Exciting transformer winding and clamp defect detection method fusing global-local attention mechanism and slice reasoning

Abstract

The invention relates to an exciting transformer winding and clamp defect detection method fusing a global-local attention mechanism and slice reasoning, and belongs to the technical field of intelligent operation and maintenance of power equipment and computer vision. According to the method, for the problems of strong background interference, high missing detection rate of tiny defects, insufficient positioning precision and the like in the prior art, an improved YOLOv8 model is constructed, and an auxiliary bounding box is introduced to optimize a loss function to enhance the positioning capability of the irregular defects; a receptive field convolution block attention module and a reverse residual attention module are adopted to fuse local features and global dependence so as to suppress complex background interference, space-to-depth convolution is utilized to realize lossless downsampling and reserve fine texture features, and a slice-assisted super reasoning algorithm is combined to improve the detection performance of small targets in a high-resolution image. According to the method, the accuracy and the positioning precision of defect identification are remarkably improved, the omission ratio is effectively reduced, and reliable technical support is provided for intelligent operation and maintenance of the exciting transformer.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance of power equipment and computer vision technology, and relates to a method for detecting defects in excitation transformer windings and clamps that integrates global-local attention mechanism and slice reasoning. Background Technology

[0002] As a core component of the power generation system, the excitation transformer's operating status directly affects the safety and stability of the entire power grid. During long-term operation, its critical components are highly susceptible to physical damage due to electromagnetic vibration, thermal stress, and the external environment. In particular, core vibration transmitted to the clamps can easily cause loosening and displacement of the clamp nuts. Furthermore, under prolonged heat loads, the surface insulation layer of the windings is prone to aging, cracking, peeling, and creepage marks. These early defects in the windings and clamps are often precursors to serious internal equipment failures; therefore, accurate monitoring of the condition of the excitation transformer's critical components is crucial.

[0003] However, existing technologies still have significant shortcomings in the field of defect detection for key components of excitation transformers. Traditional detection methods mostly rely on manual periodic data recording and offline instrument testing, which suffers from problems such as discontinuous data, numerous monitoring blind spots, and inability to respond in real time to sudden faults. With the development of computer vision technology, general-purpose target detection algorithms, represented by YOLOv8, have been introduced into this field, but they face multiple severe challenges in the practical application of building fixed online monitoring systems.

[0004] First, the background interference characteristics at fixed monitoring locations are complex. Excitation transformers are typically installed in confined indoor spaces, and online monitoring cameras are usually fixed to the outside of security fences or on top brackets. In the actual acquired images, fixed metal grilles, cable shadows, and reflections from metal clamps highly obscure the features of the defects to be detected. Existing general network backbone layers lack targeted attention focusing mechanisms, making it difficult to effectively isolate minute insulation defects on the winding surface from the complex textures and lighting of a fixed perspective.

[0005] Secondly, subtle defect features are easily lost. Early insulation aging on the winding surface or slight displacement of clamps typically result in a very small pixel count and blurred texture, with low contrast against the background. The original YOLOv8 network frequently uses strided convolutions or pooling operations with a stride of two for downsampling during feature extraction. While this reduces computation, it inevitably leads to the loss of fine-grained texture information. Consequently, the model cannot retain enough effective pixel information when processing low-resolution or minute defects, resulting in missed detections.

[0006] Furthermore, small target detection is challenging. In wide-angle fixed monitoring images, the fault characteristics of excitation transformer components exhibit significant multi-scale features. In particular, small target components such as clamp nuts and pads located at the edges or far ends of the image often degenerate into a few pixels in the entire high-resolution monitoring image. The receptive field of general models has limitations in handling long-distance dependencies under such a fixed field of view, and directly scaling the image further blurs the features of small targets, resulting in insufficient ability of the model to capture weak defect signals.

