Model training method and device and power transmission line multi-target detection method and device
By improving the YOLOv11n neural network model, replacing the C3K2 module with the CES module, and introducing FFDPN and TADDH, the problem of missed detection caused by visual fatigue in drone inspections was solved, and efficient detection of transmission line faults was achieved.
Patent Information
- Application Number
- CN202510991776.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-18
- Publication Date
- 2025-10-17
AI Technical Summary
In the existing technology, when drones are used to inspect power transmission lines, workers are prone to visual fatigue due to long periods of concentration, which may lead to missed inspections.
Improve the YOLOv11n neural network model, replace the C3K2 module with the CES module, introduce the FFDPN structure and TADDH, and enhance multi-scale feature processing and target detection capabilities.
The accuracy and efficiency of transmission line fault detection are improved, missed detection is avoided, and the overall performance of inspection is improved.
Smart Images
Figure CN120808073A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the technical field of power transmission line detection, and particularly relates to a model training method and device and a multi-target detection method and device for a power transmission line. BACKGROUND
[0002] A power grid is an efficient and fast energy transmission channel and an optimized configuration platform, is a key link for sustainable development of energy and power, and plays an important pivotal role in a modern energy supply system and is related to national energy security. As the most basic component of a power grid, the integrity of a line component is a guarantee for stable operation of the power grid, and therefore a power transmission line needs to be regularly inspected.
[0003] Although the prior art has greatly improved the inspection efficiency by using a UAV for line inspection, the entire inspection process still needs a worker to focus on finding and distinguishing relevant components, and such long-time focus is prone to cause visual fatigue, thereby leading to missed inspection. SUMMARY
[0004] Based on the above problems in the prior art, the purpose of the embodiments of the present application is to provide a model training method and device and a multi-target detection method and device for a power transmission line, by improving a neural network model of YOLOv11n, replacing a C3K2 module thereof with a self-developed CES module, designing an FFDPN structure instead of an FPN module of an original model, and replacing a detection head of the original model with a TADDH, more effective integration and processing of multi-scale features are realized, the extraction of small target edge information and the selection ability of features are enhanced, the overall performance of target detection and classification is improved, and the ability of the detection head to accurately locate and classify targets is improved.
[0005] The technical scheme adopted by the present application to solve its technical problems is: In a first aspect, the present application provides a model training method. The model training method comprises: obtaining a plurality of component defect images of a power transmission line, pre-processing the component defect images to obtain a training data set, a verification data set and a test data set; replacing a C3K2 module in an original YOLOv11n model with a CES module, introducing an FFDPN network structure designed based on an FF module into the original YOLOv11n model to replace an original FPN, and introducing a TADDH to constitute an ipre-YOLOv11n model; inputting the training data set into the ipre-YOLOv11n model for model training, improving the ipre-YOLOv11n model training parameters based on a training result and in combination with a loss function, and iteratively training to obtain a power transmission line multi-target detection model taking the power transmission line image as input and taking a target fault defect image as output; inputting the verification data set into the power transmission line multi-target detection model for model evaluation; inputting the test data set into the power transmission line multi-target detection model for model testing to obtain the target fault defect; The power transmission line image is used to represent a target detection power transmission line, and the target fault defect image is used to represent an image of a component defect of a fault-occurring part of the target power transmission line.
[0006] Preferably, the step of pre-processing the multi-target image further comprises: segmenting the component defect image into a plurality of image feature blocks, defect labeling the image feature blocks, and establishing an initial data set; performing data enhancement on the images in the initial data set to obtain a target data set; dividing the target data set into the training data set, the verification data set, and the test data set according to a preset ratio.
[0007] Preferably, the step of replacing the C3K2 module with the CES module in the original YOLOv11n model comprises: designing a brand-new edge information extraction module, replacing the Bootleneck module in the C3K2 module with the edge information extraction module to form the CES module; The CES module specifically works as follows: adjusting the input feature map to a corresponding size through four adaptive average pooling layers with different output scales, and extracting four features with different scales through large convolution; adjusting the channel number of the four features with different scales through small kernel convolution, and restoring four shallow semantic information with different scales through an upsampling module, further extracting four edge information with different scales from the shallow semantic information through the edge information extraction module; strengthening the output of the edge information, and fusing and interacting the strengthened edge information and the original feature information.
