YOLO-SM-based method for detecting ultraviolet corona discharge targets in insulators

Abstract

This invention relates to the fields of artificial intelligence and power equipment inspection technology, specifically a method for detecting ultraviolet corona discharge targets on insulators based on YOLO-SM. The invention includes constructing a feature extraction network integrating a lightweight SCCDG feature extraction module, a feature fusion network integrating a multi-scale dual attention mechanism module (MSDAM), and a prediction network, trained using a composite bounding box regression loss. The SCCDG lightweight feature extraction module averages the input channels and feeds the sub-features into CDG multi-path convolution to extract multi-scale features; the MSDAM multi-scale dual attention mechanism module generates attention weights and performs weighted fusion; the prediction network uses a multi-scale detection head, and outputs the detection results after confidence thresholding and non-maximum suppression. This invention balances detection accuracy and computational efficiency, is suitable for real-time monitoring scenarios such as UAV inspections, and is beneficial for improving the detection rate and positioning accuracy of small corona discharge targets.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and power equipment testing technology, specifically to a method for detecting ultraviolet corona discharge targets in insulators based on YOLO-SM. Background Technology

[0002] Insulators are critical components of power transmission lines, and their ultraviolet corona discharge is a significant precursor to line faults. Timely and accurate detection of corona discharge is crucial for ensuring the safe and stable operation of power systems. With the development of artificial intelligence technology, deep learning-based target detection methods have provided a new technical approach for the automatic identification of corona discharge.

[0003] Existing corona discharge detection methods suffer from numerous technical limitations: traditional YOLO algorithms lack sufficient feature extraction capabilities when handling corona discharge from small targets, resulting in low detection accuracy; while existing lightweight networks offer high computational efficiency, their ability to identify weak corona signals in complex backgrounds is limited, easily leading to missed detections and false detections; existing attention mechanisms often employ a single-scale design, failing to effectively integrate multi-scale feature information and exhibiting poor adaptability to irregularly shaped corona discharge targets; traditional bounding box regression loss functions, primarily based on IoU calculations, struggle to accurately fit the irregular boundaries of corona discharge targets, resulting in low localization accuracy. These technical limitations severely restrict the application effectiveness of corona discharge detection systems in practical engineering. Summary of the Invention

[0004] This invention provides a YOLO-SM-based method for detecting ultraviolet corona discharge targets in insulators, aiming to improve the detection accuracy of small targets, suppress background noise, optimize positioning loss, reduce computational overhead, and support real-time inspection deployment.

[0005] A YOLO-SM-based method for detecting ultraviolet corona discharge targets in insulators includes:

[0006] S100: Obtain the trained YOLO-SM detection network, which includes a feature extraction network, a feature fusion network, and a prediction network, wherein:

[0007] The feature extraction network integrates a lightweight SCCDG feature extraction module. The lightweight SCCDG feature extraction module separates multiple sub-features through average channel separation operation, and performs three parallel processing steps on the input features through the CDG module: ordinary convolution, dilated convolution, and group convolution.

[0008] The feature fusion network integrates the MSDAM multi-scale dual attention mechanism module, which includes two parallel branches, each using a pooling kernel of different sizes, and processed by Mish and Swish activation functions to generate complementary attention weights.

[0009] The YOLO-SM detection network is trained using a composite bounding box regression loss function;

[0010] S200: Input the ultraviolet corona discharge image of the insulator into the YOLO-SM detection network, and extract multi-scale features by performing channel splitting and multi-branch convolution processing on the input features through the SCCDG lightweight feature extraction module integrated in the feature extraction network;

[0011] S300: Input the multi-scale features into the feature fusion network, perform pooling operations at different scales through the dual-branch structure of the MSDAM multi-scale dual attention mechanism module to generate attention weights, and perform weighted fusion of the multi-scale features to obtain fused features;

[0012] S400: Input the fused features into the prediction network for target detection, and output the bounding box coordinates, category label, and confidence score of the corona discharge target.

[0013] As a preferred embodiment of the present invention, the composite bounding box regression loss function includes a weighted sum of the following four losses:

[0014] The intersection-union ratio loss is calculated using the loss value based on the overlap between the predicted bounding box and the target bounding box;

[0015] The shape deviation loss is calculated based on the aspect ratio difference between the predicted bounding box and the target bounding box;

[0016] Corner distance loss is calculated by the distance between the corner coordinates of the predicted bounding box and the target bounding box;

[0017] Density-weighted loss sets a weighting factor based on the target distribution density.

[0018] As a preferred embodiment of the present invention, the feature extraction network includes a convolutional module and a lightweight SCCDG feature extraction module. The convolutional module consists of two-dimensional convolution, batch normalization, and activation functions. The outputs of multiple lightweight SCCDG feature extraction modules are connected to a feature fusion network to provide multi-level features.

[0019] As a preferred embodiment of the present invention, the SCCDG lightweight feature extraction module includes:

[0020] The input features are separated into multiple sub-features through the average channel separation operation. The sub-features are then stacked after convolution. The stacked features are then input into the CDG module for processing. The output of the CDG module forms a residual connection with the stacked features. Finally, the processed features are output after convolution.

