Power equipment anomaly detection algorithm for power inspection

Abstract

The invention relates to a power equipment anomaly detection algorithm for power inspection, which is characterized by comprising the following steps of: automatically shooting and establishing a power equipment data set; designing a CD module and a GS module, constructing a network structure of an abnormal equipment detection algorithm for power inspection based on a YOLO v8 algorithm, and naming the network structure as a CG-YOLO network structure; training the abnormal equipment detection algorithm for power inspection by using a power equipment data set, setting a training round, a learning rate and an optimizer, and obtaining an optimal weight file; an optimal weight file is loaded for an abnormal equipment detection algorithm for power inspection, a corresponding target detection result is output, and the target detection result comprises the position of the target area in the to-be-detected image and the category corresponding to each target area; according to the method, the abnormal defects of the power equipment can be accurately positioned and identified, and the optimal balance of precision, efficiency and speed is realized on a self-built power equipment data set and a public data set.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) power inspection technology, and in particular to an electric power equipment anomaly detection algorithm for power inspection. Background Art

[0002] my country's outdoor power facilities are often located in inaccessible areas such as forests, hills, and mountains. These facilities are constantly exposed to harsh weather conditions (such as rain, snow, lightning, and hail), as well as the impact of high voltage and high current, and natural aging. This can easily lead to problems such as corrosion, short circuits, and spontaneous combustion, which can affect the stable operation of the power grid. Analysis shows that improper transmission system maintenance and missed equipment inspections are the main causes of these accidents, making regular inspections a necessary maintenance measure for power systems.

[0003] Drone inspections have alleviated some of the pain points of inspection tasks, such as high risk, difficulty, and labor-intensiveness. However, efficiently and accurately processing the massive amounts of inspection images captured by drones remains a significant challenge. Therefore, equipping computers with human-like cognitive abilities, enabling them to fully, comprehensively, and rapidly understand the semantic information contained in images, is a key research topic in China. Deep learning-based image recognition technology offers a potential solution for analyzing inspection images of power grid facilities.

[0004] While existing methods have achieved some improvement in the detection performance of power targets and anomalies, they are limited in the number of defect types they detect, typically focusing on insulator defects with little attention to other defects. Consequently, existing algorithms are only reliable for transmission network inspections. However, distribution network power equipment, characterized by a wide variety of defects, small defect targets, and complex background environments, places higher demands on the algorithms. When targets are located in complex backgrounds, these algorithms struggle to effectively extract features from a variety of anomalies, exhibiting poor environmental adaptability, leading to missed and false detections of anomalies. This requires ensuring the accuracy of the algorithms. Furthermore, existing algorithm models are complex, necessitating effective parameter compression methods to address the deployment challenges of drone edge devices. Furthermore, existing equipment has high requirements for the operating environment, resulting in poor nighttime defect detection performance. Cloud computing relies on data transmission, while local edge computing platforms are limited in accuracy and detection speed, making them inadequate for wide-area distribution network inspections. Summary of the Invention

[0005] To solve the above problems, the study of a lightweight anomaly detection algorithm based on multi-granularity feature fusion has high practical significance and has important realization value for building smart grids, improving the automation level of power grids, and reducing the inspection pressure of staff. The present invention designs an abnormal equipment detection algorithm for power inspection for online inspection, achieving the optimal balance between accuracy and speed with the highest detection accuracy and competitive detection speed, and has significant comparative advantages compared with other advanced algorithms; at the same time, starting from the multi-resolution image detection task, the present invention aims at the target offset problem caused by image scaling in actual scenes, and constructs a multi-resolution, multi-category power equipment data set, providing the algorithm with target offset data at multiple resolutions, and providing a reliable quantitative basis for evaluating the robustness and generalization of the abnormal equipment detection algorithm. The specific plan is as follows: An electric power equipment anomaly detection algorithm for electric power inspection, characterized by comprising: Step S1: Self-photograph and establish a power equipment dataset for qualitative and quantitative evaluation of the detection effect of the abnormal equipment detection algorithm; Step S2: Design the CD module and the GS module, and build a network structure for the abnormal equipment detection algorithm for power inspection based on the YOLO v8 algorithm, and name it the CG-YOLO network structure; Step S3: using the power equipment data set to train the abnormal equipment detection algorithm for power inspection, setting the training rounds, learning rate and optimizer, and obtaining the optimal weight file; Step S4: Load the optimal weight file into the abnormal device detection algorithm for power inspection, and output the corresponding target detection result, which includes the position of the target area in the image to be detected and the category corresponding to each target area.

