Overhead transmission line hidden danger real-time detection method and system

Abstract

The invention discloses a method and a system for detecting hidden dangers of an overhead transmission line in real time, relates to the technical field of intelligent inspection of a power system, and has the advantages of improving small target recognition precision, enhancing model robustness, reducing calculation complexity and improving real-time performance compared with the prior art. Lightweight deployment can be realized on edge equipment with limited resources, and the method is suitable for intelligent inspection tasks of power transmission lines in complex power scenes.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection technology for power systems, specifically to a method and system for real-time detection of hidden dangers in overhead transmission lines. Background Technology

[0002] Overhead transmission lines are a critical component of the power system. Long-term exposure to the natural environment makes them susceptible to factors such as severe weather, bird activity, and mechanical stress, leading to potential hazards such as insulator damage, flashovers, bird nesting, and vibration damper detachment. If these hazards are not detected and addressed promptly, they can cause line faults or even large-scale power outages, severely impacting the safe and stable operation of the power grid. Therefore, efficient and accurate transmission line hazard detection technology is crucial for ensuring the reliability of the power system. Current traditional methods of manual and helicopter inspections suffer from inherent drawbacks such as low efficiency, high cost, and strong subjectivity, making them unsuitable for the high-efficiency inspection requirements of modern power systems.

[0003] With the development of drone technology and computer vision, deep learning-based intelligent detection methods have gradually become the mainstream solution for power transmission line inspection. Currently, deep learning-based target detection algorithms are mainly divided into two categories: single-stage detectors represented by the YOLO series and Transformer architecture detectors represented by the DETR series. While the YOLO series algorithms have a fast detection speed, their anchor box mechanism is unstable in complex backgrounds and has limited ability to detect small targets. Although the DETR series algorithms achieve end-to-end detection through the Transformer architecture, avoiding the computational overhead of non-maximum suppression, they suffer from severe loss of spatial details in high-level features, making them equally unsuitable for small target detection.

[0004] In terms of feature fusion, traditional feature pyramid networks recursively pass features between adjacent levels, causing distant-level features to undergo multiple nonlinear transformations, resulting in severe attenuation of semantic information during the transmission process. Although some studies have proposed improved schemes such as bidirectional feature pyramids, these methods often involve a significant increase in computational cost, making it difficult to achieve lightweight models while maintaining accuracy. Regarding the imbalanced sample problem, existing methods mostly use fixed-weight loss functions, failing to dynamically adjust the training focus according to sample difficulty, leading to insufficient learning of difficult samples.

[0005] The comparison file CN111339882A uses an improved Mask R-CNN scheme, which can detect hidden dangers in power transmission lines. However, its model has problems such as being heavyweight, having a large amount of computation, and having a slow inference speed, which is not conducive to real-time applications of edge devices such as drones. At the same time, it has insufficient accuracy in detecting small targets, relies on NMS post-processing, and has limited coverage of hidden danger types, resulting in poor generalization.

[0006] In summary, existing technologies for detecting hidden dangers in power transmission lines still have significant shortcomings in terms of accuracy in detecting small targets, adaptability to complex backgrounds, computational efficiency, and sample balance. There is an urgent need for a lightweight detection solution that can maintain high detection accuracy while meeting real-time requirements, in order to promote the practical application of intelligent inspection technology in power systems. Summary of the Invention

