figure 6a This is the change graph of the validation set loss function of YOLOv8;

[0050] Figure 6b To improve the loss function change graph of YOLOv8;

[0051] Figure 7 To improve the YOLOv8 dry transformer surface crack recognition result map. DETAILED DESCRIPTION

[0052] It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0053] The features and performance of the present invention are further described in detail below in conjunction with the embodiments.

[0054] Embodiment 1

[0055] Epoxy resin dry-type transformers will be affected by many factors during installation and operation, causing cracks. For example, during operation, they will be affected by temperature changes. When the temperature rises, the windings and epoxy resin will expand, and during the cooling process, they will shrink, resulting in stress inside the windings and between the windings and the epoxy resin. If the stress exceeds the bearing capacity of the material, cracks will be formed; at the same time, mechanical stress during transportation, installation and operation will also cause deformation and displacement of the windings and epoxy resin, thereby causing cracks; as an insulating material, epoxy resin will experience insulation aging after long-term operation and exposure to environmental factors such as electric fields, heat and humidity. Insulation aging will cause the physical properties and mechanical strength of epoxy resin to decrease, thereby increasing the risk of cracks. The generation of cracks will destroy the continuity of epoxy resin, resulting in a decrease in the insulation performance of the windings. This may cause discharge between windings or between windings and ground, increasing the risk of equipment failure. In addition, under mechanical stress such as external force or vibration, cracks may expand and eventually lead to the breakage of the winding. At present, cracks in epoxy resin of dry-type transformers are mainly detected by naked eye. However, this manual detection method is inefficient and objective, and inevitably leads to missed detection, which poses a great safety hazard. Therefore, it is urgent to develop an artificial intelligence method for detecting and identifying cracks in epoxy resin surface of dry-type transformers.

[0056] Deep learning algorithms can build high-performance image classification and target detection models to achieve automatic recognition and positioning of cracks on the epoxy resin surface of dry-type transformers. Common deep learning models include convolutional neural networks (CNN), recurrent neural networks (RNN), and Transformer-based models. Through training with a large amount of labeled data, the accuracy and robustness of the models in crack detection tasks can be improved. YOLO (You Only Look Once) network is a target detection algorithm based on deep learning. Compared with traditional CNN (convolutional neural network) and RNN (recurrent neural network) models, YOLO network adopts a single-stage detection method to convert the target detection task into an end-to-end regression problem, achieving faster detection speed and suitable for application scenarios with high real-time requirements. Since cracks on the surface of dry-type transformers may be relatively small targets, traditional target detection methods may have difficulties in processing small targets. The YOLO network can effectively detect small-sized targets, including small surface cracks of dry-type transformers, by detecting on multiple feature layers. However, even though the YOLO network has advantages in speed and global perception, the single-stage detection method of the YOLO network may cause some positioning errors.

[0057] Therefore, this embodiment is applied to surface crack detection of epoxy resin of dry-type transformer on the basis of improving the YOLOv8 algorithm to ensure the stable operation of the dry-type transformer and improve the detection efficiency and accuracy.

[0058] See also Figure 1-3 , a method for identifying cracks on the epoxy resin surface of a dry-type transformer, comprising:

[0059] Step S1: constructing a dry-type transformer epoxy resin surface image dataset;

[0060] Step S2: constructing a YOLOv8 network model for identifying cracks on the epoxy resin surface of a dry-type transformer; and training the YOLOv8 network model based on a dataset of epoxy resin surface images of a dry-type transformer;

[0061] Step S3: input the image information collected by the camera into the YOLOv8 network model, and use the YOLOv8 network model to determine whether the surface of the dry-type transformer is cracked and the crack position.