[0021] The CDG module performs three-way convolution processing on the input features in parallel: the first path performs ordinary convolution, the second path performs dilated convolution, and the third path performs grouped convolution. The results of the three convolutions are stacked by channels, and after convolution, normalization and activation function processing, they are further processed by convolution and pooling operations, and then connected with the intermediate processing results to form residual connections.

[0022] As a preferred technical solution of the present invention, the feature fusion network includes an upsampling module, a channel overlay module, a convolution module, a SCCDG lightweight feature extraction module, and an MSDAM multi-scale dual attention mechanism module. The feature scale is adjusted through upsampling and downsampling operations, and feature stitching is achieved through channel overlay. Multiple MSDAM multi-scale dual attention mechanism modules are respectively connected to detection heads of different scales.

[0023] As a preferred technical solution of the present invention, the MSDAM multi-scale dual attention mechanism module includes: after preprocessing the input features, two parallel branches are input respectively; the first branch is processed by a first activation function after performing average pooling and max pooling operations at the first scale; the second branch is processed by a second activation function after performing average pooling and max pooling operations at the second scale; the two branches are respectively connected with the input features to form residual connections to generate attention weights; and finally, weighted fusion is performed.

[0024] As a preferred embodiment of the present invention, the first branch and the second branch are pooled using pooling kernels of different sizes. The first branch is pooled using the Mish activation function, and the second branch is pooled using the Swish activation function, resulting in different attention weight features between the two branches.

[0025] As a preferred embodiment of the present invention, the prediction network includes three detection heads, each of which includes a bounding box regression branch, a category classification branch, and a confidence prediction branch. The three detection heads respectively handle large-scale, medium-scale, and small-scale target detection tasks.

[0026] As a preferred technical solution of the present invention, the output results of the three detection heads are merged and then post-processed, including setting a confidence threshold to filter low-confidence detection boxes, using a non-maximum suppression algorithm to remove overlapping redundant detection boxes through an IoU threshold, and outputting the final corona discharge target detection result.

[0027] The beneficial effects of this invention are:

[0028] 1. This invention achieves an optimal balance between feature extraction capability and computational efficiency through the channel separation-multi-branch convolution design of the SCCDG lightweight feature extraction module and the dual-branch architecture of the MSDAM multi-scale dual attention mechanism module. The SCCDG lightweight feature extraction module uses three parallel processing paths—ordinary convolution, dilated convolution, and grouped convolution—to capture local details, expand the receptive field, and reduce computational complexity. The MSDAM multi-scale dual attention mechanism module uses dual-branch processing with pooling kernels of different scales and different activation functions to achieve complementary fusion of local detail features and global contextual information, significantly enhancing the ability to identify weak corona signals in complex backgrounds.

[0029] 2. This invention integrates four dimensions—cross-union ratio loss, shape deviation loss, corner distance loss, and density-weighted loss—to form a composite bounding box regression loss function, systematically solving the detection challenge of irregularly shaped and blurred-boundary corona discharge targets. This loss function optimizes aspect ratio fitting through shape deviation loss, achieves refined boundary localization through corner distance loss, and ensures balanced learning in dense regions through density-weighted loss. Together with the aforementioned modules, it forms an end-to-end collaborative optimization system, effectively overcoming the technical bottlenecks of low detection accuracy and inaccurate localization of irregular corona discharge targets using traditional methods. Attached Figure Description

[0030] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0031] Figure 1 This is a flowchart illustrating the insulator ultraviolet corona discharge target detection method based on YOLO-SM of the present invention;

[0032] Figure 2 This is the YOLO-SM network model structure of the present invention;

[0033] Figure 3 This is a schematic diagram of the SCCDG lightweight feature extraction module in the YOLO-SM-based insulator ultraviolet corona discharge target detection method of the present invention;

[0034] Figure 4 This is a schematic diagram of the CDG structure in the YOLO-SM-based insulator ultraviolet corona discharge target detection method of the present invention;

[0035] Figure 5 This is a schematic diagram of the MSDAM multi-scale dual attention mechanism module in the YOLO-SM-based insulator ultraviolet corona discharge target detection method of the present invention. Detailed Implementation

[0036] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0037] Example 1: As Figure 1 As shown, the present invention provides a method for detecting ultraviolet corona discharge targets in insulators based on YOLO-SM, comprising:

[0038] S100: Obtain the trained YOLO-SM detection network, which includes a feature extraction network, a feature fusion network, and a prediction network, wherein:

[0039] The feature extraction network integrates a lightweight SCCDG feature extraction module. The lightweight SCCDG feature extraction module separates multiple sub-features through average channel separation operation, and performs three parallel processing steps on the input features through the CDG module: ordinary convolution, dilated convolution, and group convolution.

[0040] The feature fusion network integrates the MSDAM multi-scale dual attention mechanism module, which includes two parallel branches, each using a pooling kernel of different sizes, and processed by Mish and Swish activation functions to generate complementary attention weights.

[0041] The YOLO-SM detection network is trained using a composite bounding box regression loss function;

[0042] Furthermore, the composite bounding box regression loss function comprises a weighted sum of the following four losses:

[0043] The intersection-union ratio loss is calculated using the loss value based on the overlap between the predicted bounding box and the target bounding box;

[0044] The shape deviation loss is calculated based on the aspect ratio difference between the predicted bounding box and the target bounding box;

[0045] Corner distance loss is calculated by the distance between the corner coordinates of the predicted bounding box and the target bounding box;

[0046] Density-weighted loss sets a weighting factor based on the target distribution density.