[0006] Furthermore, in step S1, the self-photographing and establishing of the power equipment data set includes: Using drones to obtain images of power equipment in actual operation, including at least two different types of insulators; The acquired power equipment images are finely annotated to obtain a power equipment dataset with reference value.

[0007] Furthermore, in step S2, the CG-YOLO network structure includes a feature extraction backbone network, a feature fusion neck network and a target detection head; the feature extraction backbone network is used to process the input power equipment image, output three feature maps of large, medium and small scales, and input them into the feature fusion neck network; the target detection head performs detection based on the fusion features of large, medium and small scales output by the feature fusion neck network, and outputs the final target detection result.

[0008] Furthermore, the feature extraction backbone network includes a CD module and a convolution module; wherein the CD module completes the targeted feature extraction without changing the size of the feature map and the number of overall channels; the convolution module is used to adjust the size of the feature map; the feature extraction backbone network first passes through a convolution module, then repeats the above process four times according to the rule of convolution first and CD module later, and finally passes through the SPPF module and outputs; Among them, the second group of convolution-CD modules outputs the P3 feature map for detecting small targets, the third group of convolution-CD modules outputs the P4 feature map for detecting medium-sized targets, and the output of the fourth group of convolution-CD modules is output through the SPPF module as the P5 feature map for detecting large targets.

[0009] Furthermore, the feature fusion neck network includes an upsampling module, a splicing module, a convolution module and a GS module; wherein the upsampling module and the convolution module are used to adjust the size of the feature map, and respectively include the first and second upsampling layers and the first and second convolution layers; the splicing module is used to splice feature maps of the same size together, and includes the first to fourth splicing layers; the GS module performs targeted extraction and fusion on the spliced ​​feature maps to obtain the feature map that is finally input to the detection head, and includes the first to fourth splicing modules and the first to fourth GS modules; First, the P5 feature map is upsampled and spliced ​​with the P4 feature map, and then processed by the GS module to obtain the P6 feature map; secondly, the P6 feature map is upsampled and spliced ​​with the P3 feature map, and then processed by the GS module to generate the P7 feature map; the P7 feature map is downsampled and spliced ​​with the P6 feature map, and then processed by the GS module to obtain the P8 feature map; finally, the P8 feature map is downsampled and spliced ​​with the P5 feature map, and then processed by the GS module to obtain the P9 feature map.

[0010] Furthermore, the target detection head includes a first detection head, a second detection head and a third detection head; the first detection head is used to receive the fusion features output by the feature fusion neck network for detecting large targets, the second detection head is used to receive the fusion features output by the feature fusion neck network for detecting medium-sized targets, and the third detection head is used to receive the fusion features output by the feature fusion neck network for detecting small targets.

[0011] Furthermore, step S3 includes: constructing a network structure of an abnormal equipment detection algorithm for power inspection based on step S2, using the power equipment data set obtained in step S1 for training, setting the optimizer to SGD, adjusting the initial learning rate, momentum, and non-maximum suppression threshold, and retaining all model weights throughout the training process, and obtaining the optimal weight file through comprehensive evaluation of the three evaluation indicators of mAP, mAP50:95, and loss function.

[0012] Furthermore, in step S4, the optimal weight file is loaded into the abnormal equipment detection algorithm for power inspection, and the corresponding target detection result is output, including: firstly, the optimal weight file obtained in step S3 is loaded into the abnormal equipment detection algorithm for power inspection for initialization, and then the images and videos to be detected are input into the abnormal equipment detection algorithm after initialization. The abnormal equipment detection algorithm completes the positioning and classification of power equipment. The target categories of normal insulators include pillar-type ceramic insulators, composite insulators, glass insulators, and suspended ceramic insulators. The defects of insulators include flashover damage and fracture of a single shed. The beneficial effects of the present invention are as follows: 1. The power equipment anomaly detection algorithm of the present invention can accurately locate and identify abnormal defects of power equipment, and achieves the optimal balance of accuracy, efficiency and speed on self-built power equipment datasets and public datasets; at the same time, the present invention improves the algorithm's detection effect on small targets, solves the balance problem of existing algorithms in terms of accuracy and computational complexity, and is more suitable for deployment in scenarios with limited computing power, such as edge computing platforms.