[0062] In this embodiment, specifically, step S1 is to collect a dataset of epoxy resin surface images of a dry-type transformer, process and expand the dataset, and then divide the expanded dataset, which specifically includes the following steps:

[0063] Step S11: collecting an image of the epoxy resin surface of a dry-type transformer;

[0064] Step S12: cropping and enlarging the epoxy resin surface layer image of the dry-type transformer, and enlarging the crack position to form a sample set of the epoxy resin surface layer image of the dry-type transformer;

[0065] Step S13: annotating the defect degree of the dry-type transformer epoxy resin surface image sample set based on the surface crack degree and position to generate a VOC2007 data set;

[0066] Step S14: The VOC2007 dataset with surface cracks is expanded by using a GAN generative adversarial network written in Python to achieve data balance and make the model more accurate;

[0067] Step S15: Based on the expanded data set, the data set is divided into a training set, a test set, and a validation set according to a preset ratio;

[0068] The step S13 comprises:

[0069] During the marking process, the presence of cracks on the surface of the dry-type transformer and the location of the cracks are marked in an XML format file;

[0070] Specifically, in this embodiment, an industrial camera is used to obtain a sample of the surface image of the epoxy resin of a dry-type transformer. First, the dry-type transformer image collected on site is cropped and the crack position is enlarged. The resolution of the crack position image of the dry-type transformer used is set to 800×600, and the defect degree of the dry-type transformer epoxy resin sample set is annotated based on the degree and position of the surface cracking. During the annotation process, whether there is cracking on the surface of the dry-type transformer and the cracking position are annotated in an XML format file to generate a VOC2007 data set. Then, the GAN generative adversarial network written in python is used to expand the dry-type transformer data set with surface cracking to achieve data balance and make the model more accurate. The processed dry-type transformer epoxy resin data set is divided into a training set, a test set, and a validation set in a ratio of 6:2:2.

[0071] In this embodiment, the specific structure of the YOLOv8 network model is further explained. Figure 1 The YOLOv8 network model consists of four parts: input layer (Input), backbone network layer (Backbone), neck network layer (Neck) and detection head (Head); among which:

[0072] The input layer is used to receive the original image as input and pass the image information to the subsequent network layer. Usually, a 416x416 or 608x608 input image is used. The input image will be normalized to the range of 0 to 1 so that the neural network can learn more easily.

[0073] The backbone network layer is used to extract image features and capture semantic information at different levels for subsequent target detection tasks. YOLOv8 uses CSPDarknet53 as the default backbone network, which combines CSPNet (CrossStage Partial Network) and Darknet53 to improve the effect of feature extraction.

[0074] The neck network layer fuses features at different levels to enhance the model's ability to detect objects of different scales. YOLOv8 uses PANet (Path Aggregation Network) as the default feature fusion layer. PANet helps process multi-scale information, allowing the model to better adapt to objects of different sizes.

[0075] The detection head generates the results of target detection, including the location and category information of the target. The detection head part of the YOLOv8 network model uses prediction layers of different scales to adapt to targets of different sizes; each prediction layer is responsible for generating a set of bounding boxes and predicting the category probability, location (coordinates) and confidence of the existence of the target for each bounding box.

[0076] In this embodiment, it should also be noted that the YOLOv8 network model used above is an improved model, that is, the receptive field convolution block attention RFCBAM Conv (such as Figure 2 The original network structure is replaced by the convolutional block attention module (shown in Figure 2); the convolutional block attention module is the basis of RFCBAM Conv, and the convolutional block attention module consists of an independently running channel attention module and a spatial attention module; Squeeze-and-Excitation Networks is used in RFCBAM Conv to replace the channel attention module in the convolutional block attention module to reduce the amount of calculation; K with a step size of K is used 2 The convolution operation extracts feature information and performs weighted calculations on channel and spatial attention at the same time to further enhance RFCBAM Conv's attention to the spatial features of the receptive field;

[0077] The specific improvement process is as follows:

[0078] Although YOLOv8 performs well in target detection tasks, there are still some defects in the recognition of small targets. For example, when the target is small, insufficient receptive field may cause the network to be unable to capture the global or larger contextual information of the target. In this case, the network may only focus on the local features of the target, ignoring the relationship between the target and the surrounding environment, resulting in information loss; the insufficient receptive field of YOLOv8 may cause the network to estimate the target position inaccurately. The network can only perceive the local features near the target, but cannot accurately determine the center position of the target, affecting the accuracy of target positioning; insufficient receptive field may reduce the network's adaptability to targets of different scales. For small targets, the network may not have enough contextual information to correctly understand the overall shape and structure of the target.