[0047] Specifically, obtain the trained YOLO-SM detection network, such as Figure 2As shown, this network employs an improved YOLO architecture, specifically optimized for the detection of small targets in insulator ultraviolet corona discharge. The overall structure of the YOLO-SM detection network consists of three main parts: a feature extraction network, a feature fusion network, and a prediction network. The feature extraction network is responsible for extracting multi-level depth features from the input 640×640×3 ultraviolet corona discharge image. It includes 5 ordinary convolutional modules and 12 SCCDG lightweight feature extraction modules, progressively reducing the input image from 640×640 to 20×20 through layer-by-layer downsampling, while increasing the number of channels from 3 to 1024. The feature fusion network achieves multi-scale feature fusion through upsampling, downsampling, and channel concatenation operations, integrating the MSDAM multi-scale dual attention mechanism module to enhance the feature representation of the target region. The prediction network contains three detection heads of different scales, used to detect large, medium, and small-scale corona discharge targets, respectively.

[0048] The YOLO-SM detection network is trained using a specially designed composite bounding box regression loss function, optimized for the irregular shapes and blurred boundaries of corona discharge targets. The composite bounding box regression loss function consists of a weighted sum of four loss terms:

[0049] The first term is the Cross-Union Ratio (IoU) loss. IoU is a core metric for evaluating the overlap of bounding boxes, but for irregularly shaped targets, using only IoU loss is insufficient to optimize the model. To improve accuracy, 1-IoU is introduced as a loss term, encouraging the model to reduce localization errors by increasing the overlap of bounding boxes. During training, the loss function prioritizes optimizing the overlap between the predicted and target boxes, ensuring that the predicted boxes surround the target as much as possible. The 1-IoU formula is as follows:

[0050] ;

[0051] in: The intersection-union ratio (IoU) loss value ranges from [0,1]. 1-IoU represents the loss value for the overlap between the predicted bounding box and the target bounding box. The larger the loss value, the smaller the overlap between the predicted bounding box and the target bounding box. To predict the area of ​​the frame; The area of ​​the target bounding box; This represents the area of ​​the overlapping region between the predicted bounding box and the target bounding box; This represents the total area of ​​the merged region between the predicted bounding box and the target bounding box.

[0052] when When the predicted bounding box completely overlaps with the target bounding box, it indicates that the predicted bounding box and the target bounding box are completely aligned; when... A value of 1 indicates that the two boxes do not overlap at all. During backpropagation, the gradient of this loss term drives the predicted box to move towards the target box and resize to maximize the overlap area.

[0053] In deep learning training, the optimizer aims to minimize the loss function. If IoU is used directly as the loss, the larger the IoU (the higher the overlap), the greater the loss, and the optimizer will incorrectly reduce the overlap. By using 1-IoU, as the IoU increases, 1-IoU decreases, which conforms to the optimization logic of "reducing the loss means improving performance".

[0054] The second term is shape deviation loss. Corona discharge targets often exhibit irregular shapes, and their aspect ratio is an important morphological feature. Shape deviation loss, by constraining the difference in aspect ratio between the predicted bounding box and the target bounding box, enables the model to more accurately capture the target's morphological features. The calculation formula is:

[0055] ;

[0056] in: This represents the shape deviation loss value. The width of the prediction box; The height of the predicted bounding box; The width of the target bounding box; The height of the target bounding box; This represents the number of samples, i.e., the number of bounding boxes.

[0057] Shape deviation loss and cross-union ratio loss complement each other: cross-union ratio loss focuses on the overall overlap and may allow for cases with large shape differences but still high IoU; shape deviation loss strictly constrains the aspect ratio to ensure that the predicted box is not only correctly positioned but also has a shape consistent with the target, which is especially important for the accurate detection of irregular corona discharge targets.

[0058] The third term is corner distance loss. Traditional bounding box regression loss struggles to accurately capture the coordinates of the four corners of the target box, especially when dealing with complex shapes. Corner distance loss calculates the distance difference between the predicted box and the target box's corners, enabling finer localization adjustments and improving bounding box prediction accuracy. It is suitable for asymmetrical or angularly complex scenes, effectively enhancing the model's ability to perceive target boundary details and improving overall detection performance. The formula is as follows:

[0059] ;

[0060] in: This represents the corner distance loss value. The coordinates of the top-left corner of the prediction box; The coordinates of the top-left corner of the target bounding box; The coordinates of the bottom right corner of the prediction box; The coordinates of the bottom right corner of the target bounding box; superscript Indicates top-left (top left corner); superscript This indicates bottom-right (lower right corner); This represents the number of samples, i.e., the number of bounding boxes.

[0061] This loss term directly optimizes corner positions, making it particularly effective for asymmetrical or tilted targets. Compared to methods that only optimize the center point and width / height (such as the intersection-union loss), the corner distance loss can more accurately constrain the spatial position of the bounding box. For example, two predicted boxes with the same IoU may have different corner deviations: one may be translated overall, while the other may be distorted. The corner distance loss can distinguish between these two cases, ensuring that all four corners are accurately aligned, which is crucial for detecting corona discharge targets with blurred boundaries and complex shapes.

