Insulator icing identification and detection method based on convolutional network LDKA-NET

Abstract

The invention discloses an insulator icing identification and detection method based on a convolutional network LDKA-NET, and the method comprises the steps: firstly, providing a wide-field view convolution module, employing a larger convolution kernel to improve the perception capability and generalization capability of a network, and enabling the network to better adapt to a complex power transmission line scene; secondly, proposing a full-dimension dynamic convolution feature fusion network; finally, an expectation maximization dynamic convolution attention mechanism is adopted to pay attention to and utilize important information in input data, the model is helped to better distribute attention and pay attention to important targets or scene parts, more generalized feature representation is learned, the generalization ability and robustness of the model are improved, and the position of the target is located more accurately.

Description

Technical Field

[0001] The invention belongs to the field of electric power technology, and in particular relates to a method for identifying and detecting ice coating on insulators of power transmission lines. Background Art

[0002] With the increasing demand and dependence on electricity in modern society, the safe and stable operation of transmission lines as important channels for power transmission is particularly important. However, a large number of insulators inevitably have to pass through complex environments such as high altitudes and heavy ice, and are exposed to the wild all year round. In addition, the climate conditions have gradually deteriorated in recent years, and haze, rain, freezing snow and other weather conditions have occurred from time to time, making the weather and geographical environment where insulators are located increasingly complex, which can easily cause insulators to be covered with ice, defective, flashover and other accidents, bringing serious safety hazards to the operation of the power system, and also causing significant losses to the social economy.

[0003] In view of this, in order to timely discover and solve insulator problems and improve the safety and reliability of transmission lines, insulator monitoring technology has gradually become one of the focuses of the power industry. By real-time monitoring of the status and performance of insulators, abnormal conditions of insulators can be discovered in time, and corresponding maintenance and replacement measures can be taken to ensure the safe operation of transmission lines.

[0004] At present, the methods for insulator defect detection include ultrasonic, infrared thermal imaging, vibration, partial discharge and visual inspection. Ultrasonic requires professional equipment and operators, and the cost is relatively high; infrared thermal imaging detection is greatly affected by environmental factors, is insensitive to certain defects, and cannot directly detect internal problems of insulators; vibration detection is greatly affected by external interference and has limited accuracy; partial discharge detection equipment is complex and requires professional operation, which is costly. Visual inspection is simple and intuitive, and can quickly find obvious defects on the surface of insulators.

[0005] There are two main methods for visual insulator defect detection: one based on machine learning image processing algorithms and the other based on deep learning algorithms. Machine learning image processing algorithms usually involve feature extraction and classification of images. First, image data is collected from the surface of the insulator, and then edge detection, color analysis, shape matching and other techniques are used to extract feature information from the image. Finally, support vector machines, random forests and other methods are used to classify and identify the extracted features to determine whether the insulator has defects.

[0006] Based on the deep learning algorithm, the deep neural network is used to learn and process the original image end-to-end. By building models such as deep convolutional neural network (CNN) and recurrent neural network (RNN), features and patterns are learned directly from image data to achieve automatic detection and identification of insulator defects.

[0007] However, the existing algorithms do not consider the problem of insufficient capture of local details caused by a small perception field of view, which may lead to the neglect of global context information. Secondly, the depth of the existing network model is still relatively shallow, and the ability to extract images with particularly complex environmental backgrounds or target occlusion is limited. At the same time, due to the complex background of the transmission line and the monotonous scenery on snowy days, background information interferes with defect detection. At the same time, the damage range of some insulators is small, and the damaged part occupies fewer pixels in the original image. Summary of the invention

[0008] Aiming at the problem of low target detection accuracy in complex environments, the present invention proposes an insulator ice detection method based on convolutional network LDKA-NET.

[0009] The technical solution of the present invention is an insulator ice recognition and detection method based on the convolutional network LDKA-NET, which is specifically implemented according to the following steps:

[0010] Step 1: Collect the defect image data of ice-covered insulators on the transmission line, annotate and divide the data set into a training set, a validation set and a test set;

[0011] Step 2: Establish the convolutional network LDKA-NET network model;

[0012] Step 3, input the images in the training set into the convolutional network LDKA-NET network model for training, and obtain the optimized LDKA-NET network model through training;

[0013] Step 4: Input the images in the test set into the optimized LDKA-NET network model for detection to obtain the ice-covered insulator defect detection result map.