[0079] Therefore, in order to further improve the attention of the YOLOv8 network to the surface cracks of the epoxy resin of the dry-type transformer, suppress irrelevant background noise, and enhance the recognition accuracy and generalization ability of the model, the present invention replaces part of the original network structure with the receptive field convolution block attention RFCBAM Conv based on YOLOv8.

[0080] The convolutional block attention module (CBAM) is the basis of RFCBAMConv. As a lightweight attention module, the convolutional block attention module is mainly composed of independently running channel attention modules (CAM) and spatial attention modules (SAM), such as Figure 3 As shown. Among them, the channel attention module is responsible for paying attention to the key information in the image, and the spatial attention module is responsible for paying attention to the location information of the cracks in the dry-type transformer. In the channel attention module stage, the convolutional block attention module calculates the maximum eigenvalue and average eigenvalue of the input feature map through the parallel MaxPool layer and AvgPool layer, uses the Share MLP module to compress and expand the number of channels, and uses the Sigmoid activation function to obtain two activated results. Finally, the activation results are added element by element, processed by the Sigmoid activation function, and then multiplied with the original image to ensure that the channel dimension remains unchanged. In the SAM stage, CBAM generates two 1×H×W feature maps through the MaxPool layer and the AvgPool layer respectively, and then splices them and processes them through the convolution layer to generate spatial attention weights, which are processed by the Sigmoid activation function and then multiplied with the original image to restore C×H×W (C: number of channels, H: height, W: width).

[0081] In order to further enhance the attention of the convolutional block attention module to the spatial features of the receptive field, the convolutional block attention module is improved. The constructed RFCBAM Conv is as follows Figure 2 As shown in the figure. SE (Squeeze-and-Excitation Networks) is used to replace the channel attention module in the convolutional block attention module to reduce the amount of calculation; K with a step size of K is used. 2 The convolution operation extracts feature information and performs weighted calculations on channel and spatial attention at the same time to further enhance RFCBAM Conv’s attention to the spatial features of the receptive field.

[0082] In YOLOv8, the C2f structure is used to replace the C3 structure to enrich the gradient flow and optimize the number of channels to better adapt to the multi-scale model. The neck network layer still uses the commonly used PAN-FPN, which preserves information from different feature levels. However, by introducing the attention module, the correlation between global and local features is further enhanced, which promotes more effective information fusion and strengthens the propagation of information at high and low levels. In order to improve the attention of the YOLOv8 network to the cracks on the surface of epoxy resin, suppress irrelevant background noise, and enhance the recognition accuracy and generalization ability of the model, the receptive field convolution block attention (RFCBAMConv, Rceptive-Field Convolution Block Attention Module Conv) is used to replace part of the original network structure. The output network mainly outputs the location and classification features of the target through classification and regression.

[0083] In this embodiment, specifically, step S2 includes:

[0084] Step S21: randomly select an image in the training set as the input of the YOLOv8 network, and obtain feature maps of different scales after the backbone network layer extracts the features of the input image;

[0085] Step S22: The neck network layer fuses feature maps of different scales;

[0086] Step S23: predicting the probability and location of cracks in the image in the detection head, and generating a bounding box based on the prediction information;

[0087] Step S24: Iterate steps S21 to S23 until all training set images are traversed, calculate the YOLOv8 loss function in each iteration, and use the gradient descent method to update the network parameters to obtain an initial YOLOv8 model;

[0088] Step S25: Identify the validation set images according to the initial YOLOv8 model, iteratively count the average precision values ​​of the validation sets of different YOLOv8 models until the statistical mean average precision is stable, and obtain the trained YOLOv8 model. Use this model to identify all samples in the test set, obtain the surface cracking situation and position prediction box of the dry-type transformer in the test set image, and map it to the test set image.

[0089] In the iterative process, DIou is used as the positioning loss function of YOLOv8. DIoU takes into account the distance between the predicted box and the target box, not just IoU. This makes DIou loss more accurate for box positioning, especially when the boundaries are close but not overlapping. Its formula is:

[0090]

[0091] IoU is expressed as follows:

[0092]

[0093] Among them, b gt represents the true box at the time of prediction, B represents the predicted box, b and b gt Indicates B and B gtt The center point of , ρ(.) is the Euclidean distance, and c is the diagonal length.

