A transmission line insulator defect detection method based on adaptive enhanced module

Abstract

The present invention provides a transmission line insulator defect detection method with an adaptive enhancement module, comprising the following steps: obtaining image data of insulators of a high-altitude transmission line: using a high-definition camera of an unmanned aerial vehicle to shoot insulator defect images at different angles, performing preprocessing, and forming an insulator defect image dataset; constructing an insulator defect target detection model: a high-altitude transmission line insulator defect target detection model based on a DyHead and an information adaptive enhancement module; training the insulator defect target detection model; and detecting defects in insulator images: inputting the preprocessed insulator defect image to be detected into the trained target detection model, obtaining a classification and positioning structure of various insulator defects in the defect image, and detecting the defect position.

Description

Technical Field

[0001] The present invention relates to the field of transmission line defect detection, and in particular to a transmission line insulator defect detection method with an adaptive enhancement module. Background Art

[0002] Overhead transmission lines are a vital component of modern power systems, and their stable operation is directly related to the safe transmission of electricity. However, over the long term, and due to environmental influences such as wind, rain, ice, snow, and dirt, insulators along these lines are prone to various defects, such as cracks, breakage, and contamination. These defects can degrade insulator performance and, in severe cases, even cause power system failures, impacting the normal operation of the grid. Therefore, timely and accurate detection and diagnosis of insulator defects is crucial to ensuring the safety and reliability of power systems.

[0003] Traditional insulator defect detection methods often rely on manual inspections and simple electrical tests, often requiring specialized personnel to work at height to perform repairs and replacements. This is not only inefficient and costly, but also susceptible to human error.

[0004] In recent years, deep learning technology has made significant progress in image recognition and classification, providing a new solution for detecting insulator defects on high-altitude transmission lines. Furthermore, with the development of drone technology, researchers often first use drones to film high-altitude transmission lines. The drones' high-definition cameras can capture high-resolution images, and then use pre-trained deep learning models to classify and identify insulator defects. However, complex and diverse backgrounds, combined with insulators of varying types and shapes, can lead to weak resolution, missed detections, and false detections, making it difficult for the model to detect the target.

[0005] Therefore, how to make the network have stronger detection and differentiation capabilities for insulator defects and improve the detection accuracy of insulators in complex scenarios has become a technical problem that needs to be solved urgently. Summary of the Invention

[0006] The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and provide a method for detecting transmission line insulator defects with an adaptive enhancement module, thereby solving the problems of difficulty in identifying transmission line insulator defect models and low detection accuracy in the prior art.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] The present invention provides a method for detecting defects in insulators of a transmission line using an adaptive enhancement module, comprising:

[0009] S1. Acquisition of image data of insulators on high-altitude transmission lines: Use a high-definition camera on a drone to capture images of insulator defects from different angles, perform preprocessing, and form an insulator defect image dataset;

[0010] S2. Constructing an insulator defect target detection model: A high-altitude transmission line insulator defect target detection model based on DyHead and information adaptive enhancement module;

[0011] S3. Training of insulator defect target detection model;

[0012] S4. Insulator image defect detection: The pre-processed insulator defect image to be detected is input into the trained target detection model to obtain the classification and positioning structure of various insulator defects in the defect image and detect the defect location.

[0013] Furthermore, in said S2, an insulator defect target detection model is constructed including a deep residual network, a feature fusion network and a detection head network;

[0014] The deep residual network is based on the construction of the deep convolutional neural network ResNet50, specifically:

[0015] The deep convolutional neural network ResNet50 introduces a residual block, which includes a convolutional layer and an activation function, specifically:

[0016] y=F(x)+x (1);

[0017] Among them, x is the input; F(x) is the output after the convolution layer and activation function, through the addition operation;

[0018] The feature fusion network is based on the construction of the FPN network and the information adaptive enhancement module, specifically:

[0019] In the information adaptive enhancement module, the feature map is processed using adaptive average pooling AdaptiveAvgPool2d(×). For a feature map X of a certain layer, ij :