[0014] The step 1 is specifically implemented according to the following steps:

[0015] Step 1.1, use the high-definition camera carried by the drone to shoot a large number of inspection videos, including images of ice-covered insulators of transmission lines under various complex backgrounds, as well as images of distant views, close views and close-ups, and keep the number of images of each type of ice-covered insulators consistent;

[0016] Step 1.2, rotating, flipping, scaling, and arbitrarily cropping the ice-covered insulator image obtained in step 1.1, thereby obtaining an expanded ice-covered insulator image;

[0017] Step 1.3: The sum of the ice-covered insulator image obtained in step 1.1 and the expanded ice-covered insulator image obtained in step 1.2 is used as the ice-covered insulator sample image library. A corresponding label file is created for each image sample in the ice-covered insulator sample image library. The label file meets the xml label file standard of the Pascal VOC format. The xml label file includes the image ID, image path, image name, pixel height and width of the image, where the pixel height and width of the image are represented by four coordinates of a rectangular box, which are x min ,y min ,x max ,y max , where (x min ,y min ) is the coordinate of the upper left vertex of the rectangle, (x max ,y max ) is the coordinate of the lower right vertex of the rectangular box;

[0018] Step 1.4: Divide the ice-covered insulator sample image library obtained in step 1.3 into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0019] The step 2 is specifically implemented according to the following steps:

[0020] Step 2.1, establish the convolutional network LDKA-NET network model, including image preprocessing, backbone network (Backbone), feature extraction network (Neck) and prediction network (Head); Step 2.2, the wide-field convolution backbone network structure is as follows:

[0021] The wide-field convolution backbone network is mainly composed of four stages, each of which contains the initial layer (stem), stage layer (Stage) and transition layer (Transition) convolution layer; the Depth Wise Conv layer is a depth-separable convolution, which consists of two parts, namely Depth Wise (DW) convolution and Point Wise (PW) convolution; each convolution kernel of the DW convolution only calculates the corresponding feature map, but it may weaken the information interaction between feature pixels in different layers, which may lead to accuracy loss; the PW convolution is a 1×1 convolution with an output of n channels, which aims to compensate for the lack of feature interaction in the DW convolution. Through point-to-point multiplication and addition operations, the PW convolution can combine the information of different feature maps to enhance the interaction between features;

[0022] Assume that the input feature map size is D w ×D w ×C, the convolution kernel size is D k ×D k ×C, the output feature map size is D w ×D w×C out , the number of parameters of the standard convolutional layer is as follows:

[0023] P Standard =(D k ×D k ×C)·C out Formula (1)

[0024] Among them, C is the number of input feature channels, C out Indicates the number of output feature channels and convolution kernels;

[0025] In addition, the depthwise convolution in the depthwise separable convolution is responsible for filtering, with a size of D k ×D k ×1, a total of C, acting on each channel of the input; point-by-point convolution is responsible for converting channels, size 1×1×C, a total of C out , acts on the output feature map of the deep convolution;

[0026] The depth-separable convolution parameter is as follows:

[0027]

[0028] The ratio of the number of depth convolution parameters to the number of standard convolution parameters is:

[0029]

[0030] Based on this, the use of depthwise separable convolution can solve the problem of increased parameters and floating-point operations caused by the use of large-size convolution kernels. The stem layer consists of 3×3 convolution with a stride of 2, depthwise separable convolution, and DW convolution with a stride of 2, which are used to increase the dimension and reduce the size of the input image; as shown in the following formula:

[0031]

[0032] As shown in formula (4), w is the feature weight, D represents the depth separable volume; assuming the input size is H and W, where H is the image height and W is the image width, the output feature size of the stem layer is H / 4×W / 4×C1, where C1 is the number of channels after dimensionality increase;

[0033] The Stage layer is composed of a stack of RepLK Block and ConvFFN. The RepLK Block contains a normalization layer, a 1×1 convolution, a depth-separable convolution, and an important residual connection. ConvFFN uses a 1×1 convolution to replace the fully connected layer and uses residual connections between layers. The Transition layer mainly uses PW convolution and DW convolution with a stride of 2 for image downsampling.

[0034] In addition, the use of a large core backbone network can obtain a larger receptive field of view with fewer layers and a higher shape deviation, as shown in the following formula:

[0035] RF n =RF n-1 ×S n +(K n -1) Formula (5)

[0036] The above formula is the calculation formula for the receptive field, where RF n is the receptive field size of the nth layer, RF n-1 is the receptive field size of the n-1th layer, S n is the step size of the nth layer, K n is the convolution kernel size of the nth layer;

[0037] Step 2.3, the network structure of the full-dimensional dynamic convolution feature fusion module is as follows:

[0038] The full-dimensional dynamic convolution is implemented as follows:

[0039]

[0040] In the above formula, α wi , α fi , α ci and α si They are the attention coefficients of the convolution kernel in space, channel, convolution filter and convolution kernel respectively. The four types of attention weights enable the convolution operation to be performed for all spatial positions, all input channels, all filters and all kernels of different inputs, providing performance guarantee for capturing rich contextual clues.

[0041] The full-dimensional feature fusion (Optimized Dynamic Convolution Coarse-to-Fine, ODC2F) is as follows:

[0042] It is necessary to learn the weights and biases of feature fusion, which will introduce a large number of additional parameters and may lead to the risk of overfitting. The formula is as follows:

[0043]

[0044] In the above formula, F in is the input feature, F out is the output feature; when Bottleneck is True, τ is 1, otherwise it is 0; λ is the number of feature fusion modules, and ω is the residual weight bias coefficient. Due to the relatively limited information exchange between feature maps, the module may lead to low target detection accuracy. The ODC2F formula is as follows:

[0045]

[0046] In the above formula, F o ” ut is the output feature; is the full-dimensional dynamic convolution described above. Through the full-dimensional dynamic convolution, the attention features are learned for the four dimensions of the convolution kernel space by using a parallel strategy, which can provide better performance to capture rich context information.