[0062] The fourth term is density-weighted loss. In dense target scenes, models tend to overemphasize isolated targets while neglecting details in dense regions. Density-weighted loss achieves balanced optimization during training by assigning a weight factor to each target based on the density of its surrounding targets. The calculation formula is:

[0063] ;

[0064] in: This is the density-weighted loss value; : This is a weighting factor associated with the target box, which depends on the density of the target box; regions with higher density receive higher weights. : The center coordinates of the prediction box; : The center coordinates of the target bounding box; : The width of the target bounding box; : is the height of the target bounding box; : This represents the number of samples, i.e., the number of bounding boxes.

[0065] The density weighting factor ensures that targets in dense regions receive higher learning weights. For example, an isolated target has a weight of 1.0, while a target in a dense region with five neighboring targets has a weight of 1.5, thus balancing the learning effects of regions with different distribution densities. This design increases the loss contribution of targets in dense regions by 50%, forcing the model to pay more attention to them during training, thereby balancing the learning effects of regions with different distribution densities. In scenarios where multiple strings of insulators are densely arranged in transmission lines, this loss term can effectively prevent the model from ignoring corona discharge targets in dense regions, significantly reducing the false negative rate.

[0066] The four loss terms are weighted and summed to form the final composite bounding box regression loss function:

[0067] ;

[0068] in, The total loss for the composite bounding box regression; The weighting factor is preferably 0.4, 0.2, 0.2, or 0.2, used to balance the contribution of each loss term.

[0069] The synergistic mechanism of each loss term: the Cross-Union-Ratio (CIRR) loss provides an overall optimization direction, ensuring that the predicted bounding box moves closer to the target bounding box; the Shape Deviation loss further constrains the aspect ratio based on the CIRR loss to prevent shape distortion; the Corner Distance loss achieves fine-grained localization, ensuring accurate alignment of the four corner points; and the Density Weighted Loss balances the learning of regions with different densities, preventing training bias. This composite loss function simultaneously optimizes position, shape, corner points, and density during training, forming a synergistic optimization mechanism that significantly improves the detection accuracy of irregular corona discharge targets.

[0070] S200: Input the ultraviolet corona discharge image of the insulator into the YOLO-SM detection network, and extract multi-scale features by performing channel splitting and multi-branch convolution processing on the input features through the SCCDG lightweight feature extraction module integrated in the feature extraction network;

[0071] Furthermore, the feature extraction network includes a convolutional module and a lightweight SCCDG feature extraction module. The convolutional module consists of two-dimensional convolution, batch normalization, and activation functions. The outputs of multiple lightweight SCCDG feature extraction modules are connected to a feature fusion network to provide multi-level features.

[0072] Furthermore, the SCCDG lightweight feature extraction module includes:

[0073] The input features are separated into multiple sub-features through the average channel separation operation. The sub-features are then stacked after convolution. The stacked features are then input into the CDG module for processing. The output of the CDG module forms a residual connection with the stacked features. Finally, the processed features are output after convolution.

[0074] The CDG module performs three-way convolution processing on the input features in parallel: the first path performs ordinary convolution, the second path performs dilated convolution, and the third path performs grouped convolution. The results of the three convolutions are stacked by channels, and after convolution, normalization and activation function processing, they are further processed by convolution and pooling operations, and then connected with the intermediate processing results to form residual connections.

[0075] Specifically, an ultraviolet corona discharge image of an insulator with dimensions of 640×640×3 is input into the YOLO-SM detection network for feature extraction. The feature extraction network uses a hierarchical and progressive approach to extract deep feature representations of the image step by step.

[0076] The overall structure of the feature extraction network consists of 5 convolutional modules and 12 SCCDG lightweight feature extraction modules. The convolutional module, as the basic feature extraction unit, is composed of a two-dimensional convolutional layer (Conv2d), a batch normalization layer (BatchNorm), and a SiLU activation function in sequence. The kernel size is set to 3×3, and the stride is 2, which is used to achieve feature downsampling and channel expansion.

[0077] The input image is first processed by two convolutional modules, reducing the image size from 640×640 to 160×160 and increasing the number of channels from 3 to 128. It then undergoes the following processing: the first group contains two consecutive lightweight feature extraction modules, with an output size of 80×80×256; after one ordinary convolution operation, it enters the second group of two consecutive lightweight feature extraction modules, with an output size of 40×40×512; after another ordinary convolution operation, it enters the third group of four consecutive lightweight feature extraction modules, with an output size of 20×20×512; after one more ordinary convolution operation, it is processed by the fourth group of two consecutive lightweight feature extraction modules, finally outputting a size of 20×20×1024.

[0078] The entire feature extraction network comprises 10 lightweight SCCDG feature extraction modules. Among them, the output features of the first, second, and third groups (with sizes of 80×80×256, 40×40×512, and 20×20×512, respectively) are directly connected to the feature fusion network as features at three different levels to provide multi-level feature information.

[0079] The SCCDG lightweight feature extraction module is the core innovation of this invention, specifically designed for small target detection tasks involving corona discharge, as detailed below. Figure 3 As shown. This module first separates the input features into n sub-features through an average channel separation operation, with each sub-feature having C / n channels, resulting in... Each of the n sub-features is subjected to a 1×1 convolution operation with a stride of 2. The convolutional features are then stacked by channel to form a unified feature representation.

