Insulator defect detection method and device, medium and equipment

Abstract

The invention discloses an insulator defect detection method and device, a medium and equipment, and relates to the technical field of power system equipment detection. The method comprises the following steps: integrating a star network StarNet into a YOLO11n network, adding a small target feature pyramid network into a neck network of the YOLO11n network, applying an ADown module into a backbone network and the neck network of the YOLO11n network, and replacing an IOU evaluation index of a loss function of the YOLO11n network with an NWD evaluation index to obtain an improved YOLO11n network; taking the training data set marked with the insulator state category as input, taking the insulator state category as output, training the improved YOLO11n network, and obtaining an insulator defect detection model; and inputting insulator image data acquired in real time into the insulator defect detection model to obtain a defect detection result of the insulator. According to the scheme, the defect detection accuracy of the insulator can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power system equipment detection, and particularly relates to a method, device, medium and equipment for insulator defect detection. Background Art

[0002] With the continuous expansion of the scale of the power system, transmission lines are widely distributed in various complex geographical environments. As a key component of the transmission line, the operating state of the insulator is directly related to the safety and stability of the transmission line. Insulators that are long-term exposed to the natural environment are subjected to multiple tests of mechanical stress, electrical load and harsh climate conditions, and are extremely prone to defects such as self-explosion, breakage, and flashover. If these defects cannot be detected and processed in time, it may cause transmission line failures and even large-scale power outages, resulting in huge economic losses.

[0003] Traditional insulator defect detection methods mainly rely on manual inspection or robot detection based on simple image processing technology. The manual inspection method not only has a large labor intensity and low efficiency, but also is affected by the subjective factors of the inspectors, and it is easy to miss and misdetect. Although robot detection has improved the detection efficiency to a certain extent, its detection accuracy still needs to be improved in the face of complex environments and diverse defect types. In recent years, deep learning technology has made remarkable progress in the field of computer vision, providing a new solution for insulator defect detection. Object detection algorithms based on deep learning, such as Faster RCNN and YOLO series algorithms, have been widely used in insulator detection.

[0004] However, the existing insulator defect detection methods based on deep learning still face many challenges. On the one hand, the complex natural environment and changing weather conditions (such as light changes, cloud and fog occlusion, rain and snow weather, etc.) seriously affect the accuracy of the detection model, making it difficult for the model to stably identify insulator defects in practical applications. On the other hand, there are various types of insulator defects, especially small target defects such as flashover and breakage, and their features are difficult to effectively extract, resulting in low detection accuracy for insulator defects. In addition, some detection algorithm models are complex, have a large number of parameters, and require high computing resources, making it difficult to meet the needs of real-time detection tasks. Especially in the resource-constrained UAV inspection scenario, the real-time performance and adaptability of the detection model become key issues.

[0005] Therefore, the existing insulator defect detection methods cannot accurately identify the categories of insulator defects. Summary of the Invention

[0006] Accurate and timely insulator defect detection can effectively prevent transmission line failures, reduce power outages caused by insulator defects, reduce economic losses, improve the reliability and stability of the power system operation, and play an important role in promoting the safe and efficient operation and maintenance of the power system.

[0007] Based on this, it is necessary to provide an insulator defect detection method, device, medium and equipment for the technical problem that the existing insulator defect detection methods cannot accurately identify the insulator defect categories.

[0008] The present invention adopts the following technical solutions:

[0009] In the first aspect, the present invention provides an insulator defect detection method, and the method includes:

[0010] Integrate the StarNet into the YOLO11n network, add a small object feature pyramid network to the neck network of the YOLO11n network, apply the ADown module to the backbone network and the neck network of the YOLO11n network, and replace the IOU evaluation index of the loss function of the YOLO11n network with the NWD evaluation index to obtain an improved YOLO11n network;

[0011] Use the training data set labeled with insulator status categories as the input and the insulator status categories as the output to train the improved YOLO11n network to obtain an insulator defect detection model;

[0012] Input the real-time collected insulator image data into the insulator defect detection model to obtain the defect detection result of the insulator.

