Power transmission line insulator defect detection method based on deep learning

Abstract

The invention discloses a power transmission line insulator defect detection method based on deep learning, and the method comprises the steps: S1, obtaining an aerial image, carrying out the preprocessing of the aerial image, and generating an insulator defect detection data set; s2, based on a YOLOv8n model, introducing a space object attention module (RCS-OSA), and improving the feature extraction capability under a complex background; s3, adding an extrusion excitation network (SENet) in the neck network of the model to enhance the feature expression ability between channels; s4, constructing a small target detection layer, and adding a high-resolution feature map (160 * 160) to capture small defect target information; s5, optimizing a detection head by using multi-scale feature fusion and a dynamic label distribution strategy, and improving the recognition precision of the model on insulator defects under a complex background; and S6, outputting defect positions, types and confidence coefficients based on a detection result, and generating an electric power inspection report. According to the method, the accuracy and robustness of insulator defect detection under the complex background of the power transmission line are effectively improved, and the method has relatively high applicability and real-time performance.

Description

Technical Field

[0001] The present invention relates to the field of target detection technology, and in particular to a transmission line insulator defect detection method based on deep learning, which is used to improve the accuracy and efficiency of transmission line insulator defect detection and is suitable for scenarios such as power system inspection and transmission line equipment maintenance. Background Art

[0002] As a key component in the transmission line, insulators are responsible for mechanical support and electrical insulation. Their performance directly affects the stable operation of the transmission line and even the entire power system. However, after long-term exposure to the natural environment, insulators are easily affected by factors such as wind and sand, high temperature, ice and snow, and external force damage, and may have defects such as self-explosion, defect or flashover. These defects will lead to a decrease in insulation performance, which in turn threatens the safe operation of the power grid.

[0003] Traditional insulator detection methods mainly include manual inspection and detection technology based on image processing. Manual inspection methods are time-consuming and labor-intensive, and it is difficult to fully cover transmission lines in complex terrain. Traditional methods based on image processing rely on edge detection, histogram equalization and other technologies, which are difficult to adapt to defect target extraction under complex backgrounds. In recent years, with the popularization of drone inspections and the rapid development of deep learning, target detection technology has gradually been applied to the field of insulator defect detection. However, existing deep learning-based detection methods still have the problems of high false detection rate and high missed detection rate when dealing with small target defects and complex backgrounds, which makes it difficult to meet the needs of actual power inspections.

[0004] In view of the above problems, designing an insulator detection algorithm that can efficiently and accurately detect small target defects in complex backgrounds is the focus and difficulty of current research. This paper improves the existing YOLO framework by introducing an attention mechanism and a small target detection layer, significantly improving the comprehensive performance of insulator defect detection. Summary of the invention

[0005] The problem to be solved by the present invention is that the existing transmission line insulator defect detection method has problems such as high false detection rate and high missed detection rate in complex background and small target detection, which is difficult to meet the needs of power inspection.

[0006] In order to solve the above technical problems, the present invention proposes a transmission line insulator defect detection method based on deep learning, which specifically includes the following steps:

[0007] S1, obtain aerial images and perform preprocessing to generate an insulator detection data set containing three types of defects: self-explosion, defect, and flashover;

[0008] S2. Based on the YOLOv8n model, the spatial object attention module (RCS-OSA) is introduced into the backbone network to replace the traditional two-dimensional convolution module to improve the feature extraction capability under complex backgrounds.

[0009] S3. Add a squeeze excitation network (SENet) to the neck network of the model to enhance the inter-channel feature expression capability and improve the focus on the target defect area;

[0010] S4, build a small target detection layer, add a feature map with a resolution of 160×160, and improve the detection performance of small target defects through multi-scale feature fusion;

[0011] S5. Adopt anchor-free detection and dynamic label allocation strategy to optimize the detection head structure and balance the requirements of detection accuracy and model lightweight;

[0012] S6. Based on the detection results, the detection information including the defect location, category and confidence level is output, and the report required for the power inspection is generated.