[0080] The stacked features are then input into the CDG module for deep feature extraction. The CDG module is as follows: Figure 4 As shown, the CDG module adopts a multi-branch parallel processing architecture, simultaneously performing three different convolutional processes on the input features: the first path uses a convolutional operation with a kernel size of 1×1 and a stride of 1 to obtain the features. The second approach uses dilated convolution with a kernel size of 3×3 and an inflation rate of 2 to obtain features. By introducing holes to expand the receptive field, a wider range of contextual information can be captured without increasing computational cost; the third path uses grouped convolution operations with a kernel size of 3×3 and 2 groups to obtain features. By using grouped convolutions, computational overhead and the number of parameters are significantly reduced.

[0081] The three convolution results Features are obtained by performing channel stacking operations. And then Features are obtained by sequentially performing 2D convolution with a kernel size of 1×1 and a stride of 1, batch normalization, and Gaussian error activation function operations. Then the features Features are obtained by sequentially performing convolution operations with a kernel size of 1×1 and a stride of 1, average pooling operations with a kernel size of 3×3 and padding of 1, and fully connected layer operations. Finally, features and Output features are obtained by constructing residual connections. .

[0082] The output of the SCCDG lightweight feature extraction module integrates features through a final convolution operation with a 1×1 kernel and a stride of 1, ultimately outputting a multi-scale feature map of size H×W×C. This module, through a combination of channel separation, multi-branch convolution, and residual connections, ensures effective feature extraction for small targets while maintaining lightweight computational characteristics. This allows the network to operate efficiently on resource-constrained embedded devices, significantly improving the detection accuracy for small targets such as corona discharge.

[0083] S300: Input the multi-scale features into the feature fusion network, perform pooling operations at different scales through the dual-branch structure of the MSDAM multi-scale dual attention mechanism module to generate attention weights, and perform weighted fusion of the multi-scale features to obtain fused features;

[0084] Furthermore, the MSDAM multi-scale dual attention mechanism module includes: after preprocessing the input features, two parallel branches are input respectively. The first branch is processed by a first activation function after performing average pooling and max pooling operations at the first scale. The second branch is processed by a second activation function after performing average pooling and max pooling operations at the second scale. The two branches form residual connections with the input features to generate attention weights, and finally, weighted fusion is performed.

[0085] Furthermore, the first branch and the second branch use pooling kernels of different sizes for pooling operations. The first branch uses the Mish activation function after pooling, while the second branch uses the Swish activation function after pooling, resulting in different attention weight features between the two branches.

[0086] Specifically, the multi-scale features output from the feature extraction network are input into the feature fusion network for feature integration and optimization. The feature fusion network achieves effective fusion of multi-level features through upsampling, downsampling, and feature concatenation operations, and integrates the MSDAM multi-scale dual attention mechanism module to enhance the feature representation of the target region.

[0087] The overall architecture of the feature fusion network is as follows Figure 2 As shown, the network comprises two upsampling modules, four channel stacking modules, two convolutional modules, four lightweight feature extraction modules, and three multi-scale dual attention mechanism modules. The network adjusts and fuses image features in both spatial scale and channel dimension through upsampling and downsampling operations combined with channel stacking. The three MSDAM multi-scale dual attention mechanism modules connect to three output detection heads at different scales, providing attention-weighted high-quality features for subsequent target prediction.

[0088] The MSDAM multi-scale dual attention mechanism module is a core component of the feature fusion network, such as... Figure 5 As shown, this module is specifically designed to address the problem of indistinct corona discharge target features and susceptibility to background noise interference in complex backgrounds. It employs a dual-branch parallel processing architecture, adaptively enhancing the network's focus on important feature regions through multi-scale information extraction and attention weighting mechanisms.

[0089] The specific implementation process of the MSDAM multi-scale dual attention mechanism module is as follows: The input features first undergo a convolution operation with a kernel size of 1×1 and a stride of 1 to obtain the features. This is used to unify feature dimensions and prepare for subsequent processing. Then, the features... The process proceeds to the two parallel branches, one above the other, for processing.

[0090] In the first branch, features The algorithm sequentially performs average pooling with a kernel size of 3×3 and padding of 1, and max pooling with a kernel size of 3×3 and padding of 1. Average pooling preserves the overall information of local regions, while max pooling highlights salient features. The combination of these two pooling methods effectively extracts different types of spatial information. The results of these two pooling operations are merged through a channel stacking module, and then input into a pointwise convolution module with a kernel size of 1×1 and a stride of 1 for feature integration. After processing by the Mish activation function, the final feature is obtained. The Mish activation function possesses smooth non-linear properties, providing better gradient flow and feature representation capabilities. With the original input Construct residual connections to generate features with attention weights. .

[0091] In the second branch, the processing method is similar to that of the first branch, but different scale parameters are used. Features Average pooling and max pooling operations with a kernel size of 5×5 and padding of 2 are performed respectively. Compared with the 3×3 pooling kernel in the first branch, the 5×5 pooling kernel can capture a wider range of spatial context information. Similarly, the results of these two pooling operations are stacked by channels and fed into a pointwise convolutional module with a kernel size of 3×3 and padding of 1, and then processed by the Swish activation function to obtain features. The self-gating property of the Swish activation function allows for dynamic adjustment of information flow, enhancing the model's expressive power. Features With input Construct another residual connection to output features with attention weights. .

