A defect detection method for unmanned aerial vehicle power line inspection

Abstract

The application discloses a kind of defects detection methods for unmanned aerial vehicle power transmission line inspection, belong to electric power inspection technical field.The application is collected by unmanned aerial vehicle transmission line defect image and completes labeling and data enhancement, input backbone network with pre-training weight extracts multi-scale feature.Multi-scale feature is respectively input into double-branch structure, constructs joint loss function and is iteratively trained in reverse, and scene adaptive defect detection model is obtained after convergence.In the process of inspection, rely on geographic location identification to distinguish scene type, call corresponding branch detection and output defect location and category.According to defect position correlation mapping intersection section line, generate defect correlation candidate line set;Combined with fault mechanism library, match defect cause dimension and quantify contribution degree, determine the line set to be adjusted by weighted calculation.Combined with defect category, construct multi-objective optimization model, check the rationality of power grid power flow of multi-class parameter scheme, solve constraint condition, obtain transmission line operation adjustment scheme.

Description

Technical Field

[0001] This invention relates to the field of power line inspection technology, specifically a defect detection method for unmanned aerial vehicle (UAV) power line inspection. Background Technology

[0002] Transmission lines are the core carriers of electrical energy transmission, and their safe and stable operation is directly related to the reliability of the power grid. With its advantages of high flexibility, wide inspection range, and high safety, drone inspection has gradually replaced manual inspection and become the mainstream method for defect detection in transmission lines. Currently, deep learning-based target detection algorithms are widely used in drone-based transmission line defect detection scenarios, effectively achieving automatic identification of common defects such as insulator damage, broken conductor strands, and hardware corrosion.

[0003] However, existing transmission line defect detection technologies still have many significant shortcomings in practical engineering applications, making it difficult to adapt to complex inspection scenarios and the closed-loop requirements of power grid operation and maintenance. First, most mainstream detection models adopt a single, fixed network structure, failing to distinguish between routine transmission line inspection scenarios and challenging scenarios involving dense small defects, complex backgrounds, and target occlusion, resulting in poor model generalization ability. Second, existing models lack targeted multi-scale density feature enhancement mechanisms, resulting in insufficient multi-scale feature extraction capabilities for small, dense defects in transmission lines. The fusion effect between shallow detail features and deep semantic features is poor, easily leading to the loss of small target defect features. Finally, traditional power grid parameters often employ a single optimization criterion, lacking multi-objective constraint optimization and power grid power flow rationality verification mechanisms. Parameter adjustment schemes are poorly targeted and have low adaptability, easily leading to problems such as power flow exceeding limits and operational instability. Summary of the Invention

[0004] The purpose of this invention is to provide a defect detection method for unmanned aerial vehicle (UAV) power line inspection, so as to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: Firstly, this application provides a defect detection method for unmanned aerial vehicle (UAV) power line inspection, comprising the following steps: Images of defective transmission lines are sampled using drones to generate annotation files; the annotation files are then augmented with data and input into a backbone network with pre-trained weights to extract multi-scale features. The multi-scale features are input into the first branch and the second branch respectively; a joint loss function is constructed and the network parameters are updated through error backpropagation; the network is iteratively trained until it converges to obtain a scene-adaptive defect detection model; during the inspection process, the geographic location identifier of the current location is obtained, the scene category of the current section is determined, and the defect location and defect category are output through the scene-adaptive defect detection model. Based on the location of the defect, the defect is associated with all the transmission lines in the crossing section to generate a set of candidate lines associated with the defect; a transmission line fault mechanism library is constructed to match the causal association dimension of the current defect category and quantify the causal contribution. A weighted summation algorithm is used to calculate the comprehensive causal contribution of each line in the defect-associated candidate line set to the defect, generating a set of transmission lines to be adjusted. A multi-objective optimization objective function is constructed in combination with the defect category, and a constrained multi-objective optimization equation system is constructed to verify the rationality of the power flow distribution under different parameter adjustment schemes. The constrained equation system is solved to obtain the transmission line adjustment scheme.

[0006] In conjunction with the first aspect, in a first embodiment of the first aspect of this application, the step of data augmentation of the labeled file, inputting it into a backbone network equipped with pre-trained weights, and extracting multi-scale features includes: The image size is adjusted, and random photometric distortion, random resizing, random IoU cropping, random horizontal flipping, and mosaic data enhancement are applied to the image. The image format is then converted to Tensor format. Multi-scale features are extracted from the image using a backbone network with pre-trained dinov3 weights. The backbone network uses ConvNext v2.