[0020]

[0021] in, is the feature after pooling; i, j are the ordinal numbers of the input feature level and the enhanced feature level respectively; σ and μ are the mean and standard deviation in x; ε is a small positive number added to stabilize the division; γ and β are trainable parameters;

[0022] For features lower than the feature layer, a 3×3 convolution with a stride of 2 is used to reduce the feature until the feature scale is the same; for features higher than the feature layer, a 3×3 transposed convolution with a stride of 2 is used to enlarge the feature until the feature scale is the same:

[0023]

[0024] Among them, conv(·) is a 3×3 convolution with a stride of 2; conv trans (·) is a 3×3 transposed convolution with a stride of 2; is the output feature;

[0025] right Perform point-by-point convolution PWconv(·) to obtain the weight parameter W ij , to express the attention of the feature map of the level to a certain level:

[0026]

[0027] After obtaining the initial weight parameters of all layers, they are mapped to the range (0, 1) through the softmax function to obtain the final weight parameters

[0028]

[0029] Among them, a and b are the level ordinal numbers;

[0030] Perform weighted summation on the scaled hierarchical features and the corresponding final weight parameters to obtain the enhanced feature map Y j :

[0031]

[0032] in, is element-wise multiplication; c is the total number of all levels;

[0033] Add a shortcut connection path to the enhanced features to obtain the final output features

[0034] Among them, X j It is the input feature of the information adaptive enhancement module.

[0035] Furthermore, the detection head network is constructed based on the DyHead module;

[0036] The DyHead module includes a scale-aware attention module, a space-aware attention module, and a task-aware attention module;

[0037] The scale-aware attention module π L(·) Dynamically fuse features based on the semantic importance of features at different scales:

[0038]

[0039] Where F is the input feature vector; S and C are the total number of width-height products and the total number of channels, respectively; f(·) is a linear function approximated by a 1×1 convolution; α(·) is a hard-sigmoid function;

[0040] The spatial perception attention module π S (·) Specifically:

[0041]

[0042] Where L is the number of feature levels; l is the current feature level number; K is the number of sparse sampling positions; k is the current sparse sampling position number; p k +Δp k Focus on a specific area by position shift; Δm k is the position p k The self-learnable importance measurement factor of

[0043] The task-aware attention module π C (·) Select different tasks by dynamically opening or closing feature channels:

[0044] π L (F)·F=max(α 1 (F)·F+β 1 (F),α 2 (F)·F+β 2 (F));

[0045] Among them, [α 1 ,β 1 ,α 2 ,β 2 ] T =θ(·) is a hyperfunction used to learn and control the activation threshold.

[0046] Furthermore, in S3, the training of the insulator defect target detection model includes processing of the data set, definition of the loss function, and training of the model parameters:

[0047] The definition of the loss function includes classification loss and regression loss;

[0048] The classification loss uses the cross entropy loss L cls :

[0049]

[0050] Among them, yi is the label of the true category; p i The category probability predicted by the model; C is the target for each prediction;

[0051] The regression loss uses the GIoU loss function:

[0052]

[0053] Among them, A and B are the areas of the predicted bounding box and the true bounding box respectively, and C is the area of ​​the minimum bounding rectangle that can contain A and B.

[0054] Furthermore, the model parameters are trained using the stochastic gradient descent method, where only part of the data is used for calculation each time the parameters are updated, and the parameters are updated by calculating the gradient of the loss:

[0055]

[0056] Among them, θ is the model parameter; η is the learning rate; is the gradient of the loss function with respect to the parameters.