[0008] Preferably, the step of introducing the FFDPN network structure based on the FF module design to replace the original FPN in the original YOLOv11n model comprises: the FF module extracts features from the input feature map through multi-scale feature fusion and deep convolution operation; a forward focusing diffusion pyramid network is constructed based on the FF module to realize multi-scale feature extraction and fusion, and the specific working process comprises: three feature maps of different scales are processed through upsampling, downsampling and convolution operation; parallel deep convolution is used to further capture multi-scale information to generate comprehensive feature maps.
[0009] Preferably, the step of introducing the TADDH into the original YOLOv11n model comprises the working process of the TADDH: two shared group normalization convolution blocks are used to first extract feature maps of different dimensions; a function extractor is used to learn task interaction functions from multiple convolution layers, and the concatenated interaction functions are used for classification and localization tasks; the DCNV2 is generated by using the interactive function of the positioning branch to generate the offset and mask of the DCNV2, the features are selected by using the DCNV2, and the features are expanded by using the scaling layer.
[0010] Preferably, the step of inputting the verification data set into the power line multi-target detection model for model evaluation and outputting multiple evaluation indexes comprises: the verification data set is input into the power line multi-target detection model, and the power line multi-target detection model is evaluated based on the precision P, recall R and average precision mean mAP; the index formula for evaluating the model performance is as follows:
[0011]
[0012]
[0013] wherein, TP is the number of correctly predicted positive samples, FN is the number of incorrectly predicted negative samples, and FP is the number of incorrectly predicted positive samples.
[0014] In a second aspect, the present application provides a power line multi-target detection method. The multi-target detection method comprises: acquiring a power line image to be detected; inputting the power line image into a power line multi-target detection model to obtain an image of a component defect of the target power line; The power transmission line multi-target detection model is obtained through the model training method of any one of the first aspect.
[0015] In a third aspect, the present application provides a model training device. The model training device comprises: a data acquisition module configured to acquire a plurality of component defect images of a power transmission line, and to pre-process the component defect images to obtain a training data set, a verification data set, and a test data set; a model improvement module configured to replace a C3K2 module with a CES module in an original YOLOv11n model, and to introduce an FFDPN network structure based on an FF module design into the original YOLOv11n model to replace an original FPN, and to introduce a TADDH to form an ipre-YOLOv11n model; a model training module configured to input the training data set into the ipre-YOLOv11n model for model training, to improve the ipre-YOLOv11n model training parameters based on a training result and in combination with a loss function, and to iteratively train to obtain a power transmission line multi-target detection model that takes the power transmission line image as input and takes a target fault defect image as output; a model evaluation module configured to input the verification data set into the power transmission line multi-target detection model for model evaluation; a detection result generation module configured to input the test data set into the power transmission line multi-target detection model to obtain the target fault defect; The power transmission line image is used to represent a target detection power transmission line, and the target fault defect image is used to represent an image of a component defect of the target power transmission line that has occurred.
[0016] In a fourth aspect, the present application provides a power transmission line multi-target detection device. The multi-target detection device comprises: an image acquisition module configured to acquire a detection image of a power transmission line to be detected; a result generation module configured to input the detection image into a power transmission line multi-target detection model to acquire an image of a component defect of the power transmission line to be detected that has occurred; The power transmission line multi-target detection model is obtained through the model training device of the third aspect.
[0017] In a fifth aspect, the present application provides a computer readable storage medium storing a computer program, the computer program being executed by a processor to implement the model training method of any one of the first aspect and the power transmission line multi-target detection method of the third aspect.
