Construction method of power transmission tower pole electric power fitting damage detection network enhanced through high and low frequency cooperative guidance based on single-stage detection

Abstract

The invention belongs to the field of image target detection. A construction method of a power transmission tower pole electric power fitting damage detection network enhanced by high and low frequency cooperative guidance based on single-stage detection is characterized by comprising the following steps: step S1, constructing a single-stage detection framework for target detection of power transmission tower pole electric power fitting damage; s2, a plurality of high and low frequency collaborative guidance enhancement modules WCGE are added into the detection framework, and the discrimination capability of the model is improved through effective interaction of high and low frequency components; and S3, inputting a power transmission tower pole electric power fitting damage data set into the detection network, optimizing the model through a training process of loss constraint, and finally outputting a defect detection result. According to the method, the detection performance is remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of image target detection, and particularly relates to a construction method of a high-low frequency cooperative guidance enhanced power fitting damage detection network for a power transmission tower based on a single-stage detection. BACKGROUND

[0002] Power fitting damage detection for a power transmission tower is an important safety monitoring problem for power equipment, especially in the inspection and maintenance of power transmission lines. The existence of power fitting damage not only threatens the operation of power facilities, but in some cases can also cause power failures or safety accidents such as fires. Therefore, how to quickly and accurately detect power fitting damage on the tower of a power transmission line has become a key technical requirement in the power industry.

[0003] Traditional power transmission tower inspection usually relies on manual inspection or visual image acquisition by unmanned aerial vehicles and other equipment, and then relies on manual image analysis. Manual inspection and manual image analysis not only consume time and effort, but are also easily affected by external factors such as weather and light, making it difficult to ensure the comprehensiveness and accuracy of the detection.

[0004] In recent years, with the rapid development of computer vision and deep learning technology, image analysis-based power fitting damage detection methods have gradually become a hot topic in the research field. Through automated analysis of power transmission line tower images, efficient power fitting damage detection can be achieved, thereby providing timely safety hazard warnings for power operation and maintenance departments and reducing the risk of potential accidents. Existing target detection methods have certain applications in this field. For example, Zhang et al. proposed an intelligent detection method based on a refined mask region convolutional neural network, which effectively realized the automatic identification and analysis of power fitting damage on a power transmission tower. Qiu et al. combined the image dataset of power fitting damage scenarios and harmful bird species for data enhancement, and used deep learning technology to realize bird target classification and positioning in power transmission line inspection images. Although these methods have achieved significant results in power fitting damage detection and provide relatively accurate detection results, their processing speed and computational efficiency still face great challenges. Therefore, single-stage target detection algorithms have gradually become the focus of research, especially the YOLO series. Miao et al. implemented accurate identification of power fitting damage on a power transmission line based on a multi-scale training YOLOv3 target detection network. Hao et al. further improved the accuracy of power transmission line fault detection by introducing an attention module into the YOLOv3 feature extraction network. However, although existing target detection methods have made certain progress in power transmission tower inspection, there are still some limitations when applying single-stage target detection frameworks.

[0005] The existing power hardware damage detection algorithm of the transmission tower mainly relies on the spatial domain feature extraction method. However, due to the complex structure of the power hardware damage, the similar color and background and the fuzzy edge, the traditional convolutional neural network (CNN) has limitations in capturing high-frequency texture features. Especially for the power hardware damage with subtle texture difference, the traditional convolution operation is difficult to effectively extract the rich texture details, resulting in insufficient feature expression. This insufficient feature extraction capability directly affects the detection accuracy of the model in a complex background, especially when the power hardware damage is similar to the surrounding environment, which is prone to false detection or missed detection.

[0006] In order to solve the limitations of the backbone network in the single-stage detection framework when using the single-stage detection framework for power hardware damage detection of transmission tower, the existing method mainly relies on spatial domain feature extraction, and ignores the effective use of other domain features such as frequency domain, resulting in limited detection accuracy when dealing with complex background and subtle targets. Traditional convolution has certain limitations in capturing high-frequency texture details, especially in the detection of power hardware damage with complex structure, similar color and background and fuzzy edge, which is prone to false detection or missed detection, thereby affecting the accuracy and reliability of the detection result. In order to solve this problem, the present application proposes a construction method of high-low frequency collaborative guidance enhanced power hardware damage detection of transmission tower based on single-stage detection. SUMMARY

[0007] The purpose of the present application is to provide a construction method of high-low frequency collaborative guidance enhanced power hardware damage detection network of transmission tower based on single-stage detection, which significantly improves the detection performance.