[0007] Finally, sample quality and boundary regression are challenging. In industrial defect datasets, defects such as insulation peeling have highly irregular contours, resulting in inconsistent sample quality. The traditional Intersection over Union (IoU) loss function lacks an adaptive gradient adjustment mechanism when handling prediction boxes of different quality. This leads to slow convergence of the model during training when faced with indistinguishable samples, and the generated bounding boxes often fail to accurately fit the irregular defect contours, affecting the final quantitative evaluation accuracy. Summary of the Invention

[0008] In view of this, the purpose of this invention is to provide a method for detecting defects in excitation transformer windings and clamping parts that integrates global-local attention mechanism and slice reasoning.

[0009] To achieve the above objectives, the present invention provides the following technical solution: A method for detecting defects in excitation transformer windings and clamping components that integrates global-local attention mechanisms and slice reasoning includes the following steps: Acquire monitoring images of the excitation transformer; The monitored image was sliced ​​using the Slice Assisted Hyper Inference (SAHI) algorithm to obtain multiple overlapping slices; The multiple overlapping slices are input into the improved YOLOv8 model, which is constructed by introducing an Inner Intersection over Union (Inner-IoU) loss function, a Receptive Field Convolutional Block Attention Module (RFCBAMConv), an Inverted Residual Mobile Block (iRMB) attention module, and a Space to Depth Convolution (SPD-Conv) lossless downsampling module. Based on the output of the improved YOLOv8 model, the detection results are merged using a non-maximum suppression algorithm to output the defect location and type.

[0010] Furthermore, the introduction of the auxiliary bounding box optimization Inner-IoU loss function includes: During the model training phase, the scaling factor ratio is set according to the matching degree of the samples, where ratio is the scale control parameter of the auxiliary bounding box; For samples with high matching degree, set the ratio to be less than 1 to generate small auxiliary bounding boxes to accelerate convergence; For samples with low matching degree, set the ratio to be greater than 1 to generate a large auxiliary bounding box to enhance the regression effect; The calculation of the Inner-IoU loss function includes generating auxiliary bounding boxes and calculating intersection and union regions.

[0011] Furthermore, the introduction of the receptive field convolutional block attention module RFCBAMConv includes: In the backbone network of the YOLOv8 model, some convolutional layers are replaced with the RFCBAMConv module. The RFCBAMConv module calculates channel weights through the SE module and utilizes... K 2 Convolution operations generate spatial weights, and the RFCBAMConv module integrates a multi-head attention mechanism, the calculation formula of which is as follows: For each head n , n = 1,2 ,..., N ,have ,and ,in Q For querying the matrix, K The key matrix, V For value matrices, d For the dimension of the key, N For the number of heads, W This is the weight matrix.

[0012] Furthermore, the introduction of the inverse residual attention module iRMB includes: In the feature extraction stage of the YOLOv8 model, the iRMB module is integrated. The iRMB module extracts local static features through deep convolution and captures global dynamic features through expanded window multi-head self-attention (EW-MHSA). EW-MHSA calculates long-range dependencies based on the multi-head attention mechanism.

[0013] Furthermore, the introduction of the spatial-to-depth convolutional SPD-Conv lossless downsampling module includes: In the downsampling stage of the backbone and neck network of the YOLOv8 model, the SPD-Conv module is used to replace the traditional strided convolution. The SPD-Conv module samples and divides the input feature map into multiple sub-maps at row and column intervals and concatenates them in the channel dimension, and then fuses them through a standard convolution with a stride of 1.

[0014] Furthermore, the slicing process of the Slice-Assisted Hyper-Inference Algorithm (SAHI) includes: The high-resolution monitoring image is divided into multiple overlapping slices according to a set step size; Perform inference detection independently on each slice and map the results back to the coordinates of the original image.

[0015] Furthermore, the merged detection results are processed to remove redundancy from the detection boxes of all slices using a non-maximum suppression algorithm.

[0016] Furthermore, the training of the improved YOLOv8 model includes augmenting the training samples using Mosaic data augmentation techniques, which include random cropping, scaling, and color dithering.

[0017] Furthermore, the method also includes preprocessing the monitoring images, which includes manually annotating defect areas and converting them into VOC format datasets.

[0018] Furthermore, the defect types include winding insulation aging, loose clamp nuts, and metallic foreign objects.