[0007] This invention relates to the field of intelligent inspection technology for power systems. In order to overcome the defects of the existing technology, the purpose of this invention is to provide a method and system for real-time detection of hidden dangers in overhead transmission lines.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: To solve the above-mentioned technical problems, the objective of this invention can be achieved through the following technical solution: a method and system for real-time detection of hidden dangers in overhead transmission lines, specifically including the following steps: S1. Collect publicly available images of potential hazards in overhead transmission lines, and extract keyframe images from videos of drone inspections of transmission lines as supplementary data to obtain sample data, constructing a dataset of potential hazards in overhead transmission lines covering multiple scenarios and categories. Then, perform data augmentation on the sampled hazard images and divide them into training, validation, and test sets in a 7:1:2 ratio. S2. Based on the target detection model RT-DETR, the backbone network is changed to a lightweight ResNet-WT network. The ResNet network is improved by using the wavelet transform convolution module WTConv to decompose the input features into high-frequency and low-frequency components, so as to enhance the ability to capture multi-scale, especially small hidden danger features, while reducing the computational complexity of the model. S3. Traditional Feature Pyramid Networks (FPNs) recursively pass features between adjacent layers, leading to attenuation of feature information in distant layers and severe loss of details in small targets. To address this, a cross-layer multi-scale aggregated feature fusion network (CMIAFPN) is proposed to replace the original CCFM cross-scale feature fusion network. It includes a multi-layer feature fusion module (MLF) and an enhanced multi-scale fusion module (EMF). This network directly connects deep and shallow features through cross-layer skip connections, shortening the information path, preserving key details, and preventing the attenuation of small target feature information during transmission. S4. Improve the downsampling module in RT-DETR to the LDown dynamic downsampling convolution module. By learning the offset, the sampling position is dynamically adjusted to avoid the loss of key features caused by fixed downsampling and to retain the features of small targets. S5. Based on the Wise-IoU loss function, a dynamic weighting mechanism of Focal Loss is added to obtain the Focal-Wise-IoU dynamic focusing regression loss function. Through the dynamic gradient gain mechanism, higher training weights are given to small target samples that are difficult to identify, thereby improving the convergence speed and detection accuracy of the model. S6. Use the above modules to build a complete detection model, train it on the processed dataset, and evaluate its performance on the validation set to determine the weight parameters that perform best in the validation set. S7. Apply the found optimal weight parameters to the improved model to perform real-time, high-precision overhead transmission line hazard detection tasks on actual inspection images or videos.

[0009] In a preferred embodiment, step S1, the process of creating a dataset of potential hazards for overhead transmission lines, includes the following steps: (1) First, collect publicly available pictures of potential hazards in overhead transmission lines, including types such as insulator damage, flashover, bird nests, and detached anti-vibration hammers.

[0010] (2) Use tools such as LabelImg to label the hazard categories and corresponding target detection boxes.

[0011] (3) Use the Albumentations tool to perform data augmentation on images of potential hazards on overhead transmission lines, such as rain and fog simulation and brightness adjustment, and perform data balancing on images of potential hazards on different types of transmission lines to improve the generalization ability of the model.

[0012] (4) Construct a dataset of hidden dangers in overhead transmission lines, which includes training set, validation set and test set.

[0013] In a preferred embodiment, step S2, the construction of the lightweight ResNet-WT network includes the following steps: (1) Replace the 3×3 convolution in the ResNet18 backbone network with the convolution block WTConv based on Haar wavelet transform to reduce redundant convolution operations and reduce the number of model parameters.

[0014] (2) WTConv decomposes the input features into one low-frequency sub-band (LL) and three high-frequency sub-bands (horizontal HL, vertical LH, and diagonal HH) through wavelet transform. The low-frequency sub-band preserves the main structure of the image, while the high-frequency sub-band captures detailed features such as edges and textures.

[0015] (3) Use depthwise separable convolution to enhance the local context for low-frequency features, and use standard convolution to perform cross-channel information interaction for high-frequency features.

[0016] (4) Finally, the optimized high and low frequency features are fused by inverse wavelet transform (IWT) and combined with the baseline features of the initial convolutional layer to reconstruct and generate the final output features.

[0017] In a preferred embodiment, step S3 involves replacing the original CCFM cross-scale feature fusion network in the original RT-DETR with the CMIAFPN cross-level multi-scale aggregated feature fusion network, which includes the following steps: (1) Replace the traditional CCFM’s recursive transmission mode of “top-down + bottom-up” and construct a direct cross-level connection path to directly associate the shallow features (S2, S3) and mid-level features (S4) output by the backbone network with the high-level features (F5) processed by AIFI, thus shortening the feature propagation path.

[0018] (2) Multi-scale feature alignment: Multi-scale features are aligned by upsampling and LDown dynamic downsampling operations to ensure that feature details are not lost during scale transformation.

[0019] (3) After splitting the aligned S2, S3, S4, and F5 features according to the channel dimension, they are then concatenated. The formula is as follows: Where C is a 1×1 convolution, Using the SiLU activation function, shallow details such as the surface texture of insulators and high-level semantics such as the overall shape of the bird's nest are preserved through cross-level splicing.

[0020] (4) Based on the MLF output, introduce features from higher levels to further enrich feature diversity: Among them, the newly added S2 downsampling feature provides supplementary local details, while the F5 upsampling feature enhances global semantic guidance.

[0021] (5) Extract three sets of features from different stages of CMIAFPN, and unify the number of channels through 1×1 convolution to adapt to the input requirements of the subsequent Transformer decoder.