[0013] Further, the step of using the training data set labeled with insulator status categories as the input and the insulator status categories as the output to train the improved YOLO11n network to obtain an insulator defect detection model specifically includes:

[0014] Collect an image data set containing various insulator statuses;

[0015] Use the imgaug library to process the image data set and simulate the weather in the image data set to obtain an image data set with enhanced features;

[0016] Divide the image data set with enhanced features into a training data set and a validation data set, and label the insulator status categories for the training data set and the validation data set to obtain a training data set labeled with insulator status categories and a validation data set labeled with insulator status categories;

[0017] Use the training data set labeled with insulator status categories as the input and the insulator status categories as the output to train the improved YOLO11n network, and adjust the parameters of the improved YOLO11n network using the validation data set labeled with insulator status categories during the training process to obtain the insulator defect detection model.

[0018] Further, after obtaining the insulator defect detection model, the method further includes:

[0019] Inputting the test data set labeled with insulator status categories into the insulator defect detection model to obtain test results;

[0020] Calculating multiple performance indicators of the insulator defect detection model according to the test results, where the multiple performance indicators include precision, recall rate, average precision, number of parameters, and computational complexity;

[0021] In the case where any one of the multiple performance indicators is less than the preset threshold, expand the training data set and adjust the parameters of the improved YOLO11n network until the multiple performance indicators of the trained insulator defect detection model are all greater than the preset threshold.

[0022] Further, the star network is stacked by a four-order hierarchical structure, and the specific composition method includes:

[0023] Using a convolutional layer for downsampling, and using a modified display block demoblock for feature extraction, and replacing the GELU function in the display block with a RelU6 function.

[0024] Further, the small target feature pyramid network includes:

[0025] SPDConv layer and CSP-Omnikernel layer;

[0026] Among them, the SPDConv layer is used to downsample the feature map in the spatial dimension and expand the channel dimension;

[0027] The CSP-Omnikernel layer includes a global branch, a large branch, and a local branch, and is used to learn feature representations from global to local.

[0028] Further, the ADown module is used to reduce the spatial resolution of the feature map, and the NWD evaluation index is used to improve the accuracy of detecting small target defects.

[0029] In a second aspect, the present invention provides an insulator defect detection device, including:

[0030] An optimization module, which is used to integrate the StarNet into the YOLO11n network, add a small object feature pyramid network to the neck network of the YOLO11n network, apply the ADown module to the backbone network and the neck network of the YOLO11n network, and replace the IOU evaluation metric of the loss function of the YOLO11n network with the NWD evaluation metric to obtain an improved YOLO11n network;

[0031] A training module, which is used to take the training data set labeled with the insulator state category as the input and the insulator state category as the output to train the improved YOLO11n network to obtain an insulator defect detection model;

[0032] A detection module, which is used to input the insulator image data collected in real time into the insulator defect detection model to obtain the defect detection result of the insulator.

[0033] The present invention provides a computer-readable storage medium, and the storage medium stores a computer program, and when the computer program is executed by a processor, the insulator defect detection method is implemented.

[0034] The present invention provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, the insulator defect detection method is implemented.

[0035] The at least one technical solution adopted by the present invention can achieve the following beneficial effects: By integrating StarNet into the YOLO11n network, the present invention can improve the inference speed of the YOLO11n network through star operations and compress the YOLO11n network. Adding a small target feature pyramid network to the neck network of the YOLO11n network can enhance the detection ability of the YOLO11n network for small target defects in insulator images by fusing multiple special layers. Applying the ADown module to the backbone network and neck network of the YOLO11n network can reduce the resolution of the feature map, optimize the parameters of the YOLO11n network, and reduce the computational amount. By replacing the IOU evaluation index in the loss function of the YOLO11n network with the NWD evaluation index, an improved YOLO11n network is obtained, enabling the YOLO11n network to accurately identify the position deviation of tiny targets (tiny targets contain very few pixels in the image). Finally, using a training data set labeled with insulator status categories to train the improved YOLO11n network, an insulator defect detection model capable of accurately identifying the categories of small target defects in insulator images is obtained. In summary, through the above solutions, the insulator defect detection model can better adapt to different lighting and weather conditions and the requirements of small target defect detection, improving the robustness and stability of the model in practical applications and enhancing the accuracy of identifying insulator defects. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The drawings described herein are used to provide a further understanding of the present invention, form a part of the present invention, and the schematic embodiments and descriptions thereof are used to explain the present invention without unduly limiting the present invention. In the drawings:

[0037] Figure 1 is a schematic flowchart of a method for detecting insulator defects provided by the present invention;

[0038] Figure 2 is a structural diagram of the improved YOLO11n network provided by the present invention;

[0039] Figure 3 is a structural diagram of the small target feature pyramid network provided by the present invention;

[0040] Figure 4 is a structural diagram of the CSP-Omnikernel provided by the present invention;

[0041] Figure 5 is a structural diagram of the ADown provided by the present invention;

[0042] Figure 6 is a schematic diagram of a device for detecting insulator defects provided by the present invention;

[0043] Figure 7Schematic diagram of a computer device for implementing an insulator defect detection method provided by the present invention. Detailed implementation manners

[0044] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0045] At present, the server mentioned in the present invention can be a server set up on a service platform, or a device such as a desktop computer or a notebook computer that can execute the solution of the present invention. For the convenience of description, only the server will be used as the execution subject for description below. The technical solutions provided by each embodiment of the present invention will be described in detail below with reference to the drawings.

[0046] Refer to Figure 1 , which is an insulator defect detection method in the present invention, specifically including the following steps:

[0047] S10: Integrate the StarNet into the YOLO11n network, add a small target feature pyramid network to the neck network of the YOLO11n network, apply the ADown module to the backbone network and the neck network of the YOLO11n network, and replace the IOU evaluation index of the loss function of the YOLO11n network with the NWD evaluation index to obtain an improved YOLO11n network.

[0048] In this embodiment, refer to Figure 2, is an improved YOLO11n network structure, including a backbone network Backbone, a neck network Neck, and a head network Head. Among them, the backbone network includes an input layer Input, a convolutional layer ConvMoudle, an ADown module, a C3K2_StarNB layer, an ADown module, a C3K2_StarNB layer, an ADown module, a C3K2_StarNB layer, an ADown module, a C3K2_StarNB layer, an SPPF layer, and a C2PSA layer connected in sequence. The neck network includes an upsampling layer Upsample, a concatenation layer Concat, a C3K2_StarNB layer, an upsampling layer Upsample, a concatenation layer Concat, a CSPOmniKernel layer, a C3K2_StarNB layer, an ADown module, a concatenation layer Concat, a C3K2_StarNB layer, an ADown module, a concatenation layer Concat, and a C3K2_StarNB layer connected in sequence. The head network includes three detection layers 11Detect.

[0049] Among them, StarNet is a lightweight neural network structure that demonstrates excellent performance and low latency under a compact network structure and efficient budget. StarNet uses the star operation to map the input data to an extremely high-dimensional non-linear feature space without increasing the computational complexity. By stacking multiple layers of star operations, StarNet can significantly increase the implicit feature dimension at each stage, thereby achieving rich feature representation while keeping the model compact.

[0050] In this embodiment, StarNet improves the inference speed of the YOLO11n network and compresses the YOLO11n network through the star operation. The star operation recursively increases the implicit dimension of the YOLO11n network. Assuming that the width of the initial YOLO11n network is c, the expression for performing a star operation once is:

[0051]

[0052] Assuming that the width of an initial network is c, * represents element-wise multiplication, and O l is the output obtained from the l-th star operation, then we get:

[0053]

[0054]

[0055] Among them, W1 and W2 are weight matrices, W1, W2, x ∈ R (C+1)×1 , X is the input feature, X ∈ R(C+1)×n , where c is the number of input channels.

[0056] By stacking multiple layers of neural networks containing star operations, the implicit dimension of the YOLO11n network can be exponentially amplified.

[0057] In one or more embodiments of the present invention, the star network is stacked by a fourth-order hierarchical structure, and the specific composition method includes:

[0058] Use a convolutional layer for downsampling, and use a modified display block demoblock (a basic building block summarized from related work) for feature extraction, and replace the GELU function in the display block with a RelU6 function.

[0059] To solve the problem of low detection accuracy for small target defects such as flashover in insulator defect detection, a Small Object Pyramid Network (SOPN) is designed in the backbone network of the YOLO11n network.

[0060] Reference Figure 4 , different from the traditional method of improving the small target detection ability by adding a P2 feature detection layer, the present invention first makes the P2 feature detection layer obtain features rich in small target information through SPDConv and fuse them with the features obtained by the P3 feature detection layer, and then integrates them through CSP-Omnikernel obtained by fusing and improving based on CSPNet and Omnikernel, thereby effectively improving the small target detection performance.