[0013] 2. The power equipment anomaly detection algorithm of the present invention combines multi-granularity information and lightweight modules, which improves the overall detection accuracy of the abnormal equipment detection algorithm in cases where complex backgrounds and small targets account for a large proportion.

[0014] 3. The present invention proposes a CD module, which effectively extracts key feature information in the backbone part and provides sufficient feature information for the deep network, significantly improving the detection accuracy of the abnormal device detection algorithm and significantly reducing the computational complexity of the abnormal device detection algorithm.

[0015] 4. The present invention designs a GS module, which reduces the number of channels in the deep layer of the network, significantly strengthens the feature information of the deep network, and gives the abnormal device detection algorithm a certain degree of flexibility, so that the abnormal device detection algorithm can determine the importance of each channel by learning the weights of different channels, ensure the accuracy of header information, and improve the final detection effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a schematic diagram of the network structure of the abnormal equipment detection algorithm for power inspection in the present invention.

[0017] Figure 2 Schematic diagram of the principle and calculation method of the convolution module of the present invention.

[0018] Figure 3 Schematic diagram of the principle and calculation method of the pooling module of the present invention.

[0019] Figure 4 Schematic diagram of the CD module structure of the present invention.

[0020] Figure 5 Schematic diagram of the GS module structure of the present invention.

[0021] Figure 6 、 Figure 7 This is a schematic diagram of the detection results of a normal insulator according to an embodiment of the present invention.

[0022] Figure 8 、 Figure 9 Schematic diagram of the detection results of defective insulators according to an embodiment of the present invention.

[0023] Figure 10 、 Figure 11 Schematic diagram of the detection effect of an embodiment of the present invention on a public dataset. DETAILED DESCRIPTION

[0024] An electric power equipment anomaly detection algorithm for electric power inspection, specifically including: Step S1: Self-photograph and establish a power equipment dataset for qualitative and quantitative evaluation of the detection effect of the abnormal equipment detection algorithm: Drones were used to capture images of power equipment in operation. The primary targets were four different types of insulators: post-type ceramic insulators, composite insulators, glass insulators, and suspension-type ceramic insulators. The dataset also included images of flashover damage and shed fracture defects, which were used to qualitatively analyze the effectiveness of the abnormal equipment detection algorithm. The insulator dataset contains 1,887 images, 80% of which are used as a training set and 20% as a validation set.

[0025] Step S2: Design the CD module and the GS module, and build a network structure for the abnormal device detection algorithm for power inspection based on the YOLO v8 algorithm, and name it the CG-YOLO network structure: like Figure 1 As shown in Figure 2, the network structure of the abnormal equipment detection algorithm for power inspection consists of a feature extraction backbone network, a feature fusion neck network, and a target detection head. The feature extraction backbone network adopts a collaborative architecture of CD modules and convolution modules. The CD module achieves selective feature extraction while maintaining the feature map size and the number of input and output channels unchanged. The convolution module achieves a step-by-step dimensionality reduction of the feature map size through downsampling with a stride of 2. After passing through a convolution module, the input image is iteratively processed four times using a "convolution module + CD module" combination unit, and finally outputs a 20×20 feature map through the SPPF module. It is worth noting that the second to fourth combination units output 80×80 (P3), 40×40 (P4), and 20×20 (P5) feature maps, respectively, constructing a multi-scale feature representation system. The feature fusion neck network realizes cross-layer feature fusion through the cascade operation of upsampling-splicing-GS modules: first, the P5 feature map is upsampled to 40×40 and spliced ​​with the P4 feature map, and processed by the GS module to obtain the P6 feature map; secondly, the P6 feature map is upsampled to 80×80 and spliced ​​with the feature map P3, and the P7 feature map is generated by the GS module; the feature map P7 is downsampled to 40×40 and spliced ​​with the feature map P6, and then processed by the GS module to obtain the P8 feature map; finally, the P8 feature map is downsampled and spliced ​​with the feature map P5, and processed by the GS module to obtain the 20×20 feature map P9; this architecture uses multiple base layers to innovatively construct multi-scale fusion features including 80×80 (P7), 40×40 (P8), and 20×20 (P9), realizing feature complementarity from shallow details to deep semantics. 1. Convolutional Layer