[0018] The beneficial effects of the present application are: the present application provides a model training method, device and a multi-target detection method and device for a power transmission line. The model training method comprises: acquiring a plurality of component defect images of a power transmission line, preprocessing the component defect images to obtain a training data set, a verification data set and a test data set; replacing the C3K2 module in the original YOLOv11n model with a self-developed CES module, introducing an FFDPN network structure based on the FF module design into the original YOLOv11n model to replace the original FPN, and introducing TADDH to form an ipre-YOLOv11n model; inputting the training data set into the ipre-YOLOv11n model for model training, improving the ipre-YOLOv11n model training parameters based on the training results combined with the loss function, and iteratively training to obtain a power transmission line multi-target detection model; inputting the verification data set and the test data set into the power transmission line multi-target detection model for model evaluation and model testing. The model training method of the present application improves the neural network model of YOLOv11n by replacing the original C3K2 module with a self-developed CES module, which can more effectively integrate and process multi-scale features, and enhance the extraction of small target edge information and the selection ability of features. The FFDPN structure is designed to replace the FPN module of the original model, the FF module is customized in the FFDPN structure to extract rich cross-scale information in parallel and integrate features at multiple scales, ensuring that the features at each scale have detailed context information; the feature diffusion mechanism effectively propagates these context information to each detection scale, enhancing the robustness and accuracy of the model, thereby improving the overall performance of target detection and classification. By replacing the detection head of the original model with TADDH, the concept of task alignment in the detection head is proposed, a novel alignment structure is developed to enhance the connection between the positioning and classification tasks, and the ability of the detection head to accurately position and classify targets is improved. BRIEF DESCRIPTION OF DRAWINGS
[0019] The present application will be further described below in conjunction with the drawings and examples.
[0020] Figure 1 is a flowchart of the model training method of embodiment 1 of the present application; Figure 2 is a structural diagram of the original YOLOv11n model of embodiment 1 of the present application; Figure 3 is a structural diagram of the improved ipre-YOLOv11n model of embodiment 1 of the present application; Figure 4 is a structural diagram of the traditional C3K2 module of embodiment 1 of the present application; Figure 5 is a structural diagram of the improved new CES edge information feature extraction selection module of embodiment 1 of the present application; Figure 6 is a structural schematic diagram of the introduced FF module of embodiment 1 of the present application; Figure 7 is a structural schematic diagram of the FF DPN module constructed based on the FF module of embodiment 1 of the present application; Figure 8 is a structural schematic diagram of the TADDH detection head (module) of embodiment 1 of the present application; Figure 9 is a structural schematic diagram of the model each module ablation experiment result record of embodiment 1 of the present application; Figure 10 is a flow schematic diagram of the power transmission line multi-target detection method of embodiment 2 of the present application; Figure 11 is a module schematic diagram of the model training device of embodiment 3 of the present application; Figure 12 is a module schematic diagram of the power transmission line multi-target detection device of embodiment 4 of the present application. DETAILED DESCRIPTION
[0021] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the present application will be described clearly and completely below in conjunction with the drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.
[0022] Embodiment 1 As shown in Figures 1-9 , the present embodiment provides a model training method, which comprises: Step S1, acquiring a plurality of component defect images of a power transmission line, and pre-processing the component defect images to obtain a training data set, a verification data set and a test data set.
[0023] In the embodiment, specifically, first, a plurality of component defect images of a power transmission line are collected by using a UAV, the plurality of components of the power transmission line include insulators, shock absorbers, grading rings and bird nests; and the component defect images are segmented into a plurality of image feature blocks, the image feature blocks are labeled with defects, and an initial data set is established; then, data augmentation is performed on the images in the initial data set to obtain a target data set; finally, the target data set is divided into the training data set, the verification data set and the test data set according to a preset ratio; the preset ratio can be a 7:2:1 ratio random division. The data augmentation method of the embodiment includes image rotation, image brightness and image noise increase; specifically, the image rotation is to rotate the images in the initial data set around a certain center point by a certain angle, so that the model learns the features of the images under different angles. The image brightness adjustment is to increase or decrease the overall brightness of the image, which is realized by adjusting the brightness channel of each pixel of the image, that is, the red, green and blue channels in the RGB color space. The image noise addition can add random noise with Gaussian distribution to the image or randomly generate white and black pixel points in the image to simulate the noise interference in the image acquisition process. The size of the data set is increased by performing data augmentation on the images in the initial data set, and then the target data set is obtained.