[0061] Specifically, the small target feature pyramid network includes:

[0062] SPDConv layer and CSP-Omnikernel layer. The SPDConv layer is used to downsample the feature map in the spatial dimension and expand the channel dimension.

[0063] The CSP-Omnikernel layer includes a global branch, a large branch, and a local branch, and is used to learn feature representations from global to local.

[0064] The structure of SOPN mainly has two key points: SPDConv layer and CSP-Omnikernel layer. Different from traditional convolution operations, the SPDConv layer is used to downsample the feature map X in the spatial dimension and expand it in the channel dimension, so as to reduce the spatial size of the feature map and increase the number of channels while retaining all information.

[0065] Specifically, assume that a feature map of arbitrary size is X, and its size is S×S×C1. The SPD layer divides this feature map into multiple sub-feature maps, each of which is formed by sampling every scale pixels, that is, a sub-feature map composed of all X(i,j) elements that satisfy the sum and can be divided by scale, thereby achieving downsampling. As shown in formula (3). Here F represents the segmentation ratio. The formula for the SPDConv layer to convert the original feature map into the intermediate feature map is:

[0066]

[0067] Through these formula transformations, the SPDConv layer converts the original feature map X(S,S,C1) into an intermediate feature map In this way, the SPDConv layer effectively reduces the spatial resolution of the feature map while increasing the number of channels, providing richer information for subsequent feature extraction and object detection tasks.

[0068] refer to Figure 4 In (b), CSP-Omnikernel is a feature module based on the fusion of CSP and OmniKernel. First, the feature is input into 1×1 convolution (Conv1×1) for processing, and then the processed result is input into the spilt function for tensor segmentation, and the segmented part is input into the OmniKernel module for processing.

[0069] refer to Figure 4 In (a), the OmniKernel module consists of three branches, namely the global branch Global, the large branch Large and the local branch Local, to effectively learn the feature representation from global to local, and finally improve the detection performance of small objects. The result processed by the OmniKernel module is integrated with the other part of the feature tensor after segmentation, and then output after Conv1×1 convolution.

[0070] S20: Taking the training data set labeled with the insulator state category as input and the insulator state category as output, training the improved YOLO11n network to obtain the insulator defect detection model.

[0071] In this embodiment, labeling software is used to annotate the insulator states of the training data set, and the state categories include self-explosion, flashover, damage, and normal.

[0072] S30: Inputting the insulator image data collected in real time into the insulator defect detection model to obtain the insulator defect detection result.

[0073] based on Figure 1An insulator defect detection method is shown. By integrating StarNet into the YOLO11n network, it can improve the inference speed of the YOLO11n network through star operations and compress the YOLO11n network. Add a small target feature pyramid network to the neck network of the YOLO11n network to improve the detection ability of the YOLO11n network for small target defects in insulator images by fusing multiple special layers. And apply the ADown module to the backbone network and neck network of the YOLO11n network, which can reduce the resolution of the feature map, optimize the parameters of the YOLO11n network, and reduce the amount of computation. By replacing the IOU evaluation metric in the loss function of the YOLO11n network with the NWD evaluation metric, an improved YOLO11n network is obtained, enabling the YOLO11n network to accurately identify the position deviation of tiny targets (tiny targets contain very few pixels in the image). Finally, use the training dataset labeled with insulator status categories to train the improved YOLO11n network to obtain an insulator defect detection model that can accurately identify the categories of small target defects in insulator images. In summary, through the above solutions, the insulator defect detection model can better adapt to different lighting and weather conditions and the requirements of small target defect detection, improve the robustness and stability of the model in practical applications, and enhance the accuracy of identifying insulator defects.

[0074] When applying an insulator defect detection method provided by the present invention, it is not necessary to execute according to Figure 1 the order of the steps shown. The specific execution order of each step can be determined as needed, and the present invention does not limit this.

[0075] In addition, in one or more embodiments of the present invention, using the training dataset labeled with insulator status categories as the input and the insulator status category as the output, training the improved YOLO11n network to obtain an insulator defect detection model specifically includes:

[0076] Collect an image dataset containing multiple insulator statuses.