[0026] As a core component of a convolutional neural network (CNN), the convolutional layer plays a key role in extracting features from the input image. By performing a convolution operation on the image—sliding the convolution kernel pixel by pixel across the image—it captures rich local visual features, such as the edges and textures of power equipment. This operation enables each convolution kernel to focus on extracting specific features of power equipment, thereby constructing a complex, high-dimensional feature representation. This provides a rich information foundation for subsequent image analysis tasks such as classification and detection.

[0027] like Figure 2 As shown in Figure 1, the convolution operation is essentially a mathematical element-wise product-sum operation, where each movement of the convolution kernel generates a single pixel value in the output feature map through weighted multiplication and summation. This process is repeated until the entire input image is covered. In the CNN architecture, the convolution layer automatically learns data features, abstracting from the most basic visual features to complex feature combinations, demonstrating its powerful ability in visual tasks such as image recognition.

[0028] Furthermore, the convolutional layer design embodies two key features: parameter sharing and sparse connectivity. Parameter sharing significantly reduces the number of model parameters, improving the model's generalization across different images. Sparse connectivity, on the other hand, ensures that the network can focus on processing local features, effectively reducing the amount of computation. These two features work together to improve network efficiency and performance. The steps for implementing a convolutional layer are briefly as follows: (1) Initialization: Define the size, number, and step size of the convolution kernel. These parameters directly affect the refinement of feature extraction and the dimension of the feature map.

[0029] (2) Convolution operation: The convolution kernel starts from the upper left corner of the image and slides across the image according to the set step size. At each position, the element-wise product is performed and the sum is calculated to obtain the corresponding value on the feature map.

[0030] (3) Activation function: Apply an activation function, such as ReLU, to the result of the convolution operation to introduce nonlinearity and enhance the learning ability of the network.

[0031] In practical applications, the design and configuration of convolutional layers significantly impact network performance. For example, parameters such as kernel size, convolution step size, and padding must be carefully selected based on the specific task and the characteristics of the input data. Therefore, in the field of drone-based power inspections, careful consideration of convolution parameters is crucial to maximize the overall computational efficiency of the abnormal device detection algorithm and reduce redundant computing power, thereby improving its performance in real-world scenarios. 2. Pooling, upsampling, and splicing

[0032] like Figure 3 As shown in Figure 1, the pooling layer is a key component of convolutional neural networks, used to reduce the dimensionality of feature maps while retaining key feature information. Pooling can reduce network computational complexity and enhance model generalization capabilities. The pooling layer has three main functions: (1) By reducing the size of the feature map of power equipment, the pooling layer can significantly reduce the parameters and computational complexity of subsequent layers, thereby improving the overall speed of anomaly detection. At the same time, pooling helps to avoid overfitting.

[0033] (2) Max pooling effectively retains the texture information of power equipment by selecting the maximum value in the region as the representative feature.

[0034] (3) The pooling layer makes the network more robust to small translations of the input image, and improves the model's tolerance to deformation and rotation of power equipment from the perspective of the drone.

[0035] Pooling operations are mainly divided into two types: maximum pooling and average pooling, such as Figure 3 As shown in Figure 2. Max pooling effectively preserves image texture information by selecting the maximum value in a region as the representative feature. Average pooling calculates the average of all values ​​in a region, retains the average feature, and smoothes the feature map. Implementing a pooling layer involves the following steps: (1) Select the pooling type: Decide whether to use maximum pooling or average pooling. Maximum pooling selects the maximum value of the covered area, while average pooling calculates the average of all values ​​in the area.