[0024] Step S2, replace the C3K2 module in the original YOLOv11n model with a self-developed CES module, introduce an FFDPN network structure based on the FF module design into the original YOLOv11n model to replace the original FPN, and introduce TADDH to form an ipre-YOLOv11n model.
[0025] Specifically, as shown in Figure 2 The structure of the existing YOLOv11n model mainly includes a backbone network, a neck network and a detection head; the backbone network includes a Conv module, a C3K2 module, an SPPF module and a C2PSA module, wherein C3K2 is a feature extraction-fusion module in YOLOv11n. The C3K2 is composed of a Conv module and a variable number of Bottleneck modules, and the number of Bottleneck modules is determined by the width factor of the model. Bottleneck is a basic module with a residual structure. The neck network accepts the feature layers from the backbone network as input, and performs feature fusion by using the C3K2 module, the Concat module and the upsampling operation, and finally outputs a feature map. The detection head includes three detection branches of large, medium and small target scales, and each branch is responsible for detecting targets in different size ranges.
[0026] In this embodiment, the original YOLOv11n network is improved, and the structure of the ipre-YOLOv11n model improved through step S2 is as shown in Figure 3 The specific improvement steps are as follows: As an optional embodiment, the C3K2 module in the original YOLOv11n model is replaced by a self-developed CES module. Specifically, the structure of the traditional C3K2 module is as shown in Figure 4 The generated intermediate feature map is split into two parts, one of which is directly passed to the final Concat block, and the other is passed to multiple C3k modules for further processing. It is worth noting that the type of Bottleneck can be controlled here through the parameter C3k: if False, it degenerates into a structure similar to C2f; if True, the improved C3k module is used. Usually, the C3k parameter is disabled in the shallow network to reduce the amount of calculation; it is enabled in the deep network to enhance feature extraction. The feature map input to these C3k blocks is processed through a series of convolution, normalization and activation operations, and finally the generated feature map is spliced with the directly transmitted feature map in the Concat block. In the C3k module, the number of Bottleneck modules is defined by the depth muliple (width factor) parameter of the original YOLOv11n model, which means that the depth and computational complexity of the module can be flexibly adjusted according to requirements. The spliced feature map is input to a final convolution block for further processing to generate the final output feature map.
[0027] In this embodiment, a new edge information extraction module (Multi Scale Edge Information Enhance Select Module) is designed, which replaces the Bootleneck module in the traditional C3K2 module to form a new CES edge information feature extraction selection module. The structure of the new CES module is as shown in Figure 5 The specific working process is as follows: first, the input feature map is adjusted to the corresponding size through four adaptive average pooling layers with output scales of [3, 6, 9, 12], second, the four feature maps with different output scales are further processed through a 3x3 convolution to extract four features with different scales, then through a 1x1 convolution to adjust the channel number, and then through an upsampling module to restore the shallow semantic information, and then through a new edge information extraction module (Edgehancer) to further extract four different scale edge information, and finally the four different scale edge information and the original feature information are fused and interacted.
[0028] In this embodiment, in order to select the locally important information in the model and assign higher weights to important objects, Figure 5 The DSM edge information selection attention module is introduced in
[15] to select highly task-relevant edge information and then perform convolution on it for output. The edge information enhancement module average-pools the input information, selects low-frequency information from the features, then subtracts the pooled information from the original information to obtain edge feature information. Convolution is used to further extract edge information, which is then enhanced and output. Finally, the enhanced edge information is added to the original feature information to form the edge feature enhancement module. Overall, the CES selection module not only extracts the original features of the object but also enhances the edge information that is easily overlooked.