[0092] The two branches, by employing pooling kernels of different sizes (3×3 and 5×5) and different activation functions (Mish and Swish), are able to extract feature information from multiple scales and different nonlinear transformation perspectives, generating complementary attention weight features. The first branch focuses on extracting local detail features, while the second branch focuses more on capturing global contextual information.

[0093] Finally, the features generated from the two branches with attention weights are combined. and A weighted fusion operation is performed. Through the learned attention weights, the module adaptively adjusts the importance of different spatial locations and channels in the feature map, highlighting the feature representation of the corona discharge target region while suppressing background noise interference. The fused features are then subjected to a final feature integration through a 1×1 convolution operation with a stride of 1, resulting in a fused feature with multi-scale dual attention weighting.

[0094] Through the processing of the MSDAM multi-scale dual attention mechanism module, the feature fusion network can effectively reduce the interference of complex background noise on target detection, enhance the network's ability to identify corona discharge targets, and significantly improve detection accuracy and robustness.

[0095] S400: Input the fused features into the prediction network for target detection, and output the bounding box coordinates, category label, and confidence score of the corona discharge target.

[0096] Furthermore, the prediction network includes three detection heads, each containing a bounding box regression branch, a category classification branch, and a confidence prediction branch. The three detection heads respectively handle large-scale, medium-scale, and small-scale target detection tasks.

[0097] Furthermore, the output results of the three detection heads are merged and then post-processed, including setting a confidence threshold to filter low-confidence detection boxes, using a non-maximum suppression algorithm to remove overlapping redundant detection boxes through an IoU threshold, and outputting the final corona discharge target detection result.

[0098] Specifically, the attention-weighted fused features output from the feature fusion network are input into the prediction network for final target detection. The prediction network is responsible for converting high-level semantic features into specific detection results, outputting the precise location, category information, and confidence score of the corona discharge target.

[0099] The prediction network employs a multi-scale detection head architecture, comprising three independent detection heads corresponding to large-scale, medium-scale, and small-scale target detection tasks, respectively. This multi-scale design effectively addresses the issue of significant size variations in corona discharge targets in real-world scenarios, ensuring accurate detection of targets of varying sizes. The three detection heads receive outputs from different levels of the MSDAM multi-scale dual attention mechanism modules within the feature fusion network, forming a detection system ranging from coarse-grained to fine-grained.

[0100] Each detection head has an identical internal structure, containing three parallel prediction branches: a bounding box regression branch, a category classification branch, and a confidence prediction branch. The bounding box regression branch predicts the spatial location information of the target, outputting four parameters including the coordinates of the bounding box center point (x, y), the width w, and the height h of the bounding box. It achieves precise localization of corona discharge targets through regression learning. The category classification branch determines which category the detected target belongs to; in the application scenario of this invention, it mainly identifies corona discharge phenomena and outputs the probability distribution of each category. The confidence prediction branch evaluates the reliability of the detection results, outputting a confidence score between 0 and 1, reflecting the probability that a target actually exists within the detection box.

[0101] The output dimensions of the three detection heads are adapted to different feature map sizes. The large-scale detection head processes smaller feature maps and is mainly responsible for detecting large targets in the image; the medium-scale detection head processes medium-sized feature maps and focuses on medium-sized targets; the small-scale detection head processes larger feature maps, has higher spatial resolution, and is specifically used to detect corona discharge phenomena in small targets. This hierarchical detection strategy significantly improves the network's ability to detect targets at different scales.

[0102] The raw outputs from the three detectors require post-processing to obtain the final detection results. The post-processing process includes two main stages: confidence threshold filtering and non-maximum suppression.

[0103] In the confidence threshold filtering stage, the system sets a confidence threshold of 0.5 and filters the detection boxes output by all three detection heads according to their confidence scores. Only detection boxes with confidence scores higher than the set threshold are retained for subsequent processing, while low-confidence detection boxes are discarded directly. This step effectively removes a large number of false positive results and improves detection accuracy.

[0104] In the non-maximum suppression (NMS) processing stage, an algorithm based on the IoU (Intersection over Union) threshold is used to remove overlapping redundant detection boxes. The specific process is as follows: First, all retained detection boxes are sorted from highest to lowest confidence score; then, the detection box with the highest confidence score is selected as the baseline box; the IoU value between this baseline box and all other detection boxes is calculated; when the IoU value exceeds a set threshold of 0.5, it is considered that the two detection boxes detect the same target, the box with the higher confidence score is retained, and the redundant box with the lower confidence score is deleted; this process is repeated until all detection boxes have been processed.

[0105] The final output after post-processing includes the bounding box coordinates of each corona discharge target. The category label "discharge" and its corresponding confidence score are included. The bounding box coordinate parameters have the following meanings:

[0106] : The x-coordinate (pixel position) of the top-left corner of the bounding box; : The vertical coordinate (pixel position) of the top left corner of the bounding box; : The horizontal coordinate (pixel position) of the bottom right corner of the bounding box; : The vertical coordinate (pixel position) of the bottom right corner of the bounding box.