[0019] The beneficial effects of this invention are as follows: (1) In terms of detection accuracy, this invention, by integrating a global-local attention mechanism, can accurately identify minute insulation defects in the windings of the excitation transformer and minor loosening features of the clamps from complex backgrounds. The model has a strong ability to suppress interference factors such as metal grid obstruction and reflection, which significantly improves the accuracy and reliability of defect identification.

[0020] (2) In terms of localization capability, the introduced auxiliary bounding box optimization strategy enables the model to adaptively adjust the bounding box regression range. For defects with irregular contours, such as insulation layer peeling, the generated detection box can fit more closely to its real edge, achieving sub-pixel level localization accuracy, which provides a solid foundation for subsequent quantitative evaluation.

[0021] (3) To address the problem of missing small targets, the application of the slice-assisted super inference algorithm effectively solves the problem of feature loss of far-end clamping components in high-resolution images. By processing and fusing the image blocks, the model can capture minute anomalies such as loose nuts and displacement of pads that are difficult to detect by traditional methods, thus achieving all-round monitoring of key components of transformers without blind spots.

[0022] (4) At the feature preservation level, the lossless downsampling module constructed by spatial-to-depth convolution completely changes the information loss problem caused by traditional downsampling methods. This module preserves the pixel-level information of the fine cracks on the winding surface to the greatest extent, ensuring the effective transmission of low-contrast defect features in the deep network and enhancing the model's sensitivity to early defects.

[0023] (5) The model’s anti-interference ability benefits from the synergistic effect of the receptive field convolutional block attention module and the inverse residual attention module. The former enhances the ability to focus on spatial features through large kernel convolution, while the latter combines the local feature extraction advantages of convolutional neural networks with the global perception capability of the Transformer architecture, enabling the model to maintain stable detection performance in real industrial scenarios with uneven lighting and interlaced shadows.

[0024] (6) In terms of training efficiency, the adaptive loss function mechanism dynamically adjusts the optimization direction according to the sample quality, which accelerates the model convergence process and improves training stability. This enables the model to quickly adapt to sample data of different qualities, shortens the training cycle, and is more suitable for the rapid deployment needs of engineering sites.

[0025] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0026] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a diagram of the overall structure of the improved YOLOv8 network. Figure 2 A schematic diagram illustrating the Inner-IoU principle of auxiliary bounding boxes; Figure 3 Here is a structural diagram of the RFCBAMConv module; Figure 4 This is a structural diagram of the inverse residual attention module iRMB. Detailed Implementation

[0027] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0028] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0029] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0030] Example 1: Overall Flow of a Defect Detection System Based on an Improved YOLOv8 Model This embodiment illustrates the overall implementation process of the method described in this invention. For example... Figure 1 As shown, the improved YOLOv8 network structure consists of four parts: a data input layer, a feature extraction backbone network, a feature fusion neck network, and a detection head.

[0031] The specific implementation steps are as follows: First, images of the excitation transformer at a hydropower station were collected during operation. The LabelImg tool was used to label defects such as aging winding insulation, loose clamp nuts, and foreign metal objects, generating a VOC format dataset. Subsequently, Mosaic data augmentation techniques were used to expand the training samples, including random cropping, scaling, and color dithering operations.

[0032] In the feature extraction stage, the spatial-to-depth convolutional SPD-Conv module replaces the strided convolutional layers in the original YOLOv8 network. This module samples and divides the input feature map into four sub-images at row and column intervals, concatenates them along the channel dimension, and then fuses them through a standard convolutional layer with a stride of one, achieving lossless downsampling.

[0033] In the middle of the backbone network, the original convolutional layers are replaced with the receptive field convolutional block attention module RFCBAMConv. For example... Figure 3 As shown, the RFCBAMConv module first calculates the channel weights through the SE module, and then uses K-squared convolution to generate spatial attention weights to weight the feature map.

[0034] In the neck network portion, an inverse residual attention module (iRMB) is introduced. For example... Figure 4 As shown, this module extracts local static features through deep convolution and then captures global dynamic features through an extended window multi-head self-attention mechanism, effectively expanding the receptive field.