[0013] Preferably, the S1 comprises the following steps:

[0014] Step 1: Obtain aerial insulator images and construct an image dataset including normal insulators and defective insulators (such as defect, flashover, self-explosion, etc.);

[0015] Step 2: Annotate defective objects in the image, including but not limited to insulator defect, insulator flashover and insulator self-explosion;

[0016] Step 3: Divide the labeled data set into training set, test set and validation set according to the preset ratio.

[0017] Preferably, the RCS-OSA module introduced in S2 includes the following optimizations:

[0018] Step 1: In the training stage, multi-branch convolution is used to extract feature maps of different scales to enhance the model's adaptability to complex backgrounds;

[0019] Step 2: In the inference phase, the branch parameters are reorganized onto the main branch to reduce computational complexity and optimize inference speed;

[0020] Step 3: Improve the model's ability to allocate weights to important defect areas by shuffling and reconstructing the channels of the feature map.

[0021] Preferably, the squeeze excitation network (SENet) added in S3 comprises the following steps:

[0022] Step 1: Perform global average pooling on the feature map to compress it into a feature vector containing global information;

[0023] Step 2: Learn the importance weight of each feature channel through a two-layer fully connected network;

[0024] Step 3: Adjust the weight ratio of feature channels to enhance key features while suppressing non-important information to increase the attention of defect areas.

[0025] Preferably, the construction of the small target detection layer in S4 includes the following steps:

[0026] Step 1: Based on the original three-layer detection structure (P3, P4, P5) of YOLOv8n, a new detection layer (P2) with a resolution of 160×160 is added to extract more detailed feature information;

[0027] Step 2: Multi-scale feature fusion is achieved through the bidirectional flow of deep and shallow features, effectively improving the detection accuracy of small target defects (such as small-area self-explosion or flashover);

[0028] Step 3: Optimize the parameter configuration of the detection layer to ensure that the added detection layer has minimal impact on the overall efficiency of the model.

[0029] Preferably, the optimization of anchor-free box detection and dynamic label allocation in S5 includes:

[0030] Step 1: Adopt anchor-free detection technology to avoid complex hyperparameter settings related to anchor boxes and reduce the amount of calculation;

[0031] Step 2: Introduce a dynamic label assignment strategy to dynamically adjust the positive sample assignment based on feature matching to improve the flexibility of label assignment;

[0032] Step 3: Combine CIoU and DFL Loss to optimize the loss function and balance detection accuracy and training stability in classification and regression tasks.

[0033] Preferably, the output and visualization of the detection results in S6 comprises the following steps:

[0034] Step 1: Screen the detection targets according to the confidence output by the detection head and eliminate low-confidence targets;

[0035] Step 2: Output the location information, category label and confidence value of the detection result;

[0036] Step 3: Combine the detection results with the original image for visualization and generate an intuitive detection report for use in power inspection decision-making.

[0037] The beneficial effects of the present invention are:

[0038] By introducing the spatial object attention module and the squeeze excitation network, the feature extraction and expression capabilities of the model in complex backgrounds are effectively improved; by adding a small target detection layer, the detection performance of small-area defect targets is significantly improved; by adopting the anchor-free frame detection and dynamic label allocation strategy, the model can improve the detection accuracy while ensuring high operating efficiency; the present invention is suitable for multi-scenario transmission line insulator defect detection, and has the advantages of high detection accuracy, strong real-time performance and wide adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a step description of the present invention;

[0040] Figure 2 It is the RCS-OSA module description;

[0041] Figure 3 It is the network structure diagram of the squeeze incentive module;

[0042] Figure 4 It is a comparison of the detection layer architecture before and after the improvement of the feature fusion stage. DETAILED DESCRIPTION

[0043] Example 1

[0044] like Figure 1 As shown, a method for detecting defects in insulators of a transmission line based on deep learning comprises the following steps:

[0045] S1. Obtain aerial insulator images and perform data preprocessing

[0046] In this embodiment, the aerial image data preprocessing includes the following steps:

[0047] Data collection: The dataset consists of two parts. The first part comes from the open-source Chinese Power Line Insulator Dataset (CPLID) on the Internet. CPLID includes 600 normal insulator images and 248 self-explosion defect insulator images taken by drones. Some images of self-explosion defects of insulators are randomly selected as one of the sources of the dataset in this paper. The second part is images downloaded from the public network, mainly including three types of defects: insulator defect, insulator flashover, and insulator self-explosion.