[0107] These four coordinate values ​​determine the precise location and extent of the corona discharge target in the original image, which can be directly used to draw a detection box on the ultraviolet image, realizing the visual annotation of the corona discharge phenomenon and providing accurate and reliable technical support for the safety monitoring and fault diagnosis of transmission lines. The design of the entire prediction network ensures high accuracy and high reliability of the detection results, meeting the needs of practical engineering applications.

[0108] Example 2:

[0109] This is a case study of a power company's practical application of ultraviolet corona discharge detection for insulators on a 220kV transmission line. The transmission line is approximately 50 kilometers long with 120 towers, each equipped with 12-15 strings of composite insulators. Due to the line's location in a coastal area with severe salt spray pollution, corona discharge frequently occurs on the insulator surface, requiring regular inspection and monitoring.

[0110] Traditional detection methods have the following problems: manual inspection is inefficient, requiring 15 days for each full-line inspection; nighttime ultraviolet imaging detection accuracy is not high, with a false detection rate of up to 25%; existing target detection algorithms have difficulty identifying small-sized corona discharge targets, with a false alarm rate as high as 30%.

[0111] The company employs the YOLO-SM-based insulator ultraviolet corona discharge target detection method of this invention, deploying a detection system on an ultraviolet imaging device mounted on a UAV. The system configuration is as follows: equipped with a Jetson AGXXavier edge computing device, 32GB of memory, and 512GB of SSD storage; ultraviolet camera resolution of 1024×768, frame rate of 30fps; detection distance of 50–200 meters.

[0112] In a one-month practical application, the method of this invention was used to continuously monitor the transmission line, acquiring a total of 5000 ultraviolet images, including 800 images of corona discharge targets. Comparative tests with traditional YOLOv5, YOLOv8, and manual detection methods yielded the following experimental results:

[0113] Table 1. Test Comparison Table

[0114]

[0115] As shown in Table 1, the comparative analysis reveals that the method of this invention significantly outperforms existing methods in all aspects: the detection accuracy of this method reaches 92.4%, an improvement of 7.1 percentage points compared to YOLOv8 and 13.9 percentage points compared to manual detection. This is mainly due to the fact that the lightweight feature extraction module of SCCDG, through multi-branch convolution processing, can better extract the subtle features of corona discharge in small targets.

[0116] Significantly improved recall: The recall rate reached 89.7%, and the false negative rate dropped to 10.3%, a reduction of 17.4 percentage points compared to traditional methods. The MSDAM multi-scale dual attention mechanism module effectively enhances the ability to identify weak corona signals in complex backgrounds through pooling at different scales and dual activation functions.

[0117] The false alarm rate was significantly reduced: the false alarm rate was only 7.6%, a decrease of 7.1 percentage points compared to YOLOv8. The shape deviation loss and corner distance loss terms in the composite bounding box regression loss function enable the model to more accurately locate irregularly shaped corona discharge regions and reduce misjudgments due to background noise.

[0118] Improved computational efficiency: Detection speed reaches 32.5ms / frame, a 16% improvement compared to YOLOv8, with a model size of only 18.9MB, facilitating deployment on edge devices. The channel separation and grouped convolution design in the SCCDG lightweight feature extraction module effectively reduces computational overhead.

[0119] Practical Application Results: In one month of actual operation, the system successfully detected 43 corona discharge fault points, of which 35 were confirmed as actual faults on-site, achieving an accuracy rate of 81.4%. Compared to the previous manual inspections that could only find 20-30 fault points per month, the detection efficiency has been significantly improved. The system can also achieve 24 / 7 continuous monitoring, promptly detecting corona discharge phenomena at night and under adverse weather conditions, providing strong protection for the safe and stable operation of transmission lines.

[0120] The effectiveness of the core technology components of this invention was verified in a power laboratory by comparing and testing 600 ultraviolet corona discharge images of insulators at a 500kV substation that were actually collected on-site.

[0121] Compare the performance differences between using the standard convolutional module and the SCCDG lightweight feature extraction module:

[0122] Table 2. Validation Table of SCCDG Lightweight Feature Extraction Module

[0123]

[0124] As shown in Table 2, the test results show that the SCCDG lightweight feature extraction module improves detection accuracy while significantly reducing model complexity through channel separation and multi-branch convolution design.

[0125] Comparison of detection performance with and without MSDAM multi-scale dual attention mechanism:

[0126] Table 3. Validation of the attention mechanism of the MSDAM multi-scale dual attention mechanism module

[0127]

[0128] As shown in Table 3, the MSDAM multi-scale dual attention mechanism module effectively enhances the ability to identify corona discharge of small targets in complex backgrounds through the combination of bi-branch multi-scale pooling and different activation functions.

[0129] Comparing the effects of traditional intersection-union loss and composite bounding box regression loss function:

[0130] Table 4. Validation Table of Composite Loss Function

[0131]

[0132] As shown in Table 4, the composite loss function significantly improves the detection and positioning accuracy of irregularly shaped corona discharge targets by fusing shape deviation loss, corner distance loss and density-weighted loss.

[0133] In actual testing at a 500kV substation, the complete YOLO-SM detection system demonstrated a significant performance advantage compared to the version that removed core components one by one:

[0134] Overall detection accuracy: improved from 84.7% in the basic version to 92.4% in the full version.