[0035] During the training phase, an auxiliary bounding box optimization strategy is employed. For example... Figure 2 As shown, a scaling factor ratio less than one is set for high-matching samples to generate small auxiliary bounding boxes, while a scaling factor greater than one is set for low-matching samples to generate large auxiliary bounding boxes. This adaptive adjustment improves the bounding box regression accuracy.

[0036] Example 2: Specific Implementation of the Slice-Assisted Hyper-Inference Algorithm This embodiment focuses on illustrating the specific implementation of the Slice-Assisted Hyperinference (SAHI) algorithm. This embodiment corresponds to the slice processing technology solution in the claims.

[0037] The implementation environment is configured as follows: a high-performance deep learning server with no fewer than fourteen processor cores, more than 16GB of video memory, Ubuntu 20.04 operating system, and PyTorch deep learning framework.

[0038] The specific workflow includes: First, the high-resolution monitoring image of 1920×1080 pixels is divided into multiple overlapping slices of 512×512 pixels with a step size of 256 pixels. Each slice is independently input into the trained improved YOLOv8 model for inference to obtain local detection results.

[0039] The detection results coordinates of each slice are then mapped back to the absolute coordinate system of the original image. The specific mapping formula is: x_original = x_slice × stride_x + offset_x, where x_original is the x-coordinate of the original image, x_slice is the x-coordinate within the slice, stride_x is the horizontal segmentation step size, and offset_x is the horizontal offset of the slice in the original image.

[0040] Finally, a non-maximum suppression algorithm was used to remove redundancy from the detection boxes of all slices. An overlap threshold of 0.5 was set to retain the detection boxes with the highest confidence, thus eliminating duplicate detections.

[0041] Example 3: Training Implementation Process of Multi-Module Collaborative Optimization This embodiment details the collaborative training process of each improved module, covering multiple technical solutions in the claims.

[0042] The training parameters were set as follows: a base learning rate of 0.01, a cosine annealing scheduling strategy, a batch size of 16, and 150 training epochs. The optimizer used was stochastic gradient descent with a momentum parameter of 0.937 and a weight decay coefficient of 0.0005.

[0043] During training, a multi-task loss function is used for loss function calculation, which includes bounding box regression loss, confidence loss, and classification loss. Among them, the bounding box regression loss adopts an improved Inner-IoU loss function, which dynamically adjusts the size of the auxiliary bounding box through a scaling factor.

[0044] The forward propagation process of the feature extraction network sequentially passes through the SPD-Conv module, the RFCBAMConv module, and the iRMB module. The SPD-Conv module performs lossless downsampling, the RFCBAMConv module enhances receptive field attention, and the iRMB module fuses local and global features. These three modules work together to significantly improve the feature extraction capability for minute defects.

[0045] Experimental results on the validation set show that, while maintaining real-time detection speed, this implementation significantly improves the detection accuracy of minor defects such as winding insulation aging, and the bounding box positioning is more accurate.

[0046] To verify the effectiveness of the multi-strategy improved YOLOv8 network proposed in this invention, this embodiment constructs a dataset based on actual collected images of the excitation transformer's appearance and trains and tests it on a deep learning workstation.

[0047] 1. Experimental Environment and Dataset Construction Data source: Data was collected from the on-site operating environment of an excitation transformer at a hydropower station, covering normal conditions as well as various typical physical faults such as scratches, insulation peeling (patches), and loose components.

[0048] Data preprocessing: The acquired images were manually labeled using the LabelImg tool to generate XML label files, which were then converted into VOC format datasets. To improve the model's generalization ability, Mosaic data augmentation techniques were used to expand the training samples, including random cropping, scaling, and color jittering.

[0049] Hardware environment: The experimental platform uses a high-performance deep learning server with the following configuration: processor cores ≥14, graphics memory >16G, operating system is Ubuntu 20.04, and deep learning framework is PyTorch.