[0094] In the iterative process, EIou is used as the YOLOv8 detection box regression loss function; EIOU further improves the distance metric between boxes by introducing embedded vectors, making it more accurate when evaluating the association between boxes and targets. Its calculation formula is:

[0095]

[0096] Among them, w and h are the width and height of the prediction box respectively, w gt 、h gt are the width and height of the real frame, c w 、c h They are the width and height of the minimum bounding rectangle of the target box and the predicted box, respectively.

[0097] In this embodiment, specifically, step S3 includes:

[0098] The acquisition point of the zoom industrial camera is selected at the upper end of the grille on both sides of the site. The image information of the camera is transmitted to the industrial computer through the CAN bus, and the YOLOv8 network model is used to judge whether the surface of the dry-type transformer is cracked and the crack position.

[0099] The acquisition point of the zoom industrial camera is selected at the upper end of the grille on both sides of the site. The distance between the transformer and the camera is appropriate, which is conducive to the camera to better collect image signals. The camera can be fixed on the grille mesh with a mounting bracket, which is convenient for disassembly and maintenance and the routing of the camera signal line. The installation diagram is as follows Figure 4 The image information of the camera is transmitted to the industrial computer through the CAN bus, and the improved YOLOv8 network model is used to judge whether the surface of the dry-type transformer is cracked and the crack position.

[0100] The 1500 images collected were tested by improving the YOLOv8 model, and the final results are as follows:

[0101] As an indicator to measure the difference between the model prediction results and the actual results, the smaller the loss function value, the closer the model prediction results are to the actual results, and the better the model performance. Figure 5a and Figure 5b As shown in the figure, in the training set, the loss function of the improved YOLOv8 decreases rapidly in the initial stage of model training. When the training reaches 60 rounds, the rate of decrease of the loss function slows down significantly, and the oscillation of the curve becomes more obvious, which is about 80 rounds earlier than the original YOLOv8 network. After 300 rounds of training, the loss function of the model tends to be stable. The bounding box loss, classification loss, and feature point loss in the training set finally stabilize below 0.3, 0.5, and 1.2, respectively. Among them, the bounding box loss decreased by 50% compared with the original YOLOv8 network, the classification loss did not change, and the feature point loss increased by 22%. Although the feature point loss increased slightly, the decrease in the bounding box loss was more obvious.

[0102] like Figure 6a and Figure 6b As shown in the figure, the oscillation of the model loss function of the YOLOv8-IRI network and YOLOv8 in the validation set is much higher than that in the training set. In the initial stage of model training, the YOLOv8-IRI network still has a higher rate of decline in the loss function than the YOLOv8 network. When the training reaches 60 rounds, the rate of decline of the loss function slows down significantly, and the oscillation of the curve becomes more obvious, which is 60 rounds earlier than the original YOLOv8 network. After 500 rounds of training, the loss function of the model in the validation set tends to be stable. In the validation set, the bounding box loss, classification loss, and feature point loss finally stabilized below 0.4, 0.9, and 1.3, respectively. Compared with the original YOLOv8 network, the bounding box loss decreased by 60%, the classification loss decreased by 10%, and the feature point loss increased by 8%, and the model oscillation has been significantly improved.

[0103] The final dry-type transformer epoxy resin surface crack identification results are as follows: Figure 7 shown.

[0104] The above-mentioned embodiments only express the specific implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the protection scope of the present application. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the technical solution concept of the present application, and these all belong to the protection scope of the present application.

[0105] This background section is provided to generally present the context of the invention, and the work of the presently named inventors, the work to the extent described in this background section, and aspects of this section that did not constitute prior art at the time of application are neither explicitly nor implicitly admitted to be prior art to the present invention.

Claims

1. A method for identifying cracks on the epoxy resin surface of a dry-type transformer, characterized in that: include: Step S1: constructing a dry-type transformer epoxy resin surface image dataset; Step S2: constructing a YOLOv8 network model for identifying cracks on the epoxy resin surface of a dry-type transformer; and training the YOLOv8 network model based on a dataset of epoxy resin surface images of a dry-type transformer; Step S3: input the image information collected by the camera into the YOLOv8 network model, and use the YOLOv8 network model to determine whether the surface of the dry-type transformer is cracked and the crack position.