[0036] (2) Set the pooling window size and step size: Define the window size and sliding step size of the pooling operation. The window size determines the area covered by each pooling operation, and the step size affects the size of the feature map after pooling.

[0037] (3) Perform pooling operation: Slidingly apply the pooling window to the feature map, and perform maximum selection or average calculation on each coverage area according to the pooling type to generate a new, smaller feature map.

[0038] It's important to note that when the convolution stride is 2 or greater, the convolution module can also implement pooling. The only difference is that the operation does not use the average or maximum value method. Instead, it uses a matrix to multiply the input. By adjusting the matrix weights, important feature information can be obtained in a targeted manner. In contrast to the pooling layer, the upsampling module (Upsample Layer) converts small feature maps back into large feature maps. It is used to amplify the size of feature maps in deep networks, allowing them to fuse features of different depths together through the concatenation layer (Concat Layer). This increases feature reuse, reduces feature loss in deep networks, and effectively reduces the depth of abnormal equipment detection algorithms, thereby improving the detection effect of power equipment. 3.CD module

[0039] The C2f module is an improved CSP neck structure proposed in YOLOv8. Through dual convolution design and hidden channels, the C2f module effectively enhances the network's processing capabilities for images of different sizes, while achieving fine-tuning of feature maps without adding additional computational burden. The module first doubles the number of feature map channels and divides them into two paths, processes one of the paths through the depth of the neck layer, and then fuses it with the other path, and finally adjusts it to the target number of channels. This design not only optimizes the parameter configuration, but also ensures the efficient integration of information flow through residual connections, significantly improving the performance and accuracy of the model in multi-scale target detection. However, the overall structural design of C2f is relatively simple, and it can only extract some feature information of power equipment. With extensive and repeated use, some power equipment features are lost as the network depth increases, reducing the overall sensitivity of the abnormal equipment detection algorithm to abnormal features and affecting the overall detection effect.

[0040] like Figure 4As shown, the input is first roughly processed by the convolution module, and the output channel is divided into two parts. One part undergoes depthwise separable convolution and ordinary convolution, and the other part undergoes GS convolution. The two channels are then spliced ​​together, and the result is then spliced ​​with the output of the first convolution. After another convolution, the result is output. The CD module provides a wide range of shortcut branches within the module, combining power equipment feature maps from different layers. This enables the abnormal equipment detection algorithm to comprehensively determine the content of abnormal features. Based on multi-granular feature information, the detection capability of small abnormal features is enhanced. The SP module is used to split the input channel into two, reducing the computational complexity of subsequent modules and ensuring overall detection speed. Finally, by learning the weight of each channel at the output, the abnormal equipment detection algorithm can weaken unnecessary background features and unimportant local features and strengthen abnormal features.

[0041] Since this part of the module is mainly used in the backbone part and the overall computational workload needs to be guaranteed, a feature method based on depthwise separable convolution is taken into consideration. At the same time, considering the speed balancing problem of the branches on both sides, this embodiment adds a convolution module on the depthwise separable convolution side, which not only improves the overall feature extraction effect of the abnormal device detection algorithm, but also ensures the overall parallel efficiency of the abnormal device detection algorithm. 4. GS module

[0042] As the network depth increases, the number of channels in the abnormal device detection algorithm will increase significantly. When the network depth increases, the number of channels at the exit of the CD module will increase significantly, reducing the efficiency of the abnormal device detection algorithm in detecting abnormal power equipment. This will seriously affect the computational complexity and detection speed of the abnormal device detection algorithm. Although it can improve the detection accuracy, it will seriously affect the overall speed. In order to adapt to the channel environment of the deep network, this embodiment designs a feature fusion method based on attention, such as Figure 5 shown.