[0007] In conjunction with the first aspect, in a second embodiment of the first aspect of this application, the step of inputting the multi-scale features into the first branch and the second branch respectively includes: The first branch is the difficult scenario branch, and the second branch is the normal scenario branch; The first branch constructs a multi-scale feature pyramid to complete multi-scale hybrid encoding, generates a pixel-level density map, calculates the batch adaptive query number, and outputs the defect detection result; the second branch constructs a lightweight multi-scale feature pyramid, inputs it into a lightweight decoder to complete decoding inference with a fixed number of queries, and outputs the defect detection result.

[0008] In conjunction with the first aspect, in the third embodiment of the first aspect of this application, the first branch constructs a multi-scale feature pyramid to complete multi-scale hybrid encoding, generates a pixel-level density map, calculates the batch adaptive query number, and outputs the defect detection result; the second branch constructs a lightweight multi-scale feature pyramid, inputs it into a lightweight decoder to complete decoding inference with a fixed number of queries, and outputs the defect detection result, including: Multi-scale feature maps are extracted from different depth stages of the first branch. After unifying the channel dimensions through a convolutional module, a multi-scale feature fusion encoder is constructed for the first branch to perform multi-scale hybrid encoding. The multi-scale feature fusion encoder outputs a first feature pyramid containing five levels. Each layer of features has different channel dimensions and spatial resolution, among which, This is a shallow, native feature layer with medium to low resolution. This is a medium-to-high resolution shallow native feature layer. This is the highest resolution shallow native feature layer. This is a mesoscale deep fusion feature layer. For large-scale deep fusion feature layers; Based on the first feature pyramid, the features of each layer are mapped to a unified hidden dimension H through the input projection layer of the feature fusion encoder of the first branch, resulting in a projected feature sequence. Each projected feature in the projected feature sequence is flattened along the spatial dimension and stitched together to construct the encoder output sequence. Where R represents the real number field, B is the batch size, and i is the index of the first feature pyramid level. Let be the spatial height of the feature map of the i-th layer. Let be the spatial width of the feature map of the i-th layer; record the spatial shape of the features of each layer. In M The corresponding sequence is input to the density regression module, which uses an encoder and decoder structure to generate a pixel-level density map and obtains a continuous estimate of the number of image instances through integration. The continuous estimate is then converted into an adaptive query number through a nonlinear mapping function. Based on the continuous estimate, a regression head is used to map the density value to a continuous regression value, and the adaptive query quantity for each batch of samples is calculated. ,in This represents the number of queries from the first sample to the Bth sample in the current batch. The maximum number of queries in this batch is determined as the unified number of queries for the current batch. Based on the pixel-level density map, a multi-scale density feature is generated using a multi-scale feature extractor. The multi-scale density feature and M are then input into a multi-scale density feature enhancement module. This module includes a channel attention mechanism and a spatial gating unit, which adaptively weights feature channels at different scales according to the category distribution, and outputs the enhanced memory feature. ; Will Instead of M, input the Transformer decoder and detector head of the RT-DETR network, and set the number of queries generated by the decoder to the uniform number of queries for the current batch; generate the bounding box coordinate regression results and class regression results of the first branch through the decoder; Multi-scale feature maps are extracted from different depth stages of the backbone network. After unifying the channel dimensions through a convolutional module, a multi-scale feature fusion encoder is constructed for the second branch to perform multi-scale hybrid encoding, resulting in a second feature pyramid with four levels output by the encoder. ; The second feature pyramid is input into a lightweight decoder without feature map augmentation, and the number of queries generated by the decoder is set to a fixed value; the decoder generates the predicted bounding box coordinate regression results and category regression results of the second branch.

[0009] In conjunction with the first aspect, in the fourth embodiment of the first aspect of this application, the construction of the joint loss function and the updating of network parameters through error backpropagation includes: Through the loss function Calculate the class regression loss for the first branch and the second branch separately, using the following formula: ; Where N is the number of samples, and C is the number of classification categories in the dataset. One-hot encoding of the real label. This refers to the results of category regression, specifically the predicted samples. The probability of belonging to a type C defect; Loss function based on coordinate regression Calculate the coordinate regression loss for each predicted bounding box in the first and second branches separately, using the following formula: ; in, Let x be the x-coordinate of the center point of the true bounding box of the k-th defect target. Let y be the center point of the true bounding box of the k-th defect target. Let be the width of the true bounding box of the k-th defect target. Let the height be the actual bounding box height of the k-th defect target. The x-coordinate of the center point of the k-th defect target prediction box output by the model. The ordinate of the center point of the k-th defect target prediction box output by the model. Let be the width of the predicted bounding box for the k-th defect target output by the model. The height of the predicted bounding box for the k-th defect target output by the model; Calculate the generalized intersection-union ratio (GUU) loss for the first and second branches respectively. The smaller the loss value, the higher the overlap between the predicted box and the true box, the closer the position, and the more accurate the defect localization. Specifically, the coordinate parameters of all predicted bounding boxes of defects output from the first and second branches, as well as the coordinate parameters of the corresponding ground truth bounding boxes, are obtained to achieve one-to-one matching and alignment between the predicted and ground truth bounding boxes. For each set of matched predicted and ground truth bounding boxes, the areas of their intersection and union regions are calculated. The ratio of the intersection to the union is used to obtain the basic cross-union ratio (CUNR), which characterizes the basic overlap between the two boxes. The area of ​​the minimum bounding rectangle that can completely enclose both the predicted and ground truth bounding boxes is calculated, and the generalized CUNR is obtained by combining it with the basic CUNR. The CUNR loss value for a single set of boxes is obtained by subtracting the generalized CUNR from 1. The loss values ​​of all defect targets within a single branch are averaged to generate the total CUNR loss for the first and second branches, respectively.