[0022] In a preferred embodiment, step S4 involves replacing the original RT-DETR fixed downsampling module with an LDown dynamic downsampling convolution module, including the following steps: (1) Define a uniform set of sampling points on the input feature map and generate a dynamic offset matrix corresponding to the sampling points through 1×1 convolution; (2) The offset is superimposed on the initial sampling point to obtain the dynamically adjusted sampling coordinates, so that the sampling point can adaptively focus on the high response region; (3) Perform bilinear interpolation sampling based on the optimized sampling coordinates to obtain feature values ​​and retain sub-pixel level detail information; (4) Apply 3×3 depthwise separable convolution to the sampled feature map to enhance the local context and compress channel redundancy, and finally output the downsampled feature map, so as to retain the key features of small targets to the greatest extent while reducing the resolution.

[0023] In a preferred embodiment, step S5, which improves the original Wise-IoU loss to a Focal-Wise-IoU dynamic focusing regression loss function, includes the following steps: (1) Define the dynamic gradient gain coefficient based on IoU. When IoU is low, the coefficients increase exponentially, causing the model to prioritize optimizing the predicted boxes for hard samples, where δ and γ are adjustment parameters; (2) Introduce FocalLoss scaling factor to further suppress the gradient contribution of easy samples, so that small targets and difficult-to-identify hidden dangers account for a larger proportion in training; (3) Based on the center offset penalty term of Wise-IoU, scale sensitivity adjustment for small targets is added to strengthen the constraint on the positioning accuracy of small targets. ,in, Indicates the center coordinates of the prediction box. The coordinates of the center of the true bounding box are represented by W and H, which are the width and height of the true bounding box, respectively.

[0024] (4) The dynamic gradient gain coefficient, the center offset penalty term, and the IoU regression loss are fused together to form the final Focal-Wise-IoU loss function. This effectively improves the accuracy and convergence speed of the model for small target detection in complex backgrounds.

[0025] A real-time detection system for hidden dangers in overhead transmission lines includes a data preprocessing module, a backbone network improvement module, an improved feature fusion module, a loss function optimization module, a model training and weight selection module, and an inference detection module. Data preprocessing module: Collect publicly available datasets of potential hazards in power transmission lines and images obtained by extracting key frames from drone inspection videos as datasets. Label and enhance the data to build a dataset of potential hazard images covering multiple categories and scenarios, and split it into training set, validation set and test set in a ratio of 7:1:2. Backbone network improvement module: Based on the target detection model RT-DETR, the backbone network is improved by replacing it with a lightweight ResNet-WT network. By introducing the wavelet transform convolution module WTConv, the input features are decomposed into high and low frequency components, thereby reducing computational complexity while improving the feature extraction capability for small targets and texture details. Improved feature fusion module: The cross-scale feature fusion network in RT-DETR is replaced with the cross-level multi-scale aggregated feature fusion network CMIAPPN. This network achieves efficient fusion of shallow details and high-level semantics through cross-level skip connections, multi-level feature fusion module MLF and enhanced multi-scale fusion module EMF, improving the detection accuracy of small targets in complex backgrounds. At the same time, the original fixed stride downsampling is replaced with LDown dynamic downsampling convolution module, which adaptively adjusts the sampling position by learning the sampling point offset, avoiding the loss of important features and enhancing the network's sensitivity to small targets. Loss function optimization module: Based on the Wise-IoU loss function, the FocalLoss dynamic weighting mechanism is introduced, and the Focal-Wise-IoU loss function is proposed. By combining the dynamic gradient gain coefficient with the center offset penalty term, the detection accuracy of the model on difficult samples and small targets is improved. Model training and weight selection module: The training set is input into the improved RT-DETR model for training, the model performance is evaluated on the validation set at different rounds, and the best weights are selected based on metrics such as mAP, recall and F1-score. Inference Detection Module: The obtained optimal weights are loaded into the improved RT-DETR model to perform real-time inference on the inspection images or videos of overhead transmission lines, and output the category and location information of hidden dangers, so as to realize real-time detection of hidden dangers of overhead transmission lines in multiple categories and multiple scenarios.