[0077] In this embodiment, multiple insulator statuses include but are not limited to: precision, recall rate, mean average precision (mAP), number of parameters, and amount of computation (GFLOPs).

[0078] Use the imgaug library to process the image dataset and simulate the weather in the image dataset to obtain an image dataset with enhanced features.

[0079] In this embodiment, by using the imgaug library to add noise, add black frames, rotate, blur, and simulate weather conditions to the image dataset, the diversity of the insulator image dataset can be enhanced, the requirement for the completeness of the image data can be reduced, and the training effect of the insulator defect detection model can be improved.

[0080] The feature-enhanced image dataset is divided into a training dataset and a validation dataset, and the insulator state categories are labeled for the training dataset and the validation dataset to obtain a training dataset labeled with insulator state categories and a validation dataset labeled with insulator state categories.

[0081] In this embodiment, considering the influence of factors such as inspection lighting conditions, camera orientation, and weather changes on the detection accuracy, the Insulator defect dataset_2024 dataset of 15,481 images is constructed and divided into a training dataset, a validation dataset, and a test dataset according to a ratio of 7:1.5:1.5. Finally, the insulator states of each image are labeled using the labeling software.

[0082] Taking the training dataset labeled with insulator state categories as the input and the insulator state categories as the output, the improved YOLO11n network is trained, and the parameters of the improved YOLO11n network are adjusted using the validation dataset labeled with insulator state categories during the training process to obtain an insulator defect detection model.

[0083] The solution shown in this embodiment can enhance the diversity of the insulator image dataset, reduce the requirement for the completeness of the image data, reduce the false detection and missed detection rates, and improve the training effect of the insulator defect detection model by using the imgaug library to add noise, add black frames, rotate, blur, and simulate weather conditions to the image dataset containing insulators, enabling the insulator defect detection model to still maintain a high detection accuracy for insulator defects in the image under complex environments.

[0084] Optionally, the parameter settings for training the improved YOLO11n network are as follows:

[0085] The size of the input image data is set to 640×640, the number of images per batch is set to 32, the number of training epochs is set to 300, the number of threads used by the cpu during data loading is set to 16, and both the initial learning rate and the termination learning rate are set to 0.01.

[0086] In addition, in one or more embodiments of the present invention, after obtaining the insulator defect detection model, it further includes:

[0087] Input the test data set labeled with insulator status categories into the insulator defect detection model to obtain the test results.

[0088] In this embodiment, the acquisition method of the test data set labeled with insulator status categories includes: using the validation data set labeled with insulator status categories as the test data set labeled with insulator status categories, or re-collecting image data about insulators and labeling the newly collected image data about insulators with insulator status categories to obtain the test data set labeled with insulator status categories.

[0089] Calculate multiple performance indicators of the insulator defect detection model according to the test results. The multiple performance indicators include precision, recall, mean average precision, number of parameters, and computational complexity.

[0090] In this embodiment, the model is evaluated using the test set, multiple indicators are calculated to measure the performance, and compared with mainstream algorithms to verify the effectiveness of the improved method, ensuring that the model is accurate, reliable, and has advantages. Among them, the following five indicators are used to evaluate the performance of the algorithm in object detection: Precision, Recall, mAP (mean Average Precision), number of parameters, and computational complexity (GFLOPs). In the comparative experiment, the following 4 evaluation indicators are mainly used: precision, recall, mAP_0.5, and mAP_0.5:0.95.

[0091] The calculation formulas are as follows:

[0092]

[0093] Among them, P is precision, R is recall, and N TP is the number of prediction bounding boxes (bounding box) in which the algorithm correctly identifies and locates the target object. N FP refers to the number of prediction bounding boxes in which the algorithm misidentifies the target object, that is, the IOU of the prediction bounding box with any real target object is lower than the set threshold. N FN refers to the number of target objects that the algorithm fails to identify, that is, the existing real target objects are not detected by the algorithm.

[0094] Specifically, the calculation formula for mean average precision is:

[0095]

[0096] Among them, R is recall, P is precision, and AP is the mean average precision of one class of samples.

[0097] Specifically, when any one of the multiple performance indicators is less than the preset threshold, the training dataset is expanded, and the parameters of the improved YOLO11n network are adjusted until all the multiple performance indicators of the trained insulator defect detection model are greater than the preset threshold.