[0043] Similar to the CD module, the GS module first uses convolution to roughly process the input, then divides the channel into two equal parts, each undergoing two GS convolutions and a bottle process to ensure the overall detection accuracy of the abnormal equipment detection algorithm. However, the feature information is fused by addition rather than splicing to obtain the power equipment features from different fields of view. The first convolution and the fused features are then spliced ​​together, and after the last convolution, the result is output. While reducing the number of output channels, the above method can effectively weaken background information and strengthen the feature information of the target itself, allowing the abnormal equipment detection algorithm to focus on small target features. Moreover, when no abnormal features are detected or some noise is present, the shortcut branch on the other side can weaken or eliminate the feature information on this side through splicing and convolution, ensuring that there is no significant loss of the overall power equipment features, improving the lower limit of the detection effect of the abnormal equipment detection algorithm, and enabling the abnormal equipment detection algorithm to be used in data sets of different environments. 5. SPPF module

[0044] The SPPF module of this embodiment includes three maximum pooling layers connected in series, and the pooling kernel sizes of the three maximum pooling layers are all 5×5.

[0045] The SPPF module processes the input 20×20 feature map through three maximum pooling layers in sequence, and concatenates the results of the three maximum pooling layers with the input feature map to obtain the output of the pooled feature map. 6. Detection head

[0046] The target detection head includes a first detection head, a second detection head and a third detection head. The first detection head is used to receive the fused features of size 20×20 output by the feature fusion neck network, the second detection head is used to receive the fused features of size 40×40 output by the feature fusion neck network, and the third detection head is used to receive the fused features of size 80×80 output by the feature fusion neck network.

[0047] Among them, the 20×20 feature map is used to detect large targets, the 40×40 feature map is used to detect medium-sized targets, and the 80×80 feature map is used to detect small targets. Finally, the abnormal device detection algorithm comprehensively considers the output results of different detection heads to obtain the final target location and target category information, and outputs it as the final result.

[0048] Step S3: Use the power equipment dataset to train the abnormal equipment detection algorithm for power inspection, set the training rounds, learning rate and optimizer, and obtain the optimal weight file:

[0049] To more scientifically compare anomaly detection algorithms for power inspections, this example determined the training parameters and operating environment for all abnormal device detection algorithms to achieve relatively fair detection results. The specific parameters are shown in Table 1. The abnormal device detection algorithms used PyTorch to build the network structure and used the same 16-batch-size to eliminate the interference of parameter settings on detection accuracy. The platform used a Windows system, an i5-12400F CPU, and an NVIDIA RTX 3070Ti GPU to eliminate variations in detection speed caused by device issues. Furthermore, to ensure consistent constraints across the abnormal device detection algorithms, the same learning rate and optimizer were set for all algorithms. Finally, based on the mAP and loss function curves, the optimal weights for the abnormal device detection algorithms were determined after comprehensive consideration.

[0050] Step S4: Load the optimal weight file for the abnormal device detection algorithm for power inspection and output the corresponding target detection results. The target detection results include the location of the target area in the image to be detected and the category corresponding to each target area: To further optimize the detection results, the power equipment anomaly detection algorithm sets constraints on the output results. The confidence threshold is set to 0.25 to prevent the abnormal equipment detection algorithm from making false detections. The IOU threshold is set to 0.7 to prevent the abnormal equipment detection algorithm from outputting multiple detection frames for a single target, thereby avoiding counting anomalies. Finally, the detection results of pillar-type ceramic insulators, composite insulators, glass insulators, and suspended ceramic insulators are obtained. Abnormal device detection algorithm effect

[0051] In addition to testing on a self-built dataset of power equipment, this embodiment was also compared with existing advanced algorithms on a public dataset to demonstrate the effectiveness and practicality of the improved abnormal equipment detection algorithm. The public dataset, AU-AIR, consists of 32,832 frames of images captured from drone video, simulating the actual operating environment of a drone. The targets include eight categories: people, cars, trucks, vans, motorcycles, bicycles, buses, and trailers. The dataset is divided into training, test, and validation sets in a ratio of 6:1:3, as required.

[0052] This example conducted comparative experiments on different types of object detection algorithms. These included convolutional neural network-based object detection algorithms such as YOLOV5n, YOLOV8n, the YOLOV10 series, and YOLOV12n; and Transformer-based object detection algorithms such as RT-DETR-L and RT-DETR-Resnet50. A comprehensive analysis comparing detection accuracy, detection speed, and parameter count demonstrated the superiority of the CG-YOLO algorithm proposed in this example.