[0057] The beneficial effects of the present invention are: by utilizing the information adaptive enhancement module, various types of insulator defects can be effectively learned;

[0058] For the insulator defect detection task, the multi-dimensional attention mechanism is introduced into the method, which can accurately identify the location of insulator defects in complex scenes and at different angles. Compared with existing methods, the recognition accuracy is higher. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 The present invention is a flow chart of a method for detecting defects in transmission line insulators using an adaptive enhanced module. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0061] See also Figure 1 , a transmission line insulator defect detection method with an adaptive enhancement module, comprising:

[0062] S1. Acquisition of image data of insulators on high-altitude transmission lines: Use a high-definition camera on a drone to capture images of insulator defects from different angles, perform preprocessing, and form an insulator defect image dataset;

[0063] In a specific embodiment, obtaining image data of an overhead transmission line insulator includes the following steps:

[0064] First, fully plan the terrain, weather conditions, and line conditions of the transmission line; specify the flight route based on the existing transmission line map; ensure that the drone and payload are in good condition and have sufficient battery power;

[0065] Afterwards, choose a multi-rotor drone or fixed-wing drone equipped with a high-definition camera and fly it along the preset route, maintaining a safe flight altitude and distance;

[0066] Finally, the system hovers over the insulator at the optimal shooting angle to take high-definition photos. After the shooting is completed, the data is downloaded from the drone to a computer to construct an insulator defect dataset.

[0067] Before the drone takes flight, a route survey is conducted to develop a suitable flight path. The drone then flies automatically along the set route, maintaining a fixed distance from the transmission line to ensure accurate image acquisition at each insulator. The acquired image data is then used to mark the locations of insulator defects and create a dataset. The dataset is then randomly selected at a ratio of 8:2 to serve as the training set and test set, respectively.

[0068] S2. Constructing an insulator defect target detection model: A high-altitude transmission line insulator defect target detection model based on DyHead and information adaptive enhancement module;

[0069] S3. Training of insulator defect target detection model;

[0070] S4. Insulator image defect detection: The pre-processed insulator defect image to be detected is input into the trained target detection model to obtain the classification and positioning structure of various insulator defects in the defect image and detect the defect location.

[0071] In said S2, an insulator defect target detection model is constructed including a deep residual network, a feature fusion network and a detection head network;

[0072] In a specific embodiment, a deep parameter network is used to extract features for an image, which is then input into a feature fusion network to further fuse the rich feature information, and finally into a detection head network to complete the classification and positioning of the task.

[0073] The deep residual network is based on the construction of the deep convolutional neural network ResNet50.

[0074] Among them, the deep residual network is based on the deep convolutional neural network ResNet50, which consists of a 50-layer deep network. It is used to solve the problems of gradient disappearance and gradient explosion in deep neural networks, so that the network can be trained deeper.

[0075] Specifically:

[0076] (1) Construct a deep residual network.

[0077] The deep residual network uses ResNet50, which is mainly composed of convolutional layers, batch normalization, ReLU activation function, pooling layers and residual blocks. It can be divided into the following stages:

[0078] 11) The first convolution layer.

[0079] The first convolution layer uses a large 7×7 kernel with 64 channels and a stride of 2, primarily for initial feature extraction of the input image. This is followed by a 3×3 max pooling layer with a stride of 2 to reduce the size of the feature map.

[0080] 12) Convolutional layers and residual blocks.

[0081] ResNet50 consists of 4 stages, each of which consists of multiple residual blocks, and each residual block has multiple convolutional layers:

[0082] Stage 1: 3 residual blocks, each block consists of 1×1 convolution, 3×3 convolution and 1×1 convolution component;

[0083] Stage 2: 4 residual blocks, the number of convolution channels gradually increases to 64, 128, 256 and 512 channels respectively;

[0084] Stage 3: 6 residual blocks, gradually increasing the depth of the network;

[0085] Stage 4: 3 residual blocks to extract deep features;

[0086] 13) Quick connection path

[0087] The deep convolutional neural network ResNet50 introduces a residual block, which includes a convolutional layer and an activation function. The input is directly added to the output, allowing the network to directly transmit the input signal, avoiding the gradient problem caused by the deep network.

[0088] Specifically:

[0089] y=F(x)+x (1);

[0090] Among them, x is the input; F(x) is the output after the convolution layer and activation function. Through the addition operation, the network can be more easily optimized and trained.