[0010] Calculate the count regression loss of the first branch. The formula is: ; Where m is the total number of image samples in the current batch. For the first The total number of real defects in the images. For the first Total number of images showing predicted defects; Calculate the two-branch consistency loss The formula is: ; in, The feature vector output by the first branch. This is the feature vector output by the second branch; The total loss is obtained by weighting and summing the category regression loss, coordinate regression loss, and generalized intersection-union loss of the two branches with the count regression loss and bi-branch consistency loss of the first branch. The total loss is then input into the optimizer, and the network parameters are optimized through error backpropagation.

[0011] In conjunction with the first aspect, in the fifth embodiment of the first aspect of this application, the step of obtaining the geographic location identifier of the current location, determining the scene category of the current segment, and outputting the defect location and defect category through a scene adaptive defect detection model during the inspection process includes: A transmission line segment difficulty database is pre-established, including difficult scenario segment identifiers and normal scenario segment identifiers. The trained model weights are exported into the ONNX open neural network exchange format. After operator fusion and inter-layer optimization using the TensorRT inference engine, the model is deployed on a UAV platform. Before inspection, the UAV loads the transmission line segment difficulty database and determines the segment type of the current shooting location in real time based on GPS positioning or preset waypoint information. If the current segment is marked as a difficult scenario, the first branch is activated for detection; if the current segment is marked as a normal scenario, the second branch is activated for detection.

[0012] In conjunction with the first aspect, in the sixth embodiment of the first aspect of this application, the step of associating and mapping the defect with all crossing section transmission lines based on the defect location to generate a defect-associated candidate line set includes: The data of the transmission line ledger is obtained. The three-dimensional spatial topology construction algorithm and the nearest neighbor matching algorithm are used to delineate the three-dimensional spatial ownership boundary of the core components of each transmission line in the crossing section, mark the line attributes, multi-line shared attributes of each component, as well as the physical adjacency relationship and electrical coupling relationship between the lines, restore the spatial layout and ownership boundary of all transmission lines in the crossing section, and generate a three-dimensional spatial topology model of the transmission lines in the crossing section. Based on the aforementioned three-dimensional spatial topology model, with the defect three-dimensional spatial coverage area as the core, a three-dimensional spatial distance matching algorithm and a spatial intersection verification algorithm are used to calculate the shortest spatial distance between the defect three-dimensional spatial coverage area and the three-dimensional spatial boundary of each transmission line core component in the topology model, and to verify the spatial intersection relationship; a preset spatial association judgment threshold is set to filter out transmission lines whose shortest spatial distance is less than the preset threshold or have spatial intersection. Based on the defect category, an ownership verification algorithm and an electrical coupling degree analysis algorithm are used. When a defect is physically attached to a component of a single transmission line, that line is marked as a core associated line. When a defect is physically attached to a common component of multiple transmission lines, all transmission lines associated with that common component are marked as core associated lines. When the defect is an electrical defect and is located within the electromagnetic coupling zone crossed by the transmission lines, all transmission lines forming that crossing zone and having electrical coupling associations are included in the core associated range. When the defect is a mechanical defect, only transmission lines with physical attachment associations are retained. A set of candidate lines for defect associations is generated.

[0013] In conjunction with the first aspect, in the seventh embodiment of the first aspect of this application, the step of constructing a transmission line fault mechanism library, matching the causal correlation dimension of the current defect category, and quantifying the causal contribution includes: The fault occurrence mechanisms of all types of defects in transmission lines are compiled, and independent entries are divided according to defect categories. Each entry clarifies the causal correlation dimension of the corresponding defect, the mapping rules between the causal factors and line operating parameters, and the causal influence weight benchmark, thus forming a fault mechanism library for transmission lines. Based on the current defect category, the fault mechanism database of transmission lines is retrieved, and the corresponding causal association dimension is obtained. Based on the causal association dimension, a weight allocation algorithm is used to determine the influence weight of each dimension. Combined with the standardized operating data of each line in the defect association candidate line set, the causal contribution of each line to the current defect is quantitatively calculated.