[0053] 1. Quantitative Analysis

[0054] This embodiment completed a comparative experiment on a self-built dataset, and the experimental results are shown in Table 2. AP1-AP4 represent four different types of insulators, namely, post-type ceramic insulators, composite insulators, glass insulators, and suspended ceramic insulators. In the insulator dataset, the small model of this embodiment significantly reduced the number of parameters while being 1.1% higher than YOLOv8n and having a similar speed. Compared with other algorithms, the model of this embodiment has significant advantages in speed and number of parameters. While maintaining the highest detection accuracy, it has a clear comparative advantage in computational complexity and detection speed, demonstrating that CG-YOLO achieves a better balance between accuracy and detection speed.

[0055]

[0056] On the public AU-AIR dataset, the mAP of the abnormal device detection algorithm of this embodiment is 0.6% and 0.1% higher than that of YOLO8n, and its parameter count is 9.8% lower, yet its overall computational speed is similar and it maintains a high detection speed, as shown in Table 3. Compared to other larger models, the abnormal device detection algorithm of this embodiment is 1.7% and 0.9% higher than RT-DETR, respectively, with a significant advantage in parameter count and approximately double the speed, demonstrating the superiority of the abnormal device detection algorithm of this embodiment. Compared to the YOLOv10 series, the abnormal device detection algorithm of this embodiment is 1.2% higher than m and 0.5% higher than x, with a clear advantage in both parameter count and speed. These experimental results demonstrate that the abnormal device detection algorithm of this embodiment has good generalization and practicality, and also demonstrates good detection results in other scenarios. 2. Qualitative analysis

[0057] The detection results of the CG-YOLO algorithm proposed in this embodiment on the insulator dataset are as follows: Figure 6 、 Figure 7 As shown in the figure, this embodiment can accurately identify the characteristic information of pillar-type ceramic insulators, composite insulators, glass insulators, and suspension-type ceramic insulators during UAV power inspections, and perform real-time identification and positioning operations on the targets during the flight of the UAV. In terms of insulator defect detection, the visualization results of insulator defect detection in this embodiment are as follows: Figure 8 、 Figure 9 As shown, this embodiment can accurately identify and locate the power target at the scale of a single insulator shed, accurately identify defective insulators, and further distinguish the types of defects. Figure 8The effect of this embodiment on detecting the breakage of a single shed is demonstrated. Figure 9 The results of this embodiment in detecting flashover damage on a single shed are shown. Figure 10 、 Figure 11 As shown, in open-world detection, this embodiment can accurately identify target information such as cars and pedestrians and complete positioning, verifying the effectiveness of the improved abnormal device detection algorithm.

[0058] Therefore, this embodiment demonstrates significant advantages in detecting abnormalities in power equipment, particularly in identifying insulator defects. It not only accurately identifies critical defects (such as flashover damage and fractures) at the scale of a single insulator shed, surpassing existing algorithms in detection accuracy, but also effectively integrates multi-scale information through a feature extraction structure and focuses on key channel features, significantly improving the reliability of detecting multi-scale and extremely small targets (such as shed defects) in complex environments. Its lightweight design and efficient computing architecture significantly reduce the computational burden and increase processing speed. These features make this embodiment particularly suitable for drone-based online power inspection scenarios with high real-time requirements, ensuring the accuracy and reliability of defect detection.

[0059] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. An electric power equipment anomaly detection algorithm for electric power inspection, characterized in that: include: Step S1: Self-photograph and establish a power equipment dataset for qualitative and quantitative evaluation of the detection effect of the abnormal equipment detection algorithm; Step S2: Design the CD module and the GS module, and build a network structure for the abnormal equipment detection algorithm for power inspection based on the YOLO v8 algorithm, and name it the CG-YOLO network structure; Step S3: using the power equipment data set to train the abnormal equipment detection algorithm for power inspection, setting the training rounds, learning rate and optimizer, and obtaining the optimal weight file; Step S4: Load the optimal weight file into the abnormal device detection algorithm for power inspection, and output the corresponding target detection result, which includes the position of the target area in the image to be detected and the category corresponding to each target area.