[0091] 14) Global Average Pooling

[0092] After the last convolution operation, global average pooling is used to reduce the dimension of the feature map to a fixed-size feature vector;

[0093] 15) Fully connected layer

[0094] After global average pooling, it is connected to a fully connected layer and outputs a vector of length equal to the number of categories, which is usually the number of categories of the image in classification tasks.

[0095] 16) Activation function and normalization.

[0096] Each convolutional layer is followed by a ReLU activation function and batch normalization, which together help speed up the training process and stabilize the performance of the model.

[0097] The feature fusion network is based on the construction of FPN network and information adaptive enhancement module.

[0098] Among them, the FPN network consists of a bottom-up feature extraction path and a top-down feature fusion path, and then the information adaptive enhancement module is used to further enhance the fused features;

[0099] The bottom-up feature extraction path uses the deep residual network ResNet50 as the backbone network. The input image is pooled through multiple consecutive convolution kernels to generate feature maps of different scales, namely C2, C3, C4, and C5. These feature images gradually reduce the resolution and have deeper and deeper semantic information.

[0100] In a specific implementation, FPN uses several important feature layers of ResNet50 as the input of FPN, which come from convolution blocks C2, C3, C4, and C5 of different depths. The resolution of these feature maps decreases successively, while the number of channels increases successively, which are 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the input resolution respectively.

[0101] The top-down feature fusion path amplifies the high-level feature maps by upsampling, and then fuses them with the feature maps of the corresponding layers from the bottom up. The fusion method is element-by-element addition. To ensure that the fused feature maps have the same number of channels, 1╳1 convolution is used to swap the channels of the bottom-up feature maps C2, C3, C4, and C5 to match the number of channels of the upsampled feature maps.

[0102] In a specific embodiment, upsampling is performed starting from C5 and fused layer by layer with the feature maps of the shallow layers. The output P5 of C5 is upsampled by a factor of 2 to match its resolution with C4:

[0103] P5 ↑ =Upsample(P5);

[0104] Use 1x1 convolution to transform the channel of C4 and reduce the number of channels to 256:

[0105] L4 = conv(C4, 256);

[0106] Add the upsampled P5 to the 1x1 convolution-processed C4 to generate P4:

[0107] P4=P5 ↑ +L4;

[0108] In this way, P4 inherits deep semantic information while maintaining a high spatial resolution. By repeating similar steps, feature maps of different scales of P2, P3, P4, and P5 can be generated.

[0109] This path can transfer high-level semantic feature maps layer by layer to low-level feature maps, so that the deep semantic information in high-level features can be integrated with the detailed features of the low-level features.

[0110] The information adaptive enhancement module is used to further process the features after top-down path fusion, enhance the information content of the input features, and use the target position information of the low-resolution high-level features to enhance the low-resolution high-level features. At the same time, each feature is quickly linked to ensure that the original information of each feature is preserved while enhancing the perception of other layers of information. This improves the model's ability to identify the category and location of insulator defects in the responsible environment, improves the representativeness of insulator defect features, and facilitates the subsequent processing of the DyHead detection head.

[0111] Specifically, in the information adaptive enhancement module, the feature map is processed using the adaptive average pooling AdaptiveAvgPool2d(×). For a feature map X of a certain layer, ij :

[0112]

[0113] in, is the feature after pooling; i, j are the ordinal numbers of the input feature level and the enhanced feature level respectively; σ and μ are the mean and standard deviation in x; ε is a small positive number added to stabilize the division; γ and β are trainable parameters;

[0114] For features lower than the feature layer, a 3×3 convolution with a stride of 2 is used to reduce the feature until the feature scale is the same; for features higher than the feature layer, a 3×3 transposed convolution with a stride of 2 is used to enlarge the feature until the feature scale is the same:

[0115]

[0116] Among them, conv(·) is a 3×3 convolution with a stride of 2; conv trans (·) is a 3×3 transposed convolution with a stride of 2; is the output feature;