[0047] Step 2.4, the network structure of the Expectation-Maximization Dynamic Convolution Attention mechanism EM-DCA is as follows:

[0048] Let the input feature F ∈ R C×H×W , C represents the input channels, H and W represent the height and width of the input feature respectively. The feature F is divided into G sub-features along the channel dimension, G << C, then F can be expressed as:

[0049] F = [F 0 , F i , …, F G-1 , F i ∈ R C / / G×H×W Formula (12)

[0050] The EM-DCA attention mechanism has four branches, among which three parallel paths extract the attention weight descriptions of the grouped feature maps. Two 1x1 branches and one 3x3 branch. Two global average poolings are used to encode the channels in two spatial directions in the 1x1 branches, and a full-dimensional dynamic convolution is included in the 3x3 branch to capture multi-scale feature representations.

[0051] Two parallel 1D feature encoding vectors will share a 1x1 convolution, enabling the model to capture local cross-channel interactions. The formula for encoding the global information in the horizontal dimension of the channel C at height H is:

[0052]

[0053] The formula for the pooling output in the width W along the channel C is:

[0054]

[0055] The 3x3 branch captures local cross-channel interactions through convolution to expand the feature space.

[0056] Among them, Cross-spatial learning establishes mutual dependencies between channels and spaces to achieve richer feature aggregation. The two-dimensional global average pooling is used to encode the global spatial information in the output of the 1x1 branch, and the output of the smallest branch is transformed into the corresponding dimension shape The formula is:

[0057]

[0058] Finally, the output of each branch is calculated as an attention weight value, which combines the global semantic information to capture the key areas.

[0059] The step 3 is specifically implemented according to the following steps:

[0060] Step 3.1, input the training set images divided in step 1.4 into the LDKA-NET network model obtained in step 3 for training, set batch_size to 16, use stochastic gradient descent with momentum of 0.9, train for 400 rounds, and obtain the training results;

[0061] Step 3.2, use the validation set to validate the LDKA-NET network model, obtain the validation results, and analyze the differences between the training results and the validation results, such as whether they are overfitting or underfitting; if overfitting occurs, it means that the model performs very well on the training set, but very poorly on the validation set, that is, the generalization ability of the model is very poor, which needs to be solved by increasing the number of images in the training set, reducing the complexity of the model, and fine-tuning the hyperparameters; if underfitting occurs, it means that the model performs poorly on both the training set and the validation set, which needs to be solved by increasing the complexity of the model, fine-tuning the hyperparameters, and adding more target features;

[0062] Step 3.3, by adjusting the hyperparameters of the LDKA-NET network model on the validation set, fine-tuning the learning rate to 0.003 to obtain better robustness, setting batch_size to 16, and obtaining the optimized LDKA-NET network model. Among them, hyperparameters are parameters that need to be manually set before model training, such as learning rate and batch size.

[0063] The step 4 is specifically implemented according to the following steps:

[0064] Step 4.1, input the test set into the step 3.3 to obtain the optimized LDKA-NET model. The specific optimization process is as follows: set the hyperparameters of the LDKA-NET network model: input image pixel size 640*640, freeze training for 150 rounds, batch_size32, unfreeze training for 250 rounds, batch_size4, num_workers2, Adam optimizer, attenuation weight coefficient 5*10-4, initial learning rate 1*10-5, set the IoU threshold to 0.5 for testing when testing the test set, and fine-tune the learning rate to 0.003 during verification to obtain better robustness, set batch_size to 32, freeze training for the first 150 rounds, the loss decreases quickly, unfreeze training for the next 250 rounds, and continuously fine-tune the network. After 350 rounds out of a total of 400 rounds, the loss change of the verification set gradually decreases, and the optimized LDKA-NET model is obtained;

[0065] Step 4.2: Use the test set to test the performance of the optimized LDKA-NET model and obtain the test results.

[0066] The present invention has the following beneficial effects

[0067] First, a wide field view convolutional network (WFVCNet) is proposed, which uses larger convolution kernels to improve the network's perception and generalization capabilities and better adapt to complex transmission line scenarios. At the same time, the use of large-core CNN can obtain a larger effective receptive field with fewer layers, and has a higher shape deviation, which can better capture the shape information in the image.

[0068] Secondly, a full-dimensional dynamic convolution feature fusion network is proposed. Compared with traditional static convolution, this paper can significantly improve the accuracy of CNN by learning the linear combination of multiple convolution kernels and the weighted input-related attention, and has higher parameter efficiency and feature extraction capabilities.