[0135] Computational efficiency: Model size reduced by 25.3%, inference speed increased by 21.1%.

[0136] Practicality: The false detection rate is reduced to 7.6% in complex power environments, meeting the needs of practical engineering applications.

[0137] This test confirms the independent contribution and synergistic optimization effect of each core technical component of the present invention, providing a reliable technical solution for the detection of ultraviolet corona discharge in power system insulators.

[0138] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting target of ultraviolet corona discharge of insulator based on YOLO-SM, characterized in that, The method comprises the steps of: S100: obtaining a trained YOLO-SM detection network, which comprises a feature extraction network, a feature fusion network and a prediction network, wherein: The feature extraction network is integrated with an SCCDG lightweight feature extraction module, which separates multiple sub-features through an average channel separation operation and performs parallel ordinary convolution, dilated convolution and grouped convolution processing on the input features through a CDG module; The feature fusion network is integrated with an MSDAM multi-scale dual attention mechanism module, which includes two parallel branches that use different size pooling kernels and are processed by Mish and Swish activation functions respectively to generate complementary attention weights; The YOLO-SM detection network is trained using a composite bounding box regression loss function; S200: inputting the insulator ultraviolet corona discharge image into the YOLO-SM detection network, and performing channel splitting and multi-branch convolution processing on the input features through the SCCDG lightweight feature extraction module integrated in the feature extraction network to extract multi-scale features; S300: inputting the multi-scale features into the feature fusion network, performing different scale pooling operations through the double-branch structure of the MSDAM multi-scale dual attention mechanism module to generate attention weights, and weighting and fusing the multi-scale features to obtain fused features; S400: inputting the fused features into the prediction network for target detection, and outputting the bounding box coordinates, class label and confidence of the corona discharge target; The composite bounding box regression loss function includes the weighted sum of the following four losses: The intersection over union loss is calculated by the loss value of the overlap of the prediction box and the target box; The shape deviation loss is calculated by the aspect ratio difference between the prediction box and the target box; The corner distance loss is calculated by the distance between the corner coordinates of the prediction box and the target box; The density weighted loss sets a weighting factor based on the target distribution density.

2. The YOLO-SM-based insulator ultraviolet corona discharge target detection method according to claim 1, characterized in that, The feature extraction network includes a convolution module and an SCCDG lightweight feature extraction module, and the convolution module is composed of two-dimensional convolution, batch normalization and activation function, and the outputs of multiple SCCDG lightweight feature extraction modules are connected to the feature fusion network to provide multi-level features.

3. The YOLO-SM-based insulator ultraviolet corona discharge target detection method according to claim 1, characterized in that, The SCCDG lightweight feature extraction module comprises: The input features are separated into multiple sub-features through an average channel separation operation, the sub-features are stacked after convolution operation, and the stacked features are input into the CDG module for processing, the CDG module outputs are connected to the stacked features to form a residual connection, and finally the processed features are output through convolution operation; The CDG module performs three-way convolution processing on the input features in parallel: the first way performs ordinary convolution operation, the second way performs dilated convolution operation, and the third way performs grouped convolution operation; the three convolution results are stacked in the channel, and then processed through convolution, normalization and activation function, and finally connected to the intermediate processing results through convolution and pooling operations.

4. The YOLO-SM-based insulator ultraviolet corona discharge target detection method according to claim 1, characterized in that, The feature fusion network comprises an up-sampling module, a channel superposition module, a convolution module, an SCCDG lightweight feature extraction module and an MSDAM multi-scale dual attention mechanism module, the feature scale is adjusted through up-sampling and down-sampling operations, feature splicing is realized through channel superposition, and the multiple MSDAM multi-scale dual attention mechanism modules are respectively connected with detection heads of different scales.

5. The YOLO-SM-based insulator ultraviolet corona discharge target detection method according to claim 1, characterized in that, The MSDAM multi-scale dual attention mechanism module comprises: input features are respectively input into two parallel branches after preprocessing, the first branch is processed through a first activation function after average pooling and maximum pooling operations of a first scale, the second branch is processed through a second activation function after average pooling and maximum pooling operations of a second scale, the two branches respectively form residual connections with input features to generate attention weights, and finally weighted fusion is performed.

6. The YOLO-SM-based insulator ultraviolet corona discharge target detection method according to claim 5, characterized in that, The first branch and the second branch adopt different sizes of pooling kernels for pooling operations, the first branch adopts a Mish activation function after pooling, the second branch adopts a Swish activation function after pooling, and the two branches generate different attention weight features.

7. The YOLO-SM-based insulator ultraviolet corona discharge target detection method according to claim 1, characterized in that, The prediction network comprises three detection heads, each detection head comprises a bounding box regression branch, a class classification branch and a confidence prediction branch, and the three detection heads respectively process large-scale, medium-scale and small-scale target detection tasks.

8. The YOLO-SM-based insulator ultraviolet corona discharge target detection method according to claim 7, characterized in that, The output results of the three detection heads are combined and post-processed, including setting a confidence threshold to filter low-confidence detection boxes, removing redundant detection boxes through an IoU threshold by using a non-maximum suppression algorithm, and outputting final corona discharge target detection results.

Images