[0026] Compared with the prior art, the present invention has the following beneficial effects. 1. This invention improves the backbone network by adopting a lightweight ResNet-WT network based on wavelet transform convolution. The input features are decomposed into high and low frequency components through wavelet decomposition, which not only preserves the overall structural information of the image, but also enhances the ability to express details such as texture and edge features, effectively improving the feature extraction capability of small target hazards. At the same time, it reduces redundant convolution operations, reduces computational complexity, and enables the model to run efficiently on edge computing devices. 2. In the feature fusion part, this invention introduces a cross-level multi-scale aggregated feature fusion network CMIAPPN. It utilizes a multi-level feature fusion module (MLF), an enhanced multi-scale fusion module (EMF), and a local dynamic convolutional downsampling module (LDown) to shorten the feature transmission path through cross-level skip connections, reduce the attenuation of small target features during transmission, and achieve efficient complementarity between shallow details and high-level semantics, thereby improving the detection robustness and small target recognition accuracy in complex backgrounds. 3. This invention improves the loss function by proposing the Focal-Wise-IoU loss function. Based on Wise-IoU, a FocalLoss scaling factor is introduced to construct a dynamic gradient gain mechanism. This gives higher training weights to samples with low IoU and difficult detection, effectively improving the accuracy of small target hazard detection and improving the model convergence speed and training stability. 4. While ensuring detection accuracy, this invention also takes into account lightweight design and real-time performance, reduces the number of model parameters and computational complexity, and can be deployed and run efficiently in real time on edge devices such as drones and embedded GPUs, meeting the needs of power inspection in real-time and resource-constrained environments. Attached Figure Description

[0027] Figure 1 This is an improved network flowchart of the present invention; Figure 2 A schematic diagram of the improved ResNet-WT backbone network structure; Figure 3 A schematic diagram of the improved WTconv convolutional module; Figure 4 This is a schematic diagram of the CMIAFPN feature fusion network structure; Figure 5 This is a schematic diagram of the inference for the Focal-Wise-IoU loss function. Detailed Implementation

[0028] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0030] Example Please see Figure 1-5 The present invention provides a technical solution: Example 1: Please refer to Figure 1 (As shown in the flowchart of the overall method of the RT-DETR improved model), the real-time detection method for hidden dangers in overhead transmission lines based on DETR improvement described in this embodiment includes the following steps: S1. Construct a dataset of hidden dangers in overhead transmission lines. By collecting publicly available images of potential hazards and keyframes from drone inspection videos, and combining this with manual annotation tools (such as LabelImg), hazard categories such as insulator damage, flashover, bird nests, and vibration damper detachment were labeled. Subsequently, data augmentation was performed using the Albumentations tool (brightness adjustment, rain and fog simulation, rotation and cropping, etc.), and the data was divided into training, validation, and test sets in a 7:1:2 ratio to ensure sample diversity and data balance.

[0031] S2. Feature extraction based on the lightweight backbone network ResNet-WT Based on the object detection model RT-DETR, the original ResNet18 backbone is replaced with a lightweight ResNet-WT network. This network introduces wavelet transform convolution (WTConv) into the convolutional layers, splitting the input features into low-frequency and high-frequency subbands through Haar wavelet decomposition. The low-frequency part preserves the global structure, while the high-frequency part captures edge and texture details. The low-frequency part uses depthwise separable convolution to reduce computation, while the high-frequency part uses standard convolution to enhance cross-channel information. Finally, it is reconstructed through inverse wavelet transform (IWT) to obtain a feature representation that takes into account both global and detailed characteristics.

[0032] S3, Improved Feature Fusion Module A cross-level multi-scale aggregated feature fusion network (CMIAFPN) is adopted, which directly connects shallow features and deep semantics through cross-level skip connections, avoiding the information attenuation caused by recursive propagation in traditional FPN. The network consists of a multi-level feature fusion module (MLF) and an enhanced multi-scale fusion module (EMF): the MLF is responsible for cross-layer concatenation and feature alignment, preserving shallow texture information; the EMF enhances the complementarity of global semantics and local details by introducing far-layer features and upsampling operations. Local dynamic convolutional downsampling (LDown) is introduced in the downsampling stage, which adaptively adjusts the sampling position by learning the offset, thereby preserving key features of small targets while reducing resolution.

[0033] S4, Loss Function Optimization Module The Wise-IoU loss function is improved to Focal-Wise-IoU loss. Its loss function form is: in, It is a dynamic gradient gain coefficient that can boost the weights of low-IoU hard samples during training; A center-based penalty term constrains the geometric difference between the predicted bounding box and the ground truth bounding box. This mechanism automatically focuses on small targets and hard-to-identify hidden dangers during training, effectively improving detection accuracy and convergence speed.