[0029] As an optional embodiment, the original FPN is replaced by an FFDPN network structure based on the FF module in the original YOLOv11n model. Specifically, during the multi-scale feature fusion process, due to the inconsistency of features at different layers and the increasing depth of the network, shallow features are easily forgotten. For example, an object may be considered a positive sample in shallow features but a negative sample in high-level features. This feature forgetting affects the detection accuracy of the algorithm.
[0030] In order to solve this problem, in this embodiment, a FF module is designed and introduced. The FF module extracts features from the input feature map through multi-scale feature fusion and deep convolution operation. The structure of the FF module is as follows: Figure 6 As shown in the figure, by combining multi-scale functions with point bending, feature information is captured at different scales, thereby enhancing the network's performance, alleviating the problem of feature forgetting, and enhancing the algorithm's feature fusion capability model. The definition of the FF module is as follows: ; ; ; In the formula, x represents the input feature map, DWConv represents the depth-separable convolution operation, Conv represents the ordinary convolution operation, Upsample represents upsampling, and Concat represents the concatenation of feature maps.
[0031] In this embodiment, a forward focused diffusion pyramid network (FFDPN module) is constructed based on the FF module to realize multi-scale feature extraction and fusion. The FFDPN module structure is as follows: Figure 7 Specifically, combined with Figure 6 and 7The FFDPN module processes three feature maps of different scales through up-sampling (Upsample), down-sampling (ADown) and convolution operations. Parallel deep convolution (DWCONV) is used to capture multi-scale information to generate comprehensive feature maps. The architecture based on the FF module and the FFDPN module can ensure that the feature mapping of each scale contains rich contextual information by performing multi-focus and diffusion operations, enhances the feature representation capability, and thus improves the accuracy of object detection and classification.
[0032] As an optional embodiment, the TADDH is introduced in the original YOLOv11n model; specifically, small targets in power line multi-target detection, such as shock absorbers and insulators, are usually hidden in the overall image and are difficult to identify and detect.
[0033] To solve this problem, in the present embodiment, the TADDH (Target Dynamic Alignment Detection Head) is introduced, the structure of which is as shown in Figure 8 The purpose is to improve the detection performance of small-sized targets by introducing the TADDH detection head. Existing object detector heads usually use separate classification and localization branches, which leads to a lack of interaction between the two tasks. In the multi-task learning framework, effective alignment between different tasks is needed to ensure that the model can learn the functional representation of multiple tasks, such as classification and localization.
[0034] In the present embodiment, the introduced TADDH detection head is a dynamic, lightweight and effective detection head based on task alignment. Specifically, the TADDH uses two Conv_GN convolution blocks to first extract feature maps of different dimensions. Then, through the function extractor, the task interaction function is learned from multiple convolution layers, and the concatenated interaction function is used for classification and localization tasks; after obtaining the interaction function, task decomposition is performed. Then, through the positioning branch, the interactive function is used to generate the offset and mask of DCNV2, and DCNV2 is used for feature selection. The deformable convolution network (DCN) is a variant of convolution operation in deep learning, which allows the convolution kernel to be deformed in space to adapt to the geometric changes in the input feature map. DCNV2 has been optimized and improved based on the original DCN, so that the module is more effective and flexible. Finally, in order to solve the problem of inconsistent target scales detected by each detection head, the features are expanded through the scaling layer.
[0035] Step S3, input the training data set into the ipre-YOLOv11n model for model training, improve the ipre-YOLOv11n model training parameters based on the training results combined with the loss function, and obtain a power line multi-target detection model taking power line images as input and target fault defect images as output through iterative training; wherein the power line image is used to represent the target detection power line, and the target fault defect image is used to represent the image of the component defect of the target power line.
[0036] In the embodiment, the experimental environment includes that the deep learning training environment uses a Pytorch-GPU training framework and is implemented by using a Python language, a CUDA version is 11.7, and a GPU version is NVIDIA GeForce RTX 4060. The network parameters are set, including that an image training size is set to 640X640, an Adam optimizer is used to optimize the network to speed up the training process, and a learning rate is adjusted by using a cosine function; an epoch (training round) is set to 300, and a batch-size (batch size) is set to 8.