[0117] right Perform point-by-point convolution PWconv(·) to obtain the weight parameter W ij , to express the attention of the feature map of the level to a certain level:

[0118]

[0119] After obtaining the initial weight parameters of all layers, they are mapped to the range (0, 1) through the softmax function to obtain the final weight parameters

[0120]

[0121] Among them, a and b are the level ordinal numbers;

[0122] Perform weighted summation on the scaled hierarchical features and the corresponding final weight parameters to obtain the enhanced feature map Y j :

[0123]

[0124] in, is element-wise multiplication; c is the total number of all levels;

[0125] Add a shortcut connection path to the enhanced features to obtain the final output features

[0126] Among them, X j It is the input feature of the information adaptive enhancement module.

[0127] The detection head network is based on the construction of the DyHead module;

[0128] The DyHead module includes a scale-aware attention module, a space-aware attention module, and a task-aware attention module; the scale-aware attention module, the space-aware attention module, and the task-aware attention module are stacked to realize attention operation on input features, thereby enhancing the perception ability of the target detector for the scale-aware attention module, the space-aware attention module, and the task-aware attention module.

[0129] The scale-aware attention module π L (·) Dynamically fuse features based on the semantic importance of features at different scales:

[0130]

[0131]

[0132] Where F is the input feature vector; S and C are the total number of width-height products and the total number of channels, respectively; f(·) is a linear function approximated by a 1×1 convolution; α(·) is a hard-sigmoid function;

[0133] Among them, the spatial perception attention module consists of two steps: first, using deformable convolution for sparsification, and then aggregating features across layers at the same spatial location.

[0134] The spatial perception attention module π S (·) Specifically:

[0135]

[0136] Where L is the number of feature levels; l is the current feature level number; K is the number of sparse sampling positions; k is the current sparse sampling position number; p k +Δp k Focus on a specific area by position shift; Δm k is the position p k The self-learnable importance measurement factor of

[0137] The task-aware attention module π C (·) Select different tasks by dynamically opening or closing feature channels:

[0138] π L (F)·F=max(α 1 (F)·F+β 1 (F),α 2 (F)·F+β 2 (F));

[0139] Among them, [α 1 ,β 1 ,α 2 ,β 2 ] T =θ(·) is a hyperfunction used to learn and control the activation threshold. Specifically, global pooling is first performed in the L×S dimension, followed by two fully connected layers and a normalization layer, and finally a sigmoid function is used to normalize the output.

[0140] The dataset was processed by capturing images of high-altitude transmission line insulators through drone photography. Professionals then manually annotated the insulators and their defects in the images, including the bounding boxes of the insulators and labels of the defect types. The images were then converted into a COCO-formatted dataset.

[0141] In S3, the training of the insulator defect target detection model includes processing the data set, defining the loss function, and training the model parameters:

[0142] The definition of the loss function includes classification loss and regression loss;

[0143] The classification loss uses the cross entropy loss L cls , to measure the difference between the predicted probability of each category and the true label;

[0144]

[0145] Among them, y i is the label of the true category; p i The category probability predicted by the model; C is the target for each prediction;

[0146] The regression loss uses the GIoU loss function:

[0147]

[0148] Among them, A and B are the areas of the predicted bounding box and the true bounding box respectively, and C is the area of ​​the minimum bounding rectangle that can contain A and B.

[0149] The model parameters are trained using the stochastic gradient descent method. Only part of the data is used for calculation each time the parameters are updated, and the parameters are updated by calculating the gradient of the loss:

[0150]

[0151] Among them, θ is the model parameter; η is the learning rate; is the gradient of the loss function with respect to the parameters.

[0152] The model parameter update process is repeated over multiple training rounds. In each round, the model iterates over the entire training data until the loss function converges or the preset number of training rounds is reached.