[0050] 2. Baseline Model and Parameter Settings This invention selects the lightweight YOLOv8n as the baseline model for improvement, balancing detection accuracy and real-time performance. The parameter configuration of the basic model is as follows: Table 1 Parameter configuration of YOLOv8n model

[0051] 3. Implementation of the core improved algorithm To overcome the shortcomings of existing technologies in detecting complex background interference and small targets, this invention incorporates five-dimensional improvement strategies based on YOLOv8n. The following is the key implementation logic (pseudocode) of the core module: (1) Logic for calculating loss in the auxiliary bounding box (Inner-IoU) To address the issue of slow training convergence, a scaling factor (ratio) is introduced to control the size of the auxiliary bounding box. The following is the Inner-IoU calculation logic based on PyTorch style: Table 2. Pseudocode for Inner-IoU Loss

[0052] (2) Implementation logic of the perception field convolutional block attention (RFCBAMConv) Using K 2 Large kernel convolution enhances feature extraction for minute textures (such as scratches): Table 3. Pseudocode for Receptive Field Convolutional Block Attention (RFCBAMConv)

[0053] (3) Implementation logic of the inverse residual attention module (iRMB) By fusing local features from CNNs with global features from Transformers, we can improve our robustness against interference. Table 4. Pseudocode for the Inverse Residual Attention Module iRMB

[0054] (4) Lossless downsampling logic of spatial-to-depth convolution (SPD-Conv) To replace traditional strided convolutions and prevent the loss of feature information from minute defects: Table 5. Pseudocode for lossless downsampling from spatial to depthwise convolution (SPD-Conv)

[0055] (5) Slice-Assisted Hyper-Inference (SAHI) Detection Logic Slice inference and result fusion for small targets in high-resolution images: Table 6. Pseudocode for Slice-Assisted Hyper-Inference (SAHI) Detection

[0056] 4. Experimental Results and Analysis of Beneficial Effects The improved model proposed in this invention was compared with the original YOLOv8n model on the same test set. The results show that the proposed method has significant advantages in the following aspects: Convergence speed and stability: After introducing the Inner-IoU auxiliary bounding box, the model's loss function (Box Loss) decreases at a significantly faster rate than the original CIoU loss in the early stages of training. Experimental data show that the improved model exhibits a stable convergence trend after about 30 iterations, and the fluctuation of the validation set loss (val / box_loss) is significantly reduced, demonstrating the improvement effect of the improved loss function on localization accuracy.

[0057] Minor Defect Detection Performance: Thanks to the introduction of the SPD-Conv non-destructive downsampling module and the SAHI slicing inference algorithm, this method effectively solves the problem of missing minor scratches on the transformer winding surface and loose clamp nuts. Under complex lighting and grid background interference, the RFCBAMConv and iRMB modules successfully suppressed background noise, enabling the model to accurately focus on the defect area.

[0058] Overall performance indicators: Compared with the original YOLOv8n model, the method of this invention has significantly improved the average accuracy and greatly reduced the false negative rate while maintaining real-time detection capability (FPS≥30), which can meet the actual needs of all-weather unattended online monitoring of excitation transformers.

[0059] This invention focuses on the construction of online intelligent monitoring technology for excitation transformer windings and clamps, solving problems such as strong background interference, difficulty in identifying minor component defects, and insufficient model positioning accuracy in fixed monitoring scenarios. Through network structure reconstruction and inference strategy optimization, this invention significantly improves the real-time performance and accuracy of defect detection, providing solid technical support for fault early warning, condition assessment, and intelligent operation and maintenance of excitation transformers.

[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting defects in excitation transformer windings and clamping components that integrates global-local attention mechanisms and slice reasoning, characterized in that: Includes the following steps: Acquire monitoring images of the excitation transformer; The monitoring image was sliced ​​using the Slice-Assisted Hyper-Inference (SAHI) algorithm to obtain multiple overlapping slices; The multiple overlapping slices are input into the improved YOLOv8 model, which is constructed by introducing an auxiliary bounding box to optimize the Inner-IoU loss function, a receptive field convolutional block attention module RFCBAMConv, an inverse residual attention module iRMB, and a spatial-to-depth convolutional SPD-Conv lossless downsampling module. Based on the output of the improved YOLOv8 model, the detection results are merged using a non-maximum suppression algorithm to output the defect location and type.