[0098] In this embodiment, by calculating the multiple performance indicators of the insulator defect detection model, and comparing and verifying the multiple performance indicators with the performance of the mainstream algorithm. When any one of the multiple performance indicators is less than the preset threshold, the training dataset is expanded, and the parameters of the improved YOLO11n network are adjusted until all the multiple performance indicators of the trained insulator defect detection model are greater than the preset threshold, which can ensure the reliability of the insulator defect detection model.

[0099] In addition, in one or more embodiments of the present invention, the ADown module is used to reduce the spatial resolution of the feature map, and the NWD evaluation metric is used to improve the accuracy of detecting small target defects.

[0100] In this embodiment, the ADown module is applied to reduce the resolution of the feature map, and the parameters are optimized to reduce the computational amount; the NWD evaluation metric is introduced to improve the loss function and enhance the detection accuracy of small target defects. The ADown module is introduced in the present invention. Among them, the number of input channels is defined as c1, and the number of output channels is defined as c2. Half of the output channel number c2 is set as the output channel numbers of the two internal convolutional layers, which are stored in the self.c convolutional module. The 2×2 average pooling is applied to the input feature map x with a stride of 1, so that the spatial size of the feature map is halved. Subsequently, the pooled feature map x is split into two equal parts in the channel dimension, which are respectively called x1 and x2. Two convolutional layers self.cv1 and self.cv2 are created: for applying the self.cv1 convolutional layer to x1, the processed feature map is still x1; for x2, first apply the maximum pooling with a window size of 3 and a stride of 2, and then process the pooled x2 through the self.cv2 convolutional layer. Finally, x1 and x2 are concatenated in the channel dimension to obtain the final output.

[0101] Secondly, since there are many tiny targets in the insulator defect detection, which contain very few pixels in the image, it makes it difficult for the current object detection algorithm based on the Intersection over Union (IOU) evaluation metric to accurately identify the defects when dealing with the position deviation of tiny targets. To solve the above problems, the present invention introduces the Normalized Wasserstein Distance (NWD) to replace the IOU evaluation metric of the loss function in the model.

[0102] For a horizontal bounding box R = (cx, cy, 1, h), where (cx, cy) are the center coordinates, 1 and h are the width and height respectively, the probability density function of the two-dimensional Gaussian distribution is:

[0103]

[0104] where x, μ, and ∑ are the coordinate (x, y), the mean vector of the Gaussian distribution, and the covariance matrix respectively.

[0105] When (x - μ) T ∑ -1 (x - μ) = 1, it can be modeled as a two-dimensional Gaussian distribution N(μ, ∑), where:

[0106] When calculating the Wasserstein distance between two two-dimensional Gaussian distributions, assuming the two two-dimensional Gaussian distributions are μ1 = N(m1, ∑1) and μ2 = N(m2, ∑2), the Wasserstein distance can be defined as:

[0107]

[0108] Its simplified form is:

[0109] where ||.|| is the Frobenius norm. For the bounding box It can be further simplified to:

[0110]

[0111] However, since is a distance metric and cannot be directly used as a similarity metric. Therefore, the present invention normalizes the exponential form of to obtain the expression of NWD as:

[0112]

[0113] The above is a method for detecting insulator defects provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding insulator defect detection device, as Figure 6 shown, including:

[0114] An optimization module, configured to integrate the StarNet into the YOLO11n network, add a small target feature pyramid network to the neck network of the YOLO11n network, apply the ADown module to the backbone network and the neck network of the YOLO11n network, and replace the IOU evaluation metric of the loss function of the YOLO11n network with the NWD evaluation metric to obtain an improved YOLO11n network.

[0115] A training module, which takes a training data set labeled with insulator state categories as input, and the insulator state categories as output, trains an improved YOLO11n network to obtain an insulator defect detection model.

[0116] A detection module, which inputs the insulator image data collected in real time into the insulator defect detection model to obtain the insulator defect detection result.

[0117] For the specific limitations of an insulator defect detection device, reference can be made to the limitations of an insulator defect detection method in the above text, which will not be elaborated here. Each module in the insulator defect detection device can be implemented in whole or in part by software, hardware, and their combinations. Each module can be embedded in or independent of the processor in the computer device in the form of hardware, or stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0118] The present invention also provides a computer-readable storage medium, which stores a computer program that can be used to execute the Figure 1 insulator defect detection method provided.