[0069] Finally, the expectation maximization dynamic convolutional attention mechanism (EM-DCA) is used to focus on and utilize important information in the input data, helping the model to better allocate attention, focus on important targets or scene parts, and learn more generalized feature representations, which helps to improve the generalization ability and robustness of the model and more accurately locate the target. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1It is a network structure diagram of the insulator ice recognition and detection method based on the convolutional network LDKA-NET of the present invention, which consists of four parts: image preprocessing, backbone network (Backbone), feature extraction network (Neck) and prediction network (Head);

[0071] Figure 2 This is a structural diagram of the wide-field convolution backbone network of the algorithm of the present invention;

[0072] Figure 3 It is a full-dimensional dynamic convolution feature fusion network structure diagram of the present invention;

[0073] Figure 4 This is the network structure diagram of the expectation maximization dynamic convolutional attention mechanism EM-DCA of the present invention;

[0074] Figure 5 It is a sample diagram of iced insulator defect detection results in the insulator ice recognition and detection method based on the convolutional network LDKA-NET of the present invention;

[0075] Figure 6 This is a sample diagram of iced insulator defect detection results in rainy and snowy weather in the insulator ice recognition and detection method based on the convolutional network LDKA-NET of the present invention;

[0076] Figure 7 This is a sample diagram of iced insulator defect detection results in clear weather in the insulator icing identification and detection method based on the convolutional network LDKA-NET of the present invention. DETAILED DESCRIPTION

[0077] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0078] The insulator ice detection method of this embodiment is based on the convolutional network LDKA-NET, and is specifically implemented according to the following steps:

[0079] Step 1: Collect the defect image data of ice-covered insulators on the transmission line, annotate and divide the data set into a training set, a validation set and a test set;

[0080] Step 2: Establish the convolutional network LDKA-NET network model;

[0081] Step 3, input the images in the training set into the convolutional network LDKA-NET network model for training, and obtain the optimized LDKA-NET network model through training;

[0082] Step 4: Input the images in the test set into the optimized LDKA-NET network model for detection to obtain the ice-covered insulator defect detection result map.

[0083] Step 1 is specifically implemented as follows:

[0084] Step 1.1, use the high-definition camera carried by the drone to shoot a large number of inspection videos, including images of ice-covered insulators of transmission lines under various complex backgrounds, as well as images of distant views, close views and close-ups, and keep the number of images of each type of ice-covered insulators consistent;

[0085] Step 1.2, rotating, flipping, scaling, and arbitrarily cropping the ice-covered insulator image obtained in step 1.1, thereby obtaining an expanded ice-covered insulator image;

[0086] Step 1.3: The sum of the ice-covered insulator image obtained in step 1.1 and the expanded ice-covered insulator image obtained in step 1.2 is used as the ice-covered insulator sample image library. A corresponding label file is created for each image sample in the ice-covered insulator sample image library. The label file meets the xml label file standard of the Pascal VOC format. The xml label file includes the image ID, image path, image name, pixel height and width of the image, where the pixel height and width of the image are represented by four coordinates of a rectangular box, which are x min ,y min ,x max ,y max , where (x min ,y min ) is the coordinate of the upper left vertex of the rectangle, (x max ,y max ) is the coordinate of the lower right vertex of the rectangular box;

[0087] Step 1.4: Divide the ice-covered insulator sample image library obtained in step 1.3 into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0088] Step 2 is specifically implemented as follows:

[0089] Step 2.1. Establish the convolutional network LDKA-NET network model, which includes four parts: image preprocessing, backbone network (Backbone), feature extraction network (Neck) and prediction network (Head); among them, WFVC Net uses larger convolution kernels to enhance the network's perception and generalization capabilities, and better adapt to complex transmission line icing scenarios. Full-dimensional dynamic convolution greatly improves the accuracy of CNN by learning the linear combination of multiple convolution kernels and the weighting of input-related attention, and has higher parameter efficiency and feature extraction capabilities. Finally, in order to reduce the amount of calculation and maintain the information integrity of each channel, some feature channels are reorganized into batch data and divided into multiple groups through the expectation maximization dynamic convolution attention mechanism (EM-DCA), so that the spatial semantic features are evenly distributed in each feature group.

[0090] Step 2.2: The wide-field convolution backbone network structure is as follows:

[0091] like Figure 2 As shown in the figure, although the traditional CNN can obtain a larger receptive field by stacking multiple small convolution kernels, the computational complexity of the network will increase significantly as the number of layers increases, and it is easy to cause the problem of gradient disappearance or explosion. In order to solve this problem, the present invention proposes a wide-field convolution backbone network, which uses larger convolution kernels to improve the network's perception and generalization capabilities, and better adapt to the complex scenes in this article.