2. A dry-type transformer epoxy resin surface crack identification method according to claim 1, characterized in that: The step S1 comprises: Step S11: collecting an image of the epoxy resin surface of a dry-type transformer; Step S12: cropping and enlarging the epoxy resin surface layer image of the dry-type transformer, and enlarging the crack position to form a sample set of the epoxy resin surface layer image of the dry-type transformer; Step S13: annotating the defect degree of the dry-type transformer epoxy resin surface image sample set based on the surface crack degree and position to generate a VOC2007 data set; Step S14: The VOC2007 dataset with surface cracks is expanded by using a GAN generative adversarial network written in Python to achieve data balance and make the model more accurate; Step S15: Based on the expanded data set, the data set is divided into a training set, a test set, and a validation set according to a preset ratio.

3. A dry-type transformer epoxy resin surface crack identification method according to claim 2, characterized in that: The step S13 comprises: During the marking process, whether there are cracks on the surface of the dry-type transformer and the location of the cracks are marked in an XML format file.

4. A dry-type transformer epoxy resin surface crack identification method according to claim 3, characterized in that: The YOLOv8 network model consists of four parts: input layer, backbone network layer, neck network layer and detection head; among which: The input layer is used to receive the original image as input and pass the image information to the subsequent network layers; The backbone network layer is used to extract image features and capture semantic information at different levels for subsequent target detection tasks. The neck network layer fuses features at different levels to enhance the model’s ability to detect objects of different scales. The detection head generates the results of target detection, including the location and category information of the target. The detection head part of the YOLOv8 network model uses prediction layers of different scales to adapt to targets of different sizes; each prediction layer is responsible for generating a set of bounding boxes and predicting the category probability, location and confidence of the target existence of each bounding box.

5. A dry-type transformer epoxy resin surface crack identification method according to claim 4, characterized in that: The step S2 comprises: Step S21: randomly select an image in the training set as the input of the YOLOv8 network, and obtain feature maps of different scales after the backbone network layer extracts the features of the input image; Step S22: The neck network layer fuses feature maps of different scales; Step S23: predicting the probability and location of cracks in the image in the detection head, and generating a bounding box based on the prediction information; Step S24: Iterate steps S21 to S23 until all training set images are traversed, calculate the YOLOv8 loss function in each iteration, use the gradient descent method to update the network parameters, and obtain the initial YOLOv8 model.

6. A dry-type transformer epoxy resin surface crack identification method according to claim 5, characterized in that: During the iteration process, DIou is used as the positioning loss function of YOLOv8.

7. A dry-type transformer epoxy resin surface crack identification method according to claim 6, characterized in that: During the iteration process, EIou is used as the YOLOv8 detection box regression loss function.

8. A dry-type transformer epoxy resin surface crack identification method according to claim 7, characterized in that: The step S2 further includes: Step S25: Identify the validation set images according to the initial YOLOv8 model, iteratively count the average precision values ​​of the validation sets of different YOLOv8 models until the statistical mean average precision is stable, and obtain the trained YOLOv8 model. Use this model to identify all samples in the test set, obtain the surface cracking situation and position prediction box of the dry-type transformer in the test set image, and map it to the test set image.

9. A dry-type transformer epoxy resin surface crack identification method according to claim 4, characterized in that: In the neck network layer, the receptive field convolution block attention RFCBAM Conv is used to replace the original network structure; the convolution block attention module is the basis of RFCBAM Conv, and the convolution block attention module consists of an independently running channel attention module and a spatial attention module; in RFCBAM Conv, Squeeze-and-Excitation Networks is used to replace the channel attention module in the convolution block attention module to reduce the amount of calculation; K with a step size of K is used 2 The convolution operation extracts feature information and performs weighted calculations on channel and spatial attention at the same time to further enhance RFCBAM Conv’s attention to the spatial features of the receptive field.

10. A dry-type transformer epoxy resin surface crack identification method according to claim 1, characterized in that: The step S3 comprises: The acquisition point of the zoom industrial camera is selected at the upper end of the grille on both sides of the site. The image information of the camera is transmitted to the industrial computer through the CAN bus, and the YOLOv8 network model is used to judge whether the surface of the dry-type transformer is cracked and the crack position.

Images