[0008] To achieve the above purpose, the technical scheme adopted by the present application is a construction method of high-low frequency collaborative guidance enhanced power hardware damage detection network of transmission tower based on single-stage detection, characterized by comprising the following steps: Step S1: constructing a single-stage detection framework for target detection of power hardware damage of transmission tower; Step S2: adding a plurality of high-low frequency collaborative guidance enhancement modules WCGE (high-low frequency collaborative guidance enhancement module) to the detection framework, and improving the discrimination ability of the model through effective interaction of high-low frequency components; Step S3: inputting the power hardware damage dataset of transmission tower into the detection network, and optimizing the model through the loss constraint training process, and finally outputting the defect detection result.

[0009] Further, in step S1, a single-stage detection framework is constructed, which includes a backbone network, a neck network and three detection heads: Backbone network: The first layer standard convolution Conv (quasi-convolution) module: the convolution kernel size is 3*3, and the step is 2; the second layer standard convolution Conv module: the convolution kernel size is 3*3, and the step is 2; the third layer feature extraction module: the high-low frequency cooperative guidance enhancement module WCGE1; the fourth layer standard convolution Conv module: the convolution kernel size is 3*3, and the step is 2; the fifth layer feature extraction module: the high-low frequency cooperative guidance enhancement module WCGE2; the sixth layer standard convolution Conv module, the convolution kernel size is 3*3, and the step is 2; the seventh layer feature extraction module: the high-low frequency cooperative guidance enhancement module WCGE3; the eighth layer standard convolution Conv module: the convolution kernel size is 3*3, and the step is 2; the ninth layer spatial pyramid pooling layer: the SPPF module; Neck network: The tenth layer upsampling Upsample (upsampling) module; the eleventh layer concatenation concat operation; the twelfth layer C2F (cross-stage partial fusion) module; the thirteenth layer upsampling Upsample (upsampling) module; the fourteenth layer concatenation concat operation; the fifteenth layer C2F module; the sixteenth layer standard convolution Conv module, the convolution kernel size is 3*3, and the step is 2; the seventeenth layer concatenation concat operation; the eighteenth layer C2F module; the nineteenth layer standard convolution Conv module, the convolution kernel size is 3*3, and the step is 2; the twentieth layer concatenation concat operation; the twenty-first layer C2F module; Detection head: The twenty-second layer detection head 1; the twenty-third layer detection head 2; the twenty-fourth layer detection head 3.

[0010] Further, the specific structure of the high-low frequency cooperative guidance enhancement module WCGE (high-low frequency cooperative guidance enhancement module) in step S2 is as follows: Wavelet transform module DWT→low frequency component , three high frequency components ; Indicate the high frequency subband of the horizontal, vertical and diagonal line respectively; Low frequency component → feature extraction of low frequency component → output new low frequency component ( Indicate the new low frequency component)→self-attention→concatenate three high frequency components →3*3 standard convolution Conv→sigmoid (activation function) activation function to obtain three weight values→three weight values and low frequency component Multiplication output three weighted low frequency components→add with original high frequency components → update Addition→output a comprehensive high frequency component → Feature extraction of high-frequency components → Concatenation of enhanced low-frequency components → 3*3 standard convolution Conv → Output features .

[0011] Further, in step S3, the training process in the model optimized by the loss constraint adopts the following steps: Step 1) Input the transmission tower pole power fitting damage image into the backbone network of the single-stage detection framework to extract information-rich features; Step 2) In the neck network part of the framework, fuse features from different scales and send these fused features into three detection heads of different scales to generate the final detection results; Step 3) During network training, three loss functions are applied for constraint to optimize the performance of the network and promote detection accuracy and stability.

[0012] Further, in step 3), the three loss constraints are as follows: (1) Binary Cross Entropy Loss (BCE) is used to measure the difference between the probability predicted by the model and the actual label. By minimizing the BCE loss, the classification ability of the model can be optimized. Its calculation formula is: Where: represents the binary cross entropy loss value; represents the contribution of the logarithmic term to the predicted probability when the real label is 1; is the real label, usually 0 or 1; represents the probability predicted by the model, i.e., the predicted probability that the sample belongs to label 1; (2) Boundary box regression loss (CIoU) is used to measure the similarity between the predicted box and the real box, comprehensively reflecting their overlap. Its calculation formula is: Where: represents the boundary box regression loss value; IoU represents the intersection over union between the predicted boundary box b and the real boundary box ; d represents the Euclidean distance between the center points of the predicted boundary box b and the real boundary box ; c is the length of the diagonal of the smallest closed region enclosing the two boxes; v is a term used to measure the consistency of the aspect ratio; is a weight parameter, specifically , used to balance the importance of the ratio term and other parts; represents the probability predicted by the model.