[0092] The wide-field convolution backbone network is mainly composed of four stages, each of which contains the initial layer (stem), stage layer (Stage) and transition layer (Transition) convolution layer; the Depth Wise Conv layer is a depth-separable convolution, which consists of two parts, namely Depth Wise (DW) convolution and Point Wise (PW) convolution; each convolution kernel of the DW convolution only calculates the corresponding feature map, but it may weaken the information interaction between feature pixels in different layers, which may lead to accuracy loss; the PW convolution is a 1×1 convolution with an output of n channels, which aims to compensate for the lack of feature interaction in the DW convolution. Through point-to-point multiplication and addition operations, the PW convolution can combine the information of different feature maps to enhance the interaction between features;

[0093] Assume that the input feature map size is D w ×D w ×C, the convolution kernel size is D k ×D k ×C, the output feature map size is D w ×D w ×C out , the number of parameters of the standard convolutional layer is as follows:

[0094] P Standard =(D k ×D k ×C)·C out Formula (1)

[0095] Among them, C is the number of input feature channels, C out Indicates the number of output feature channels and convolution kernels;

[0096] In addition, the depthwise convolution in the depthwise separable convolution is responsible for filtering, with a size of D k ×D k ×1, a total of C, acting on each channel of the input; point-by-point convolution is responsible for converting channels, size 1×1×C, a total of C out , acts on the output feature map of the deep convolution;

[0097] The depth-separable convolution parameter is as follows:

[0098]

[0099] The ratio of the number of depth convolution parameters to the number of standard convolution parameters is:

[0100]

[0101] Based on this, the use of depthwise separable convolution can solve the problem of increased parameters and floating-point operations caused by the use of large-size convolution kernels. The stem layer consists of 3×3 convolution with a stride of 2, depthwise separable convolution, and DW convolution with a stride of 2, which are used to increase the dimension and reduce the size of the input image; as shown in the following formula:

[0102]

[0103] As shown in the formula, w is the feature weight, D represents the depth separable volume; assuming the input size is H and W, where H is the image height and W is the image width, the output feature size of the stem layer is H / 4×W / 4×C1, where C1 is the number of channels after dimensionality increase;

[0104] The Stage layer is composed of a stack of RepLK Block and ConvFFN. The RepLK Block contains a normalization layer, a 1×1 convolution, a depth-separable convolution, and an important residual connection. ConvFFN uses a 1×1 convolution to replace the fully connected layer and uses residual connections between layers. The Transition layer mainly uses PW convolution and DW convolution with a stride of 2 for image downsampling.

[0105] In addition, the use of a large core backbone network can obtain a larger receptive field of view with fewer layers and a higher shape deviation, as shown in the following formula:

[0106] RF n =RF n-1 ×S n +(K n -1) Formula (5)

[0107] The above formula is the calculation formula for the receptive field, where RF n is the receptive field size of the nth layer, RFn-1 is the receptive field size of the n-1th layer, S n is the step size of the nth layer, Kn is the convolution kernel size of the nth layer;

[0108] Step 2.3, the network structure of the full-dimensional dynamic convolution feature fusion module is as follows:

[0109] like Figure 3As shown, for complex scenarios of transmission lines, the application of traditional static convolution lacks flexibility. It remains unchanged during the network training process and cannot be adjusted according to changes in the feature distribution of different input data, resulting in decreased model adaptability and poor recognition of insulators in complex scenarios. At the same time, the baseline model adopts a cross-stage feature fusion strategy, which may cause feature information to be blurred or ignored, affecting the performance of the model, and it needs to learn the weights and biases of feature fusion, leading to the risk of overfitting. The present invention also proposes a full-dimensional dynamic convolution feature fusion module to learn the linear combination of multiple convolution kernels and the weighting of their input-related attention, thereby improving the accuracy of ice-covered insulator defect detection.

[0110] The full-dimensional dynamic convolution introduces a multi-dimensional attention mechanism through a parallel strategy to learn more flexible attention to the four dimensions of the convolution kernel space. For the convolution kernel, the convolution parameters at the spatial position, the convolution filters of the input channel, the convolution filters of different output channels, and the n overall convolution kernels are all assigned attention weight values; by progressively multiplying the convolution along the dimensions of position, channel, filter, and kernel by different attentions, the convolution operation will have differences in various dimensions for the input, providing better performance to capture rich contextual information.

[0111] A conventional convolutional layer has only one static convolution kernel, which is applicable to all input samples, but for a dynamic convolutional layer, it uses a linear combination of n convolution kernels and an attention mechanism for dynamic weighting, making the convolution operation dependent on the input. The traditional dynamic convolution operation can be defined as:

[0112]

[0113] In the above formula, x is (h,w,c in ) size feature input, x out is (h,w,c out ) size feature output. i represents the i-th output filter kernel, α wi is the attention coefficient. However, traditional dynamic convolution only includes the convolution kernel and the attention weight of the filter kernel. However, the detailed parameters of the convolution kernel still include kernel parameters, input and output channels, which ignores other adjustable parameters in convolution and increases the convolution parameters by n times, which is inefficient.

[0114] The full-dimensional dynamic convolution is implemented as follows:

[0115]

[0116] In the above formula, α wi , α fi , α ci and α siThey are the attention coefficients of the convolution kernel in space, channel, convolution filter and convolution kernel respectively. The four types of attention weights enable the convolution operation to be performed for all spatial positions, all input channels, all filters and all kernels of different inputs, providing performance guarantee for capturing rich contextual clues.