[0048] Data labeling: The labelme tool is used to label the defect targets (such as self-explosion, defect, and flashover) in the image; the labeling labels are insulator self-explosion, insulator flashover, and insulator defect. After the labeling is completed, the data set is divided into training set, test set, and validation set in a ratio of 7:1.5:1.5.

[0049] S2. Introducing the spatial object attention module (RCS-OSA)

[0050] In this embodiment, since insulator defects are easily affected by factors such as complex background, various types, and small target scale, which may lead to missed detection and false detection, the RCS-OSA module is used to replace the traditional C2f convolution module in the YOLOv8n backbone network to improve the feature extraction capability. Figure 2 As shown, the specific steps include:

[0051] Training phase: After the input is channel-separated, a portion is processed by RepVGG (including no processing, 1×1 convolution and 3×3 convolution, and using SiLU as the only activation function); then, the input processed by the RepVGG block and the unchanged input are concatenated and channel-shuffled;

[0052] Inference stage: In the inference stage, after the input is channel-separated, a part of it undergoes a simpler RepConv processing (including no processing and 3×3 convolution processing, and using SiLU as the only activation function).

[0053] S3. Adding Squeeze Excitation Network (SENet) to the Neck Network

[0054] In this embodiment, the SENet module is introduced to enhance the relationship between channels in the feature map, such as Figure 3 As shown, it increases the focus on defective areas;

[0055] Global feature compression: In the SENet module, each channel of the feature map is compressed into a global feature vector through the global average pooling operation (GAP); specifically, the global feature vector of each channel is first calculated by averaging the features of all spatial positions of each channel. The formula is as follows:

[0056]

[0057] Among them, z c is the global eigenvector of the cth channel, u c (m,n) is the eigenvalue of the mth row and nth column in the cth channel. H and W represent the height and width of the image, respectively, indicating the spatial dimension of the feature map. In this way, we compress the spatial information of each channel into a global information vector, which is convenient for subsequent feature processing and channel weighting operations.

[0058] Feature weight calculation: The weight of each channel is adaptively adjusted through a two-layer fully connected network; the calculation formula is as follows:

[0059] S=F(Z,W)=σ(g(Z,W))=σ(Wδ(WZ))

[0060] Among them, δ is the ReLU activation function, σ is the Sigmoid activation function, and Z is the global feature vector;

[0061] Feature recalibration: The calculated weight factor is multiplied by the original feature map channel by channel to enhance important features and suppress irrelevant background. The formula is as follows:

[0062]

[0063] Among them, s c is the scaling factor of the cth channel, u c is the original feature map, is the adjusted feature map.

[0064] S4. Construct small target detection layer

[0065] In order to improve the detection accuracy of small target defects, a small target detection layer with a resolution of 160×160 is added to the YOLOv8n model. The filter size of small targets larger than 4×4 pixels is detected with a larger field of view, which solves the problem of insufficient capture of small target defect information features such as self-explosion, defect, and flashover to a certain extent. Figure 4 In the feature fusion stage, the flow of deep semantic information to shallow semantic information realizes the feature fusion of 20×20, 40×40, 80×80, and 160×160, and the flow of shallow semantic information to deep semantic information realizes the feature fusion of 160×160, 80×80, 40×40, and 20×20; through multi-scale feature fusion, the detection capability of small target defects is further improved.