[0013] ​(3) Distribution Focus Loss (DFL) is used to solve regression problems of predicting continuous variables. The specific calculation formula is as follows: in: Represents the true value Falling on adjacent discrete points and The distribution focus loss that occurs between these points; This represents the total distributed focal loss, which is typically an aggregation of the losses of all relevant discrete point pairs. It is a true, continuous target value; and In the discretized space, compared with the true value The two closest adjacent discrete points; and The model predicts the true value. Falling point and The probability or score; The overall network loss function is expressed as: Where: L represents the total loss; , , The hyperparameter that balances different losses ( , , Specifically, they are 7.5, 0.5, and 1.5 respectively.

[0014] Compared with the prior art, the present invention has the following beneficial effects: This invention optimizes the single-stage detection framework, significantly improving detection performance (accuracy and reliability of detection results), and achieves excellent results on a power transmission tower fitting damage dataset. Comparative experiments verify the superior performance of the algorithm.

[0015] The high-low frequency collaborative guidance enhancement module proposed in this invention enhances the ability to capture high-frequency detailed information by guiding low-frequency global information, thereby improving the detection performance of the model. Attached Figure Description

[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is an overall network structure diagram of an embodiment of the present invention.

[0017] Figure 2 Schematic diagram of the high-low frequency collaborative guidance enhancement module WCGE. Detailed Implementation

[0018] As Figure 1 and Figure 2 shown, a method for constructing a power hardware fitting damage data set of a high-low frequency synergistic guiding enhanced transmission tower pole, comprising the following steps: S1, a single-stage detection framework is constructed for target detection of power hardware fitting damage of a transmission tower pole; S2, a plurality of high-low frequency synergistic guiding enhancement modules WCGE are added to the detection framework, and the discriminant ability of the model is improved through effective interaction of high-low frequency components; S3, the power hardware fitting damage data set of the transmission tower pole is input into the detection network, and the model is optimized through a loss constrained training process, and finally the defect detection result is output Step S1 specifically includes: As Figure 1 shown, a single-stage detection framework is constructed, which includes a backbone network, a neck network and a detection head. The present application refers to the structure of YOLOv8 (You Only Look Once, version 8) and modifies the backbone network. In the backbone network, 32 times down-sampling is first performed through 5 times convolution, and the feature maps of the second, third and fourth times down-sampling are sent into the WCGE module to extract features, and the outputs are , , , , , respectively represent the feature maps output by the third, fifth and seventh layer WCGE (high-low frequency synergistic guiding enhancement) modules on the backbone network; the feature map of the last down-sampling is sent into the SPPF (spatial pyramid pooling fast module), and the output is representing the feature map output by the eighth layer SPPF (spatial pyramid pooling fast module) on the backbone network, the specific formula is as follows: Among them, , , respectively represent the feature maps of the second, third and fourth times down-sampling in the backbone network.

[0019] Step S2 specifically includes: As Figure 2 shown, the WCGE module first applies discrete wavelet transform convolution (DWT) to the feature map for frequency domain decomposition, which is decomposed into a low-frequency component and three high-frequency components . Subsequently, the low-frequency component After processing by the feature extraction module, the feature output is obtained and the self-attention mechanism is used to generate enhanced low-frequency features to focus on the global information of the image. The specific formula is as follows: The three high-frequency components are spliced with the low-frequency features and feature fusion is performed through 3*3 convolution. Then, sigmoid function is used to generate three weight values, the purpose is to adaptively adjust the contribution of each high-frequency component according to the global information of the low-frequency features. Multiply the three weight values with the corresponding high-frequency components respectively, and add them with the original high-frequency components to ensure that each high-frequency component is properly enhanced. Finally, add the updated three high-frequency components to obtain a comprehensive high-frequency component , which aims to combine information of different frequencies to improve the model's ability to capture and express image details. The specific formula is as follows: wherein, respectively represent the attention weights generated by the Sigmoid function for respectively adaptive enhancement of the high-frequency components HL, LH, HH; respectively represent the enhanced horizontal, vertical, and diagonal high-frequency features; Next, feature extraction is performed on the high-frequency components to extract useful information from high-frequency details and enhance the perception of image details and local features. Then, the low-frequency features are spliced with the high-frequency components after feature extraction, and 3*3 convolution is used to fuse the global information of the low-frequency and the detailed information of the high-frequency to obtain the output features . This process helps to improve the model's ability to capture high-frequency details, thereby effectively improving detection accuracy. The specific formula is as follows: wherein, represents the features extracted by the feature extraction module; represents the output image after enhancement by the final WCGE (High-Low Frequency Cooperative Guidance Enhancement) module; represents 3*3 convolution; This indicates the feature extraction module.