[0117] The full-dimensional feature fusion (Optimized Dynamic Convolution Coarse-to-Fine, ODC2F) is as follows:

[0118] In the baseline model, although the C2F (Coarse-to-Fine) module plays a key role in the cross-stage feature fusion process, it may cause some important feature information to be blurred or ignored during the cross-stage feature fusion process, affecting the performance of the model. In addition, it needs to learn the weights and biases of feature fusion, which will introduce a large number of additional parameters and may lead to the risk of overfitting. The specific formula is as follows:

[0119]

[0120] In the above formula, F in is the input feature, F out is the output feature; when Bottleneck is True, τ is 1, otherwise it is 0; λ is the number of feature fusion modules, and ω is the residual weight bias coefficient. Due to the relatively limited information exchange between feature maps, the module may lead to low target detection accuracy. The ODC2F formula is as follows:

[0121]

[0122] In the above formula, F o ” ut is the output feature; This is the full-dimensional dynamic convolution described above. Through full-dimensional dynamic convolution, a parallel strategy is adopted to learn attention features in four dimensions of the convolution kernel space, which can provide better performance to capture rich contextual information.

[0123] Step 2.4, the network structure of the expectation maximization dynamic convolutional attention mechanism EM-DCA is as follows: Figure 4 As shown, the traditional attention mechanism uses channel dimensionality reduction to model cross-channel relationships. It is difficult for the network to retain information on each channel, and the model performance is poor, making it difficult to achieve the expected effect. Therefore, the present invention uses the expectation maximization attention mechanism to focus on and utilize important information in the input data, helping the model to better allocate attention, focus on important targets or scene parts, and learn more generalized feature representations, which helps to improve the generalization ability and robustness of the model.

[0124] Assume input feature F∈RC×H×W , where C represents the input channel, H and W represent the length and width of the input feature respectively. Along the channel dimension, the feature F is divided into G sub-features, and G << C, then F can be expressed as:

[0125]

[0126] The EM-DCA attention mechanism has four branches in total. Among them, three parallel paths are used to extract the attention weight descriptions of the grouped feature maps. Two 1x1 branches and one 3x3 branch. Among them, two global average poolings are used to encode the channels in two spatial directions in the 1x1 branch, and a full-dimensional dynamic convolution is included in the 3x3 branch to capture multi-scale feature representations.

[0127] Two parallel 1D feature encoding vectors will share a 1x1 convolution, enabling the model to capture local cross-channel interactions. The formula for encoding global information in the horizontal dimension of channel C at height H is:

[0128]

[0129] The formula for the pooling output of channel C at width W is:

[0130]

[0131] The 3x3 branch captures local cross-channel interactions through convolution to expand the feature space.

[0132] It not only encodes the information between channels to adjust the importance of different channels, but also retains the precise spatial structure information in the channels.

[0133] Among them, Cross-spatial learning establishes mutual dependencies between channels and space to achieve richer feature aggregation. Two-dimensional global average pooling is used to encode global spatial information in the output of the 1x1 branch, and the output of the smallest branch is transformed into the corresponding dimension shape The formula is:

[0134]

[0135] Finally, the output of each branch is calculated as the attention weight value, combined with global semantic information, to capture the key regions. This module can help the model capture key information in the input data more effectively and dynamically adjust the attention distribution according to its importance. At the same time, aiming to retain the information on each channel and reduce the computational overhead, part of the channels are reshaped into the batch dimension, and the channel dimension is grouped into multiple sub-features, so that the spatial semantic features are evenly distributed in each feature group.

[0136] Among them, step 3 is specifically implemented according to the following steps:

[0137] Step 3.1, input the training set images divided in step 1.4 into the LDKA-NET network model obtained in step 3 for training, set batch_size to 16, use stochastic gradient descent with momentum of 0.9, train for 400 rounds, and obtain the training results;

[0138] Step 3.2, use the validation set to validate the LDKA-NET network model, obtain the validation results, and analyze the differences between the training results and the validation results, such as whether they are overfitting or underfitting; if overfitting occurs, it means that the model performs very well on the training set, but very poorly on the validation set, that is, the generalization ability of the model is very poor, which needs to be solved by increasing the number of images in the training set, reducing the complexity of the model, and fine-tuning the hyperparameters; if underfitting occurs, it means that the model performs poorly on both the training set and the validation set, which needs to be solved by increasing the complexity of the model, fine-tuning the hyperparameters, and adding more target features;

[0139] Step 3.3, by adjusting the hyperparameters of the LDKA-NET network model on the validation set, fine-tuning the learning rate to 0.003 to obtain better robustness, setting batch_size to 16, and obtaining the optimized LDKA-NET network model. Among them, hyperparameters are parameters that need to be manually set before model training, such as learning rate and batch size.