[0037] Step S4, input the verification data set into the power line multi-target detection model for model evaluation.
[0038] In the embodiment, the verification data set is input into the power line multi-target detection model for model evaluation, including: inputting the verification data set into the power line multi-target detection model, and evaluating the power line multi-target detection model based on a precision P, a recall R, and an average precision mean mAP. Specifically, an ablation experiment is performed on the ipre-YOLOv11n model improvement module to evaluate the importance of each component or feature of the model and the influence of the component or feature on the overall performance of the model. The specific experimental results are shown in Figure 9 The final experimental results are obtained, and the model parameters before and after improvement are compared as shown in Table 1.
[0039] Table 1 Comparison of performance of model before and after improvement
[0040] The improved model improves the accuracy by 1.5%, the recall rate by 6.0%, and the average precision by 3.9%. It can be seen that the improved ipre-YOLOv11n model has a significant improvement in the multi-target detection of the power line.
[0041] The index formula for evaluating the performance of the model is as follows:
[0042]
[0043]
[0044] Wherein, TP is the number of correctly predicted positive samples, FN is the number of predicted negative samples, and FP is the number of predicted positive samples.
[0045] Step S5, input the test data set into the power transmission line multi-target detection model for model testing to obtain the target fault defect.
[0046] The model training method provided in the embodiment can more effectively integrate and process multi-scale features by replacing the original C3K2 module of YOLOv11n with a self-developed CES module, thereby enhancing the extraction of small target edge information and the selection capability of features. The FFPN structure is designed to replace the FPN module of the original model, the FF module is customized in the FFPN structure to extract rich cross-scale information in parallel and integrate features at multiple scales, ensuring that the features at each scale have detailed context information; the feature diffusion mechanism effectively propagates these context information to each detection scale, enhances the robustness and accuracy of the model, and thus improves the overall performance of target detection and classification. By replacing the detection head of the original model with TADDH, the concept of task alignment in the detection head is proposed, and a novel alignment structure is developed to enhance the connection between the positioning and classification tasks, thereby improving the ability of the detection head to accurately position and classify targets.
[0047] Embodiment 2 As shown in Figure 10 The multi-target detection method for a power transmission line provided in the embodiment includes: Step A1, obtaining a power transmission line image to be detected.
[0048] Step A2, inputting the power transmission line image into a power transmission line multi-target detection model to obtain an image of a component defect of the target power transmission line.
[0049] In the embodiment, the power transmission line multi-target detection model is obtained by the model training method provided in Embodiment 1.
[0050] The multi-target detection method for a power transmission line provided in the embodiment can accurately detect and identify components prone to failure in a power transmission line, such as insulators, shock absorbers, grading rings, and bird nests, thereby improving the inspection efficiency and avoiding missed detection.
[0051] Embodiment 3 As shown in Figure 11As shown, the embodiment provides a model training device, which comprises a data acquisition module 11, a model improvement module 12, a model training module 13, a model evaluation module 14 and a detection result generation module 15.
[0052] The data acquisition module 11 is configured to acquire a plurality of component defect images of the power transmission line, and pre-process the component defect images to obtain a training data set, a verification data set and a test data set.
[0053] The model improvement module 12 is configured to embed a CES module in a C3K2 module of an original YOLOv11n model, and introduce an FF module and a TADDH module into the original YOLOv11n model to constitute an ipre-YOLOv11n model.
[0054] The model training module 13 is configured to input the training data set into the ipre-YOLOv11n model for model training, improve the ipre-YOLOv11n model training parameters based on the training results combined with a loss function, and iteratively train to obtain a power transmission line multi-target detection model taking a power transmission line image as input and a target fault defect image as output; wherein the power transmission line image is used to represent a target detection power transmission line, and the target fault defect image is used to represent an image of a component defect of the target power transmission line.