[0034] S5, Model Training and Weight Selection Module The improved model was trained end-to-end using the dataset, and the model was evaluated at each training epoch using the validation set. Metrics such as mAP, recall, and F1-score were calculated. An early stopping mechanism was employed to avoid overfitting, and the best-performing weight file was ultimately selected as the parameters for the detection model.

[0035] S6, Inference Detection Module During the detection phase, the inspection images or videos of the transmission line to be tested are input into the improved model loaded with the best weights. The system outputs the category and location information of the hidden dangers, which can realize the real-time detection and location of various hidden dangers such as insulator damage, bird nests, and vibration damper detachment.

[0036] Furthermore, it should be noted that the specific embodiments described in this specification may be given different names, and the above description is merely illustrative of the structure of the present invention. All equivalent or simple variations made based on the construction, features, and principles of the present invention are included within the scope of protection of the present invention. Those skilled in the art can make various modifications or additions to the described specific examples or adopt similar methods, as long as they do not deviate from the structure of the present invention or exceed the scope defined by the claims, all of which should fall within the scope of protection of the present invention.

Claims

1. A real-time detection method for hidden dangers of overhead transmission lines, characterized in that, The method comprises the following steps: S1, constructing overhead transmission line hazard dataset: collecting public dataset and unmanned aerial vehicle power line inspection video frames provided by the power grid, labeling insulator damage, flashover, bird nest nesting, shock absorber shedding and other hazard categories and real boxes, and splitting training set, verification set and test set; S2, replacing the backbone network of RT-DETR with an improved lightweight backbone network ResNet-WT to extract features of the input image, wherein the ResNet-WT replaces part of the 3x3 convolution in the original ResNet backbone network with a wavelet transform convolution block WTConv, decomposes the input features into high-frequency and low-frequency subbands for processing respectively, separates high-frequency hazard features from low-frequency background features, expands the receptive field and reduces the computational complexity; S3, inputting the multi-scale features extracted by the backbone network into a cross-level multi-scale aggregated feature fusion network CMIAFPN, connecting shallow and deep features through cross-level skip connection, shortening the propagation path for efficient feature fusion; S4, using a Focal-WIoU loss function to train the model, wherein the loss function combines the scaling factor of FocalLoss and the dynamic non-monotonic focusing mechanism of Wise-IoU to give high weight to difficult samples with low IoU; S5, evaluating the model weight saved in different training rounds using the verification set during the training process, selecting the weight file with the best performance based on the average precision mean (mAP), recall rate and F1-score, and using it as the final detection model parameter; S6, loading the best weight obtained from the verification set into the improved RT-DETR model, performing inference on the input power line inspection image or video frame, and outputting the hazard detection result including the hazard category and location information, thereby realizing real-time detection and positioning of overhead transmission line hazards.

2. The real-time detection method for overhead transmission line hazards according to claim 1, characterized in that, The wavelet transform convolution block (WTConv) processing process in step S2 comprises the following steps: (1) collecting public datasets and unmanned aerial vehicle power line inspection videos provided by the power grid, and intercepting typical frames related to overhead transmission lines; (2) data labeling of the collected images, including but not limited to insulator damage, flashover, bird nest nesting, shock absorber shedding, etc., and drawing the corresponding real box; (3) data augmentation of the dataset, removing blurred, repeated and invalid samples, and using rotation, scaling, brightness adjustment, random cropping and other data augmentation methods to improve data diversity and robustness; (4) splitting the dataset according to the proportion, dividing it into training set, verification set and test set, and ensuring the balance of different category sample distribution, providing high-quality data basis for model training and evaluation.

3. The method of claim 1, wherein the method further comprises: The wavelet transform convolution block (WTConv) processing process in step S2 comprises the following steps: (1) using Haar wavelet transform to perform frequency domain decomposition on the input features to obtain low-frequency subband (LL) and high-frequency subband (horizontal HL, vertical LH and diagonal HH); (2) the low-frequency subband (LL) is processed through 3x3 depth separable convolution and 1x1 point-by-point convolution to enhance local context and compress channel redundancy; (4) High-frequency subbands (HL, LH, HH) are aligned in channels by 1x1 convolution and then 3x3 standard convolution is performed to realize cross-channel interaction and edge texture enhancement; (5) The inverse wavelet transform (IWT) is performed on each subband processed above to reconstruct, and then the reconstructed features are combined with the benchmark features obtained by initial convolution to generate the final output features.