[0140] Step 4 is specifically implemented according to the following steps:

[0141] Step 4.1, input the test set into the step 3.3 to obtain the optimized LDKA-NET model. The specific optimization process is as follows: set the hyperparameters of the LDKA-NET network model: input image pixel size 640*640, freeze training for 150 rounds, batch_size32, unfreeze training for 250 rounds, batch_size 4, num_workers2, Adam optimizer, decay weight coefficient 5*10-4 initial learning rate 1*10-5, set the IoU threshold to 0.5 for testing when testing the test set, and fine-tune the learning rate to 0.003 during verification to obtain better robustness, set batch_size to 32, freeze training for the first 150 rounds, the loss decreases quickly, unfreeze training for the next 250 rounds, and continuously fine-tune the network. After 350 rounds out of a total of 400 rounds, the loss change of the verification set gradually decreases, and the optimized LDKA-NET model is obtained;

[0142] Step 4.2: Use the test set to test the performance of the optimized LDKA-NET model and obtain the test results.

Claims

1. The insulator ice detection method based on the convolutional network LDKA-NET is characterized by: Follow the steps below to implement it: Step 1: Collect the image data of ice-covered insulator defects on the transmission line, annotate and divide the data set into training set, validation set and test set; Step 2: Establish the convolutional network LDKA-NET network model; Step 3: Input the images in the training set into the convolutional network LDKA-NET network model for training, and obtain the optimized LDKA-NET network model through training; Step 4: Input the images in the test set into the optimized LDKA-NET network model for detection to obtain the iced insulator defect detection result map.

2. The insulator ice recognition and detection method based on the convolutional network LDKA-NET according to claim 1 is characterized in that: The step 1 is specifically implemented according to the following steps: Step 1.1: Use the high-definition camera on the drone to shoot a large number of inspection videos, including images of ice-covered insulators of transmission lines under various complex backgrounds, as well as images of distant views, close-ups, and close-ups, and keep the number of images of each type of ice-covered insulator consistent; Step 1.2: rotating, flipping, scaling, and arbitrarily cropping the ice-covered insulator image obtained in step 1.1, thereby obtaining an expanded ice-covered insulator image; Step 1.3: The sum of the ice-covered insulator image obtained in step 1.1 and the expanded ice-covered insulator image obtained in step 1.2 is used as the ice-covered insulator sample image library. A corresponding label file is created for each image sample in the ice-covered insulator sample image library. The label file meets the xml label file standard in Pascal VOC format. The xml label file includes image ID, image path, image name, pixel height and width of the image, where the pixel height and width of the image are represented by four coordinates of a rectangular box, which are x min ,y min ,x max ,y max , where (x min ,y min ) is the coordinate of the upper left vertex of the rectangle, (x max ,y max ) is the coordinate of the lower right vertex of the rectangular box; Step 1.4: Divide the ice-covered insulator sample image library obtained in step 1.3 into a training set, a validation set, and a test set in a ratio of 8:1:

1.