2. The power equipment anomaly detection algorithm for power inspection according to claim 1, characterized in that: In step S1, the self-photographing and establishing of a power equipment data set includes: Using drones to obtain images of power equipment in actual operation, including at least two different types of insulators; The acquired power equipment images are finely annotated to obtain a power equipment dataset with reference value.

3. The power equipment anomaly detection algorithm for power inspection according to claim 1, characterized in that: In step S2, the CG-YOLO network structure includes a feature extraction backbone network, a feature fusion neck network and a target detection head; the feature extraction backbone network is used to process the input power equipment image, output three feature maps of large, medium and small scales, and input them into the feature fusion neck network; the target detection head performs detection based on the fusion features of large, medium and small scales output by the feature fusion neck network, and outputs the final target detection result.

4. The power equipment anomaly detection algorithm for power inspection according to claim 3, characterized in that: The feature extraction backbone network includes a CD module and a convolution module; the CD module completes targeted feature extraction without changing the size of the feature map and the number of overall channels; the convolution module is used to adjust the size of the feature map; the feature extraction backbone network first passes through a convolution module, then repeats the above process four times according to the rule of convolution first and CD module later, and finally passes through the SPPF module and outputs; Among them, the second group of convolution-CD modules outputs the P3 feature map for detecting small targets, the third group of convolution-CD modules outputs the P4 feature map for detecting medium-sized targets, and the output of the fourth group of convolution-CD modules is obtained through the SPPF module to obtain the P5 feature map for detecting large targets.

5. The power equipment anomaly detection algorithm for power inspection according to claim 4, characterized in that: The feature fusion neck network includes an upsampling module, a splicing module, a convolution module and a GS module; wherein the upsampling module and the convolution module are used to adjust the size of the feature map, and respectively include the first and second upsampling layers and the first and second convolution layers; the splicing module is used to splice feature maps of the same size together, and includes the first to fourth splicing layers; the GS module performs targeted extraction and fusion on the spliced ​​feature maps to obtain the feature map that is finally input to the detection head, and includes the first to fourth splicing modules and the first to fourth GS modules; First, the P5 feature map is upsampled and spliced ​​with the P4 feature map, and processed by the GS module to obtain the P6 feature map; secondly, the P6 feature map is upsampled and spliced ​​with the P3 feature map, and the P7 feature map is generated by the GS module; the P7 feature map is downsampled and spliced ​​with the P6 feature map, and then the P8 feature map is obtained by the GS module; finally, the P8 feature map is downsampled and spliced ​​with the P5 feature map, and processed by the GS module to obtain the P9 feature map.

6. The power equipment anomaly detection algorithm for power inspection according to claim 3, characterized in that: The target detection head includes a first detection head, a second detection head and a third detection head; the first detection head is used to receive the fusion features output by the feature fusion neck network for detecting large targets, the second detection head is used to receive the fusion features output by the feature fusion neck network for detecting medium-sized targets, and the third detection head is used to receive the fusion features output by the feature fusion neck network for detecting small targets.

7. The power equipment anomaly detection algorithm for power inspection according to claim 1, characterized in that: Step S3 includes: constructing a network structure of an abnormal equipment detection algorithm for power inspection based on step S2, using the power equipment data set obtained in step S1 for training, setting the optimizer to SGD, adjusting the initial learning rate, momentum, and non-maximum suppression threshold, and retaining all model weights throughout the training process, and obtaining the optimal weight file through comprehensive evaluation of three evaluation indicators: mAP, mAP50:95, and loss function.

8. The power equipment anomaly detection algorithm for power inspection according to claim 7, characterized in that: In step S4, the optimal weight file is loaded into the abnormal equipment detection algorithm for power inspection, and the corresponding target detection result is output, including: first, the optimal weight file obtained in step S3 is loaded into the abnormal equipment detection algorithm for power inspection for initialization, and then the images and videos to be detected are input into the abnormal equipment detection algorithm after initialization. The abnormal equipment detection algorithm completes the positioning and classification of power equipment. The normal insulator target categories include post-type ceramic insulators, composite insulators, glass insulators, and suspended ceramic insulators. The defects of the insulators include flashover damage and fracture of a single shed.

Images