[0020] Step S4 specifically includes: The power transmission tower fitting damage dataset is input into the aforementioned power transmission tower fitting damage detection model. A loss constraint training process is then employed, and finally, the defect detection results are output. This invention utilizes three loss constraint network training processes, as detailed below: The first type, binary cross-entropy loss (BCE), quantifies the difference between the model's predicted probabilities and the actual labels. Classification performance is optimized by minimizing the BCE loss, and its specific formula is as follows: The second type of bounding box regression loss (CIoU) is used to comprehensively reflect the similarity between the predicted box and the ground truth box. The specific formula is as follows: Where IoU is the prebound bounding box b and the ground truth bounding box b. The crossover ratio between them; ) represents the predicted bounding box b and the ground truth bounding box b. The Euclidean distance between the center points; c is the diagonal length of the smallest closed region containing the two boxes; 𝑣 measures the consistency of the aspect ratio; It is a weighting parameter used to balance the importance of the proportional term and other parts.

[0021] The third type of distribution focus loss (DFL) is used to handle regression problems that predict continuous variables. The specific formula is as follows: in, It is a true, continuous target value; and Is it the true value? The two closest adjacent discrete points; and The model predicts the true value. Falling point and The probability or score.

[0022] The overall network loss function is expressed as: in , , It is a hyperparameter that balances different losses. =7.5, =0.5, =1.5 Example 1 Parameter settings The experiment is carried out on the power hardware damage dataset of transmission tower poles, using NVIDIA RTX3090 GPU (NVIDIA RTX 3090 graphics card) and the PyTorch framework. In the experiment, the power hardware damage dataset of transmission tower poles has a total of 3281 images, 2297 images for the training set and 984 images for the test set. The input image size is set to 640*640, the Adam optimizer is used to optimize the proposed network, the momentum attenuation index is β1=0.9, β2=0.999, the learning rate is set to 0.01, and the batch size is set to 16 to adapt to the hardware memory. In order to better evaluate the method proposed in the application, the precision P, recall R, MAP50 and MAP50-95 are used as evaluation indexes. The proposed transmission tower pole power hardware damage detection network is compared with four advanced algorithms, namely Yolov5n, Yolov8n, Yolov10n and RT-Detr.

[0023] 2. Experimental results Table 1 Comparison results of the proposed method and the comparative methods on the power hardware damage dataset of transmission tower poles On the power hardware damage dataset of transmission tower poles, the method of the application and the other four algorithms are quantitatively evaluated, and Table 1 shows the average P, R, MAP50 and MAP50-95 values of various methods. It can be seen that the application of the application is higher than other methods in various numerical values on the power hardware damage dataset of transmission tower poles.

[0024] Therefore, the transmission tower pole power hardware damage network of the application has good detection performance in power hardware damage detection. Through the network of the application, the power hardware damage on the transmission tower pole can be effectively identified.

Claims

1. A method for constructing a power hardware damage detection network based on single-stage detection and high-low frequency collaborative guidance enhancement, characterized in that Comprise the following steps: Step S1: a single stage detection framework is constructed for the target detection of power fittings damage of transmission tower poles; Step S2: a plurality of high and low frequency cooperative guidance enhancement modules WCGE are added to the detection framework, and the discriminant ability of the model is improved through the effective interaction of high and low frequency components; Step S3: the power fittings damage dataset of transmission tower poles is input into the detection network, and the model is optimized through the loss constraint training process, and finally the defect detection result is output.