[0066] S5. Optimize the anchor-free frame and dynamic label allocation strategy of the detection head

[0067] In terms of label assignment and loss function, the anchor-free idea adopted by YOLOv8n avoids the complex calculations and hyperparameter settings related to anchor boxes, significantly improving the processing speed of the model, while the dynamic label matching strategy increases the flexibility of positive sample box selection;

[0068] In addition, using DFL Loss+CIoU Loss as regression loss and BCE Loss as classification loss optimizes model training and improves the detection accuracy of defect targets in complex backgrounds.

[0069] S6. Output defect report based on test results

[0070] Based on the detection results, a report containing the category, location and confidence of the defect target is generated to support power inspection decisions; the report content includes the location, classification and confidence value of each defect target, and is finally output through visualization for operators to judge.

[0071] The above are only some embodiments of the present invention. For ordinary technicians in this field, several modifications and improvements can be made without departing from the creative concept of the present invention. Any modifications, equivalent substitutions and improvements made within the creative concept of the present invention are within the scope of protection of the present invention.

Claims

1. A method for detecting defects in insulators of power transmission lines based on deep learning, characterized in that: The following steps are involved: S1: Acquire aerial images and preprocess them to generate insulator defect detection dataset; S2: Using the YOLOv8n model as the basic framework, the spatial object attention module (RCS-OSA) is introduced in the backbone network and the neck network to replace the traditional two-dimensional convolution module and enhance the feature extraction capability; S3: Add a squeeze excitation network (SENet) to the neck network of the model to improve the ability to capture feature relationships between channels; S4: Build a small target detection layer and add high-resolution feature maps to capture the details of small target defects; S5: Optimize the detection head using multi-scale feature fusion strategy to improve the detection accuracy of defect targets under complex backgrounds; S6: Generate classification and location information of insulator defects based on the detection results.

2. A method for detecting defects in transmission line insulators based on deep learning according to claim 1, characterized in that: The specific steps of S1 include: Step 1: Obtain aerial insulator images. The image dataset includes normal insulators and defective insulators (defects, flashovers, self-explosion, etc.); Step 2: Label the defect targets in the image, where the labeling content includes different types of defect targets; Step 3: Divide the labeled dataset into training set, test set, and validation set to support model training and performance evaluation.

3. A method for detecting defects in transmission line insulators based on deep learning according to claim 1, characterized in that: The RCS-OSA module introduced in S2 specifically includes: Step 1: In the training phase, multi-branch convolutional layers are used to extract multi-scale features; Step 2: In the inference phase, the multi-branch parameters are reorganized onto the main branch to reduce computational cost and optimize feature extraction.

4. A method for detecting defects in transmission line insulators based on deep learning according to claim 1, characterized in that: The specific steps of the squeeze excitation network (SENet) added in S3 include: Step 1: Compress the feature map into a global feature vector through a global average pooling operation; Step 2: Use a two-layer fully connected network to learn feature weights and adjust the weight distribution of each channel; Step 3: Use the adjusted feature weights to update the input feature map.

5. The method for detecting defects in transmission line insulators based on deep learning according to claim 1, characterized in that: The construction of the small target detection layer in S4 includes: Step 1: Based on the original YOLOv8n detection layer architecture, add a feature map with a resolution of 160×160; Step 2: Joint detection of multi-scale targets is achieved through the bidirectional fusion of deep and shallow feature information.

6. A method for detecting defects in transmission line insulators based on deep learning according to claim 1, characterized in that: The optimization of the detection head in S5 includes: Step 1: Adopt a dynamic label assignment strategy to improve the flexibility of positive sample matching; Step 2: Use anchor-free detection technology to reduce anchor-related calculations and hyperparameter settings.

7. A method for detecting defects in transmission line insulators based on deep learning according to claim 1, characterized in that: The method supports the detection of various defect targets, including but not limited to self-explosion, defect and flashover defects.

8. The method for detecting defects in transmission line insulators based on deep learning according to claim 1, characterized in that: The detection results of the method can output comprehensive information including the location, category and confidence level of the target defect, which can be used for subsequent power inspection decision support.

Images