4. The method of claim 1, wherein the method further comprises: The process of the cross-level multi-scale aggregated feature fusion network CMIAFPN in step S3 includes the following steps: (1) A multi-level feature fusion module (MLF) is used to directly fuse the low-level features and high-level features of the backbone network through cross-level skip connection, avoid information attenuation caused by traditional recursive transmission, fully retain shallow detail features, and alleviate the problem of small target features being covered by high-level semantics; (2) An enhanced multi-scale fusion module (EMF) is introduced in the deep layer of the fusion network, which integrates shallow, middle and deep features through a multi-directional connection mechanism, and combines up-sampling and down-sampling operations to form a complement of global information and local details, further improving the robustness and accuracy of detection; (3) A local dynamic convolution down-sampling module (LDown) is used in the down-sampling stage, which adjusts the sampling position by introducing a dynamic offset, and adaptively focuses on key areas and small target features using a convolution kernel, reducing the loss of important features in the down-sampling process, and improving the sensitivity of the detection network to small targets in complex backgrounds; (4) The above sub-modules are connected in sequence to form the cross-level multi-scale aggregated feature fusion network CMIAFPN, and a gating mechanism is used to selectively filter the cross-level feature channels to avoid redundant transmission of irrelevant features, and finally output the optimized multi-scale feature map for the detection head to call, realizing efficient recognition and positioning of different scale overhead transmission line hazard targets.

5. The method of claim 1, wherein the method further comprises: The process of the Focal-Wise-IoU loss function in step S4 includes the following steps: (1) A dynamic gradient gain coefficient r is used to assign higher training weights to difficult samples in low IoU samples during the regression process, so that the model pays more attention to low-quality prediction boxes during optimization, thereby improving the detection accuracy in small targets and complex backgrounds; (2) The overall expression of the Focal-Wise-IoU loss function is: wherein , is an anchor box center offset penalty term, and r is a dynamic gradient gain coefficient. (3) the anchor frame center offset penalty term The mathematical expression is as follows: Wherein, (x, y) represents the center coordinates of the prediction frame, (x gt ,y gt） represents the center coordinates of the target frame, and W and H are the width and height of the target frame, respectively. (4) The definition of the dynamic gradient gain coefficient r is: wherein δ is a scaling factor for adjusting the weight distribution between high and low quality samples; and γ is a focus parameter for controlling the enhancement intensity on difficult samples.

6. The method of claim 1, wherein the method further comprises: The processes of steps S5 and S6 include the following steps: (1) During training, the performance of the model parameters saved at different training rounds is evaluated using the validation set, and indicators such as mean precision, recall, and F1-score are calculated; (2) An early stopping mechanism is used to monitor the validation set loss, and when the performance of the validation set does not improve for several consecutive rounds, training is stopped to avoid overfitting; (3) The optimal weight file is selected as the parameter of the final detection model by combining various performance indicators; (4) The optimal weight is loaded into the improved RT-DETR model, and the model is used to detect hazards in the test set or actual inspection images, outputting the target class and position information of the overhead transmission line hazards.

7. The overhead transmission line hazard real-time detection system based on the improved DETR, used to realize the detection method of any one of claims 1-6, characterized in that, It includes: data preprocessing module, backbone network improvement module, improved feature fusion module, loss function optimization module, model training and weight selection module, inference detection module; Data preprocessing module: collect, label and enhance the inspection images of overhead transmission lines, construct a high-voltage transmission line hazard dataset covering multiple scenarios and multiple categories of hazards, and reasonably split it into training set, validation set and test set; Improved backbone network module: deploy a lightweight backbone network ResNet-WT based on wavelet transform convolution to extract multi-scale features from input images, and decompose images into low-frequency structural information and high-frequency detail information; Feature fusion module: use cross-level multi-scale aggregated feature fusion network CMIAFPN to optimize the cross-layer feature transmission path and strengthen the complementarity of different scale features through multi-level feature fusion module (MLF), enhanced multi-scale fusion module (EMF) and local dynamic convolution downsampling module (LDown); Model training and weight selection module: end-to-end training based on the improved model to obtain a lightweight real-time detection model; Inference detection module: input the image of the transmission line to be detected in the detection stage, and output the target box and category of different types of hazards to realize real-time and accurate detection of overhead transmission line hazards.

Images