[0055] The model evaluation module 14 is configured to input the verification data set into the power transmission line multi-target detection model for model evaluation.
[0056] The detection result generation module 15 is configured to input the test data set into the power transmission line multi-target detection model to obtain a target fault defect.
[0057] The embodiment provides a model training device. The model training device is obtained based on the model training method provided in embodiment 1; the embodiment can be mutually coordinated with embodiment 1. The related technical details mentioned in embodiment 1 are still valid in this embodiment, in order to reduce repetition, which will not be repeated here. Accordingly, the related technical details mentioned in this embodiment can also be applied in embodiment 1.
[0058] It is worth mentioning that each module involved in the embodiment is a logical module. In actual application, a logical unit can be a physical unit, or a part of a physical unit, or a combination of multiple physical units. In addition, in order to highlight the innovative part of the present application, the units and modules not closely related to solving the technical problems proposed by the present application are not introduced in this embodiment, but this does not mean that there are no other units in this embodiment.
[0059] Embodiment 4 AsFigure 12 As shown, the embodiment provides a multi-target detection device of a power transmission line, which comprises an image acquisition module 21 and a result generation module 22.
[0060] The image acquisition module is configured to acquire a detection image of the power transmission line to be detected.
[0061] The result generation module is configured to input the detection image into a power transmission line multi-target detection model to acquire an image of a component defect of the power transmission line to be detected.
[0062] In the embodiment, the power transmission line multi-target detection model is obtained by the model training device provided in Embodiment 3.
[0063] The embodiment provides a multi-target detection device of a power transmission line. The power transmission line multi-target detection model obtained by training can accurately detect and identify components prone to failure in the power transmission line, such as insulators, shock absorbers, grading rings, and bird nests, thereby improving the inspection efficiency and avoiding missed detection.
[0064] Embodiment 5 The embodiment relates to a computer readable storage medium storing a computer program. The computer program is executed by a processor to implement the model training method of Embodiment 1 and the multi-target detection method of the power transmission line of Embodiment 2.
[0065] That is, those skilled in the art can understand that all or part of the steps of the above-mentioned embodiment methods can be completed by programs instructing related hardware, the programs are stored in a storage medium, and the programs include a plurality of instructions for causing a device (which can be a single-chip microcomputer, a chip, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The foregoing storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.
[0066] Based on the above ideal embodiments according to the present application, through the above description, relevant personnel can make various changes and modifications without deviating from the scope of the present application. The technical scope of the present application is not limited to the contents in the specification, and must be determined according to the scope of the claims.
Claims
1. A model training method, characterized in that: The model training method includes: Acquire multiple component defect images of the transmission line, and preprocess the component defect images to obtain a training data set, a verification data set, and a test data set; In the original YOLOv11n model, the C3K2 module is replaced with the CES module, and the FFDPN network structure designed based on the FF module is introduced into the original YOLOv11n model to replace the original FPN, and TADDH is introduced to form the ipre-YOLOv11n model; Inputting the training data set into the ipre-YOLOv11n model for model training, improving the ipre-YOLOv11n model training parameters based on the training results and in combination with the loss function, and iteratively training to obtain a transmission line multi-target detection model that takes the transmission line image as input and the target fault defect image as output; Inputting the verification data set into the transmission line multi-target detection model for model evaluation; Inputting the test data set into the transmission line multi-target detection model to perform model testing to obtain the target fault defect; The transmission line image is used to represent the target detection transmission line, and the target fault defect image is used to represent an image of a defective component having a fault in the target transmission line.
2. The model training method according to claim 1, characterized in that The step of pre-processing the multi-target images further comprises: Segmenting the component defect image into a plurality of image feature blocks, marking defects on the image feature blocks, and establishing an initial data set; Performing data augmentation on the images in the initial dataset to obtain a target dataset; The target data set is divided into the training data set, the verification data set and the test data set according to a preset ratio.