2. The method according to claim 1, wherein the method is characterized in that In step S1, a single stage detection framework is constructed, which includes a backbone network, a neck network and three detection heads: The backbone network: The first layer standard convolution Conv module: the convolution kernel size is 3*3, and the step is 2; The second layer standard convolution Conv module: the convolution kernel size is 3*3, and the step is 2; The third layer feature extraction module: high and low frequency cooperative guidance enhancement module WCGE1; The fourth layer standard convolution Conv module: the convolution kernel size is 3*3, and the step is 2; The fifth layer feature extraction module: high and low frequency cooperative guidance enhancement module WCGE2; The sixth layer standard convolution Conv module, the convolution kernel size is 3*3, and the step is 2; The seventh layer feature extraction module: high and low frequency cooperative guidance enhancement module WCGE3; The eighth layer standard convolution Conv module: the convolution kernel size is 3*3, and the step is 2; The ninth layer spatial pyramid pooling layer: SPPF module; The neck network: The tenth layer up sampling Upsample module; The eleventh layer splicing concat operation; The twelfth layer C2F module; The thirteenth layer up sampling Upsample module; The fourteenth layer splicing concat operation; The fifteenth layer C2F module; The sixteenth layer standard convolution Conv module, the convolution kernel size is 3*3, and the step is 2; The seventeenth layer splicing concat operation; The eighteenth layer C2F module; The nineteenth layer standard convolution Conv module, the convolution kernel size is 3*3, and the step is 2; The twentieth layer splicing concat operation; The twenty-first layer C2F module; The detection head: The twenty-second layer detection head; The twenty-third layer detection head; The twenty-fourth layer detection head.

3. The method according to claim 1, wherein the method is characterized in that, The specific structure of the high and low frequency cooperative guidance enhancement module WCGE in step S2 is as follows: DWT→ low frequency component , three high frequency components ; representing horizontal, vertical, diagonal high frequency subbands, respectively low frequency component feature extraction of low frequency component → output new low frequency component self-attention → concatenate three high frequency components 3*3 standard convolution Conv → sigmoid activation function to obtain three weight values → three weight values and low frequency component multiply to output three weighted low frequency components → with original high frequency component add → updated add → output an integrated high frequency component feature extraction of high frequency component → concatenate enhanced low frequency component → 3*3 standard convolution Conv → output features .

4. The method according to claim 1, wherein the method is characterized in that In step S3, the training process of optimizing the model through loss constraint adopts the following steps: Step 1) input the power fittings damage image of transmission tower poles into the backbone network of the single stage detection framework to extract information rich features; Step 2) in the neck network part of the framework, fuse the features from different scales, and send these fused features into three detection heads of different scales to generate the final detection result; Step 3) in the network training process, three loss functions are applied for constraint to optimize the performance of the network and promote the detection accuracy and stability.

5. The method according to claim 4, wherein the method is characterized in that In step 3), the three loss constraints are as follows: (1) binary cross entropy loss (BCE) is used to measure the difference between the probability predicted by the model and the actual label; by minimizing the BCE loss, the classification ability of the model can be optimized; its calculation formula is: wherein: represents a binary cross-entropy loss value; represents the contribution to the log term of the predicted probability when the true label is 1 ; is the true label, typically 0 or 1 ; represents the probability predicted by the model, i.e. the predicted probability that the sample belongs to label 1 ; (2) The bounding box regression loss is used to measure the similarity between the predicted box and the real box, and comprehensively reflects their overlap. Its calculation formula is: wherein: represents the bounding box regression loss value; IoU represents the intersection over union between the pre-boundary box b and the real boundary box ; represents the Euclidean distance between the predicted boundary box b and the real boundary box center point; c is the diagonal length of the smallest closed area enclosing the two boxes; v is an item used to measure the consistency of the width-height ratio; is a weight parameter, specifically , used to balance the importance of the ratio item and other parts; represents the probability predicted by the model; (3) The distribution focus loss is used to solve the regression problem of predicting continuous variables, and the specific calculation formula is: where: represents the true value falls between adjacent discrete points and results in a distribution focus loss; represents the total distribution focus loss, typically an integration of the loss over all relevant pairs of discrete points; is the true continuous target value; and are the two adjacent discrete points in the discretized space that are closest to the true value ; and are the probabilities or scores of the model predicting the true value falls in discrete points and ; The overall network total loss function is represented as: wherein: L represents the total loss; , , is a hyperparameter balancing between different losses, , , are respectively 7.5, 0.5, 1.5.

Images