3. The insulator ice detection method based on convolutional network LDKA-NET according to claim 1 is characterized in that: The step 2 is specifically implemented according to the following steps: Step 2.1: Establish the convolutional network LDKA-NET network model, including image preprocessing, backbone network (Backbone), feature extraction network (Neck) and prediction network (Head); Step 2.2: The wide-field convolution backbone network structure is as follows The wide-field convolution backbone network is mainly composed of four stages, each of which contains the initial layer (stem), stage layer (Stage) and transition layer (Transition) convolution layer; the Depth Wise Conv layer is a depth-separable convolution, which consists of two parts, namely Depth Wise (DW) convolution and Point Wise (PW) convolution; each convolution kernel of the DW convolution only calculates the corresponding feature map, but it may weaken the information interaction between feature pixels in different layers, which may lead to accuracy loss; the PW convolution is a 1×1 convolution with an output of n channels, which aims to compensate for the lack of feature interaction in the DW convolution. Through point-to-point multiplication and addition operations, the PW convolution can combine the information of different feature maps to enhance the interaction between features; Assume that the input feature map size is D w ×D w ×C, the convolution kernel size is D k ×D k ×C, the output feature map size is D w ×D w ×C out , the number of parameters of the standard convolutional layer is as follows: P Standard =(D k ×D k ×C)·C out Formula (1) Among them, C is the number of input feature channels, C out Indicates the number of output feature channels and convolution kernels; In addition, the depthwise convolution in the depthwise separable convolution is responsible for filtering, with a size of D k ×D k ×1, a total of C, acting on each channel of the input; point-by-point convolution is responsible for converting channels, size 1×1×C, a total of C out , acts on the output feature map of the deep convolution; The depth-separable convolution parameter is as follows: The ratio of the number of depth convolution parameters to the number of standard convolution parameters is: Based on this, the use of depthwise separable convolution can solve the problem of increased parameters and floating-point operations caused by the use of large-size convolution kernels. The stem layer consists of 3×3 convolution with a stride of 2, depthwise separable convolution, and DW convolution with a stride of 2, which are used to increase the dimension and reduce the size of the input image; as shown in the following formula: As shown in the formula, w is the feature weight, D represents the depth separable volume; assuming the input size is H and W, where H is the image height and W is the image width, the output feature size of the stem layer is H / 4×W / 4×C1, where C1 is the number of channels after dimensionality increase; The Stage layer is composed of a stack of RepLK Block and ConvFFN. The RepLK Block contains a normalization layer, a 1×1 convolution, a depth-separable convolution, and an important residual connection. ConvFFN uses a 1×1 convolution to replace the fully connected layer and uses residual connections between layers. The Transition layer mainly uses PW convolution and DW convolution with a stride of 2 for image downsampling. In addition, the use of a large core backbone network can obtain a larger receptive field of view with fewer layers and a higher shape deviation, as shown in the following formula: RF n = RF n-1 × S n + (K n - 1) Equation (5) The above formula is the calculation formula for the receptive field, where RF n is the receptive field size of the nth layer, RF n-1 is the receptive field size of the n-1th layer, S n is the step size of the nth layer, K n is the convolution kernel size of the nth layer; Step 2.3: The network structure of the full-dimensional dynamic convolution feature fusion module is as follows: The full-dimensional dynamic convolution is implemented as follows: In the above formula, α wi , α fi , α ci and α si They are the attention coefficients of the convolution kernel in space, channel, convolution filter and convolution kernel respectively; The full-dimensional feature fusion (ODC2F) is as follows: The specific formula is as follows: In the above formula, F in is the input feature, F out is the output feature; when Bottleneck is True, τ is 1, otherwise it is 0; λ is the number of feature fusion modules, and ω is the residual weight bias coefficient; due to the relatively limited information exchange between feature maps, the module may lead to low target detection accuracy, and the ODC2F formula is as follows: In the above formula, F o ” ut is the output feature; This is the full-dimensional dynamic convolution described above; Step 2.4: The network structure of the EM-DCA expectation maximization dynamic convolutional attention mechanism is as follows: Let the input feature \(F\in\mathbb{R}\) C×H×W , where \(C\) represents the number of input channels, \(H\) and \(W\) represent the height and width of the input feature respectively. The feature \(F\) is divided into \(G\) sub - features along the channel dimension, and \(G\ll C\). Then \(F\) can be expressed as: F=[F0,F i ,…,F G-1 ],F i ∈R C / / G×H×W formula (12) The EM-DCA attention mechanism has four branches, three parallel paths extracting the attention weight description of the grouped feature map; two 1x1 branches and one 3x3 branch; two global average poolings are used to encode channels along two spatial directions in the 1x1 branch, and a full-dimensional dynamic convolution is included in the 3x3 branch to capture multi-scale feature representations; Two parallel 1D feature encoding vectors will share a 1x1 convolution, enabling the model to capture local cross-channel interactions. The formula for encoding global information along the horizontal dimension of channel C at height H is: The formula for the pooled output along channel C at width W is: Use 2D global average pooling to encode global spatial information in the output of the 1x1 branch and transform the output of the smallest branch into the corresponding dimensional shape R1 1×G ×R3 C / / G×HW , the formula is: Finally, the output of each branch is calculated as an attention weight value, which combines the global semantic information to capture the key areas.

4. The ice-covered insulator defect identification method based on the convolutional network LDKA-NET according to claim 1 is characterized in that: The step 3 is specifically implemented according to the following steps: Step 3.1: Input the training set images divided in step 1.4 into the LDKA-NET network model obtained in step 3 for training, set batch_size to 16, use stochastic gradient descent with momentum of 0.9, train for 400 rounds, and obtain the training results; Step 3.2: Use the validation set to validate the LDKA-NET network model, obtain the validation results, and analyze the differences between the training results and the validation results, such as whether they are overfitting or underfitting; if overfitting occurs, it means that the model performs very well on the training set, but very poorly on the validation set, that is, the generalization ability of the model is very poor, which needs to be solved by increasing the number of images in the training set, reducing the complexity of the model, and fine-tuning the hyperparameters; if underfitting occurs, it means that the model performs poorly on both the training set and the validation set, which needs to be solved by increasing the complexity of the model, fine-tuning the hyperparameters, and adding more target features; Step 3.3: By adjusting the hyperparameters of the LDKA-NET network model on the validation set, fine-tuning the learning rate to 0.003 for better robustness, setting the batch_size to 16, the optimized LDKA-NET network model is obtained; among them, the hyperparameters are the parameters that need to be manually set before model training.

5. According to the convolutional network LDKA-NET ice-covered insulator defect identification method of claim 1, the step 4 is specifically implemented according to the following steps: Step 4.1: Input the test set into step 3.3 to obtain the optimized LDKA-NET model. The specific optimization process is as follows: set the hyperparameters of the LDKA-NET network model: input image pixel size 640*640, freeze training for 150 rounds, batch_size 32, unfreeze training for 250 rounds, batch_size 4, num_workers 2, Adam optimizer, decay weight coefficient 5*10-4, initial learning rate 1*10-5, set the IoU threshold to 0.5 for testing when testing the test set, and fine-tune the learning rate to 0.003 during verification to obtain better robustness, set batch_size to 32, freeze training for the first 150 rounds, the loss decreases quickly, unfreeze training for the next 250 rounds, and continuously fine-tune the network. After 350 rounds out of a total of 400 rounds, the loss change of the verification set gradually decreases, and the optimized LDKA-NET model is obtained. Step 4.2: Use the test set to test the performance of the optimized LDKA-NET model and obtain the detection results.

Images