2. The method for detecting defects in excitation transformer windings and clamping components by integrating global-local attention mechanism and slice reasoning as described in claim 1, characterized in that: The introduction of the auxiliary bounding box optimization Inner-IoU loss function includes: During the model training phase, the scaling factor ratio is set according to the matching degree of the samples, where ratio is the scale control parameter of the auxiliary bounding box; For samples with high matching degree, set the ratio to be less than 1 to generate small auxiliary bounding boxes to accelerate convergence; For samples with low matching degree, set the ratio to be greater than 1 to generate a large auxiliary bounding box to enhance the regression effect; The calculation of the Inner-IoU loss function includes generating auxiliary bounding boxes and calculating intersection and union regions.

3. The method for detecting defects in excitation transformer windings and clamping components by integrating global-local attention mechanism and slice reasoning as described in claim 1, characterized in that: The introduction of the receptive field convolutional block attention module RFCBAMConv includes: In the backbone network of the YOLOv8 model, some convolutional layers are replaced with the RFCBAMConv module. The RFCBAMConv module calculates channel weights through the SE module and utilizes... K 2 Convolution operations generate spatial weights, and the RFCBAMConv module integrates a multi-head attention mechanism, the calculation formula of which is as follows: For each head n , n = 1,2 ,..., N ,have ,and ,in Q For querying the matrix, K The key matrix, V For value matrices, d For the dimension of the key, N For the number of heads, W This is the weight matrix.

4. The method for detecting defects in excitation transformer windings and clamping components by integrating global-local attention mechanism and slice reasoning as described in claim 1, characterized in that: The introduction of the inverse residual attention module iRMB includes: In the feature extraction stage of the YOLOv8 model, the iRMB module is integrated. The iRMB module extracts local static features through deep convolution and captures global dynamic features through extended window multi-head self-attention EW-MHSA, where EW-MHSA is based on the multi-head attention mechanism to calculate long-range dependencies.

5. The method for detecting defects in excitation transformer windings and clamping components by integrating global-local attention mechanism and slice reasoning as described in claim 1, characterized in that: The introduction of the spatial-to-depth convolutional SPD-Conv lossless downsampling module includes: In the downsampling stage of the backbone and neck network of the YOLOv8 model, the SPD-Conv module is used to replace the traditional strided convolution. The SPD-Conv module samples and divides the input feature map into multiple sub-maps at row and column intervals and concatenates them in the channel dimension, and then fuses them through a standard convolution with a stride of 1.

6. The method for detecting defects in excitation transformer windings and clamping components by integrating global-local attention mechanism and slice reasoning as described in claim 1, characterized in that: The slice processing of the Slice-Assisted Hyper-Inference Algorithm (SAHI) includes: The high-resolution monitoring image is divided into multiple overlapping slices according to a set step size; Perform inference detection independently on each slice and map the results back to the coordinates of the original image.

7. The method for detecting defects in excitation transformer windings and clamping components by integrating global-local attention mechanism and slice reasoning as described in claim 6, characterized in that: The merged detection results are processed using a non-maximum suppression algorithm to remove redundancy from the detection boxes of all slices.

8. The method for detecting defects in excitation transformer windings and clamping components by integrating global-local attention mechanism and slice reasoning as described in claim 1, characterized in that: The training of the improved YOLOv8 model includes augmenting the training samples using Mosaic data augmentation techniques, which include random cropping, scaling, and color dithering.

9. The method for detecting defects in excitation transformer windings and clamping components by integrating global-local attention mechanism and slice reasoning as described in claim 1, characterized in that: The method also includes preprocessing the monitoring images, which includes manually annotating defect areas and converting them into VOC format datasets.

10. The method for detecting defects in excitation transformer windings and clamping components by integrating global-local attention mechanism and slice reasoning according to claim 1, characterized in that: The defect types include aging of winding insulation, loose clamp nuts, and foreign metal objects.

Images