[0153] The above-described embodiments merely illustrate the implementation methods of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for detecting defects in insulators of transmission lines using an adaptive enhancement module, characterized in that: include: S1. Acquisition of image data of insulators on high-altitude transmission lines: Use a high-definition camera on a drone to capture images of insulator defects from different angles, perform preprocessing, and form an insulator defect image dataset; S2. Constructing an insulator defect target detection model: A high-altitude transmission line insulator defect target detection model based on DyHead and information adaptive enhancement module; S3. Training of insulator defect target detection model; S4. Insulator image defect detection: The pre-processed insulator defect image to be detected is input into the trained target detection model to obtain the classification and positioning structure of various insulator defects in the defect image and detect the defect location; In said S2, an insulator defect target detection model is constructed including a deep residual network, a feature fusion network and a detection head network; The deep residual network is based on the construction of the deep convolutional neural network ResNet50, specifically: The deep convolutional neural network ResNet50 introduces a residual block, which includes a convolutional layer and an activation function, specifically: （1）； in, is the input; is the output after the convolution layer and activation function, through the addition operation; The feature fusion network is based on the construction of the FPN network and the information adaptive enhancement module, specifically: In the information adaptive enhancement module, adaptive average pooling is used Process the feature map. For a feature map of a certain layer : ； in, is the feature after pooling; are the input feature level ordinal and the enhanced feature level ordinal respectively; and yes The mean and standard deviation of ; A small positive number added to stabilize the division; and is a trainable parameter; For features lower than the feature layer, a 3×3 convolution with a stride of 2 is used to reduce the feature until the feature scale is the same; for features higher than the feature layer, a 3×3 transposed convolution with a stride of 2 is used to enlarge the feature until the feature scale is the same: ； in, It is a 3×3 convolution with a stride of 2; It is a 3×3 transposed convolution with a stride of 2; is the output feature; right Perform point-by-point convolution , get the weight parameter , to express the attention of the feature map of the level to a certain level: ； After getting the initial weight parameters of all levels, The function is mapped to the range (0, 1) to obtain the final weight parameter ; ； in, is the hierarchical ordinal number; Perform weighted summation on the scaled hierarchical features and the corresponding final weight parameters to obtain the enhanced feature map : ； in, is element-wise multiplication; is the total number of all levels; Add a shortcut connection path to the enhanced features to obtain the final output features : ； in, It is the input feature of the information adaptive enhancement module; The detection head network is based on the construction of the DyHead module; The DyHead module includes a scale-aware attention module, a space-aware attention module, and a task-aware attention module; The scale-aware attention module Dynamically fuse features based on their semantic importance at different scales: ； ； in, is the input feature vector; and are the total number of width-height products and the total number of channels respectively; is a linear function approximated by a 1×1 convolution; is the hard-sigmoid function; The spatial perception attention module Specifically: ； in, is the number of feature levels; is the current feature level number; is the number of locations for sparse sampling; is the position number of the current sparse sampling; Focus on specific areas through position shifting; For location A self-learnable importance metric; The task-aware attention module Select different tasks by dynamically turning on or off feature channels: ； in, is a hyperfunction used to learn and control the activation threshold.

2. The method for detecting defects in insulators of a transmission line using an adaptive enhancement module according to claim 1, characterized in that: In S3, the training of the insulator defect target detection model includes data set processing, loss function definition, and model parameter training: The definition of the loss function includes classification loss and regression loss; The classification loss uses cross entropy loss : ； in, is the label of the true category; The class probability predicted by the model; For each predicted target; The regression loss uses the GIoU loss function: ； ； Among them, A and B are the areas of the predicted bounding box and the true bounding box respectively, and C is the area of ​​the minimum bounding rectangle that can contain A and B.

3. The method for detecting defects in transmission line insulators using an adaptive enhancement module according to claim 2, wherein: The model parameters are trained using the stochastic gradient descent method. Only part of the data is used for calculation each time the parameters are updated, and the parameters are updated by calculating the gradient of the loss: ； in, are model parameters; is the learning rate; is the gradient of the loss function with respect to the parameters.

Images