[0119] The present invention also provides Figure 7 a schematic structural diagram of the computer device shown, as Figure 7 shown, at the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory. Of course, there may also be other hardware required for other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the Figure 1 insulator defect detection method provided.

[0120] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the described embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the various methods. Among them, any reference to a memory, storage, database, or other medium used in the various embodiments provided by the present invention can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0121] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered to be within the scope recorded by the present invention.

Claims

1. A method for detecting insulator defects, characterized in that: include: The star network StarNet is integrated into the YOLO11n network, and the small target feature pyramid network is added to the neck network of the YOLO11n network, and the ADown module is applied to the backbone network and the neck network of the YOLO11n network, and the IOU evaluation index of the loss function of the YOLO11n network is replaced by the NWD evaluation index to obtain an improved YOLO11n network; Taking a training data set labeled with insulator state categories as input and the insulator state categories as output, training the improved YOLO11n network to obtain an insulator defect detection model; The insulator image data collected in real time is input into the insulator defect detection model to obtain the defect detection result of the insulator.

2. The insulator defect detection method according to claim 1, characterized in that: The training data set labeled with the insulator state category is used as input, the insulator state category is used as output, and the improved YOLO11n network is trained to obtain the insulator defect detection model, which specifically includes: Acquire image datasets containing various insulator states; The image dataset is processed using the imgaug library, and the weather in the image dataset is simulated to obtain a feature-enhanced image dataset; Dividing the feature-enhanced image data set into a training data set and a verification data set, and labeling the training data set and the verification data set with insulator state categories to obtain a training data set labeled with insulator state categories and a verification data set labeled with insulator state categories; The training data set labeled with the insulator state category is used as input, and the insulator state category is used as output to train the improved YOLO11n network. During the training process, the parameters of the improved YOLO11n network are adjusted using the verification data set labeled with the insulator state category to obtain the insulator defect detection model.

3. The insulator defect detection method according to claim 1, characterized in that: After the insulator defect detection model is obtained, the method further includes: Inputting a test data set marked with insulator status categories into the insulator defect detection model to obtain a test result; Calculating multiple performance indicators of the insulator defect detection model according to the test results, wherein the multiple performance indicators include accuracy, recall rate, average precision, parameter quantity and calculation quantity; When any one of the multiple performance indicators is less than a preset threshold, the training data set is expanded, and the parameters of the improved YOLO11n network are adjusted until multiple performance indicators of the trained insulator defect detection model are greater than the preset threshold.

4. The insulator defect detection method according to claim 1, characterized in that: The star network is composed of a stack of four-order hierarchical structures, and the specific construction method includes: The convolutional layer is used for downsampling, and the modified demoblock is used for feature extraction, and the GELU function in the demoblock is replaced by the RelU6 function.

5. The insulator defect detection method according to claim 1, characterized in that: The small target feature pyramid network includes: SPDConv layer and CSP-Omnikernel layer; The SPDConv layer is used to downsample the feature map in the spatial dimension and expand the channel dimension; The CSP-Omnikernel layer includes a global branch, a large branch, and a local branch, which are used to learn feature representations from global to local.

6. The insulator defect detection method according to claim 1, characterized in that: The ADown module is used to reduce the spatial resolution of the feature map, and the NWD evaluation index is used to improve the accuracy of detecting small target defects.

7. An insulator defect detection device, characterized in that: include: An optimization module is used to integrate the star network StarNet into the YOLO11n network, and add the small target feature pyramid network to the neck network of the YOLO11n network, and apply the ADown module to the backbone network and the neck network of the YOLO11n network, and replace the IOU evaluation index of the loss function of the YOLO11n network with the NWD evaluation index to obtain an improved YOLO11n network; A training module, used to take a training data set labeled with insulator state categories as input, take the insulator state categories as output, train the improved YOLO11n network, and obtain an insulator defect detection model; The detection module is used to input the insulator image data collected in real time into the insulator defect detection model to obtain the defect detection result of the insulator.

8. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, an insulator defect detection method according to any one of claims 1 to 6 is implemented.

9. A computer device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, an insulator defect detection method according to any one of claims 1 to 6 is implemented.

Images