[0014] In conjunction with the first aspect, in the eighth embodiment of the first aspect of this application, the step of employing a weighted summation algorithm to calculate the comprehensive causal contribution of each line in the defect-associated candidate line set to the defect, and generating a set of transmission lines to be adjusted, includes: A weighted summation algorithm is used to weight and fuse the causal contribution values ​​of individual lines according to preset weight coefficients, and calculate the comprehensive causal contribution of each line in the candidate line set to the current defect row by row. A preset threshold for the comprehensive causal contribution is set, and lines whose comprehensive causal contribution exceeds the threshold are identified as defect-causing lines. Combining the comprehensive causal contribution with the defect type, the lines are divided into core causal lines and co-causal lines, and organized according to the rule of core priority and co-contribution supplementation to generate a set of transmission lines to be adjusted.

[0015] In conjunction with the first aspect, in the ninth embodiment of the first aspect of this application, the step of constructing a multi-objective optimization objective function based on defect categories, constructing a constrained multi-objective optimization equation set, verifying the rationality of power flow distribution under different parameter adjustment schemes, solving the constraint equation set, and obtaining the transmission line adjustment scheme includes: For heat-related defects, the optimization sub-objective is to minimize the line load rate and inter-line circulating current to eliminate the risk of line overload and heat generation. For insulation discharge-related defects, the optimization sub-objective is to minimize the line voltage deviation and inter-line voltage difference to eliminate the risk of line insulation breakdown and abnormal discharge. For mechanical defects, the optimization sub-objective is to minimize the line load fluctuation amplitude to stabilize the line sag and inter-line safe distance. Differentiated weight coefficients are assigned to each sub-objective according to the defect category, with the sum of the weight coefficients being 1. At the same time, the weight of the sub-objective corresponding to the core causative line is amplified to construct a multi-objective optimization objective function. A constraint system is constructed with power flow balance as the mandatory equality constraint and safe operation of lines and completion of transmission tasks as the core inequality constraints. The equality constraints follow the power flow balance rules of active and reactive power of power grid nodes, requiring that the power output difference between any node and the load in the power grid be completely matched with the power transmission between nodes. The inequality constraints include line current safety constraints, node voltage safety constraints, transmission task constraints, and transmission task constraints. Among them, the line current safety constraint means that the operating current of all lines to be adjusted must not exceed the rated current of the line, and must not be lower than zero. The node voltage safety constraint means that the operating voltage of the corresponding nodes of all lines to be adjusted must be controlled within the preset voltage allowable deviation range. The transmission task constraint means that the total active power transmission of the lines to be adjusted must match the predetermined transmission task target power, and the deviation must not exceed the preset power allowable threshold. The line loss constraint means that the operating power loss of the lines to be adjusted must not exceed the preset line loss limit. For each group of candidate transmission line adjustment schemes, the Newton-Raphson power flow calculation algorithm is used to perform power flow iterative verification. During the iteration process, the node voltage magnitude and phase angle are continuously corrected until the imbalance of active and reactive power in the power grid is less than the preset convergence threshold, and the power flow iteration is determined to be converged. When the iteration is converged and all parameters do not exceed the aforementioned safety constraint boundary, the power flow distribution of the scheme is determined to be reasonable and it is a valid scheme; otherwise, the scheme is eliminated. Initialize the particle swarm, with each particle corresponding to a set of candidate operating parameters for the line to be adjusted; calculate the optimization objective value corresponding to the effective scheme, and update the individual optimal scheme and the global optimal Pareto solution set; continuously iterate and update the operating parameters of the particles until the preset maximum number of iterations is reached, or the optimization objective converges to a stable state; from the Pareto optimal solution set, select the transmission line adjustment scheme that takes into account both the defect elimination effect and the completion of the transmission task.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention constructs a dual-branch scene adaptive detection architecture, setting up a difficult scene branch that adapts to complex occlusion and dense small defects, and a regular scene branch for general inspection, to achieve scene-differentiated detection; it introduces density regression and adaptive query mechanism, which can dynamically adjust the number of decoder queries according to the defect density of a single image, thereby improving the model's generalization ability in different power transmission inspection scenarios.

[0017] 2. This invention extracts cross-scale density features based on pixel-level density maps and combines channel attention mechanism and spatial gating unit to adaptively weight feature channels of different scales according to the distribution of defect categories, so as to achieve deep fusion of shallow detail features and deep semantic features, effectively enhance the feature expression capability of small and dense defects in transmission lines, and avoid the loss of small target defect features.

[0018] 3. This invention combines different defect categories such as heat generation, insulation discharge, and mechanical defects to construct differentiated multi-objective optimization functions, and establishes a complete system of power grid flow equality constraints and equipment safety operation inequality constraints