3. The model training method according to claim 1, characterized in that The steps of replacing the C3K2 module with the CES module in the original YOLOv11n model include: Design a new edge information extraction module, and replace the Bootleneck module in the C3K2 module with the edge information extraction module to form the CES module; The specific working process of the CES module is as follows: The input feature map is adjusted to the corresponding size through four adaptive average pooling layers with different output scales, and features of four different scales are extracted through large convolution; The features of four different scales are convolved with small kernels to adjust the number of channels, and then restored through the upsampling module to obtain the shallow semantic information of four different scales. The shallow semantic information is passed through the edge information extraction module to further extract edge information of four different scales; The edge information is enhanced and output, and the enhanced edge information and the original feature information are fused and interacted.
4. The model training method according to claim 1, characterized in that The steps of introducing the FFDPN network structure designed based on the FF module to replace the original FPN in the original YOLOv11n model include: the FF module extracts features from the input feature map through multi-scale feature fusion and deep convolution operations; and constructing a forward focused diffusion pyramid network based on the FF module to achieve multi-scale feature extraction and fusion. The specific working process includes: Process three feature maps of different scales through upsampling, downsampling and convolution operations; Multi-scale information is further captured through parallel depthwise convolutions to generate comprehensive feature maps.
5. The model training method according to claim 1, characterized in that The step of introducing TADDH into the original YOLOv11n model includes: the working process of TADDH: First, feature map inputs of different dimensions are extracted through two shared group normalization convolution blocks; The task interaction features are learned from multiple convolutional layers through feature extractors, and the interaction features are concatenated for classification and localization tasks; The offset and mask of DCNV2 are generated using interactive functions through the localization branch, the DCNV2 is used for feature selection, and the features are expanded through the scaling layer.
6. The model training method according to claim 1, characterized in that The step of inputting the verification data set into the power transmission line multi-target detection model for model evaluation and outputting multiple evaluation indicators includes: The validation dataset is input into the transmission line multi-target detection model, and the transmission line multi-target detection model is evaluated based on its precision rate P, recall rate R, and mean average precision (mAP). The performance evaluation formula of the model is as follows: Among them, TP is the number of correctly predicted positive samples, FN is the number of incorrectly predicted negative samples, and FP is the number of incorrectly predicted positive samples.
7. A multi-target detection method for a transmission line, characterized in that: The multi-target detection method comprises: Acquire an image of a transmission line to be inspected; Inputting the transmission line image into a transmission line multi-target detection model to obtain an image of a defective component of the target transmission line; The transmission line multi-target detection model is obtained by the model training method described in any one of claims 1 to 6 above.
8. A model training device, characterized in that: The model training device comprises: A data acquisition module is used to acquire multiple component defect images of the transmission line and pre-process the component defect images to obtain a training data set, a verification data set, and a test data set; A model improvement module is used to replace the C3K2 module with the CES module in the original YOLOv11n model, introduce the FFDPN network structure designed based on the FF module to replace the original FPN in the original YOLOv11n model, and introduce TADDH to form an ipre-YOLOv11n model; A model training module is configured to input the training data set into the ipre-YOLOv11n model for model training, improve the ipre-YOLOv11n model training parameters based on the training results and in combination with a loss function, and iteratively train to obtain a transmission line multi-target detection model that uses the transmission line image as input and the target fault defect image as output; A model evaluation module, configured to input the verification data set into the transmission line multi-target detection model for model evaluation; A detection result generation module, configured to input the test data set into the transmission line multi-target detection model to obtain the target fault defect; The transmission line image is used to represent the target detection transmission line, and the target fault defect image is used to represent an image of a defective component having a fault in the target transmission line.
9. A multi-target detection device for a power transmission line, characterized in that: The multi-target detection device comprises: An image acquisition module, used to acquire a detection image of the power transmission line to be detected; A result generation module is used to input the detection image into a multi-target detection model for a power transmission line to obtain an image of a defective component of the power transmission line to be detected; The transmission line multi-target detection model is obtained through the model training device described in claim 8 above.
10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the model training method according to any one of claims 1 to 6 and the multi-target detection method for transmission lines according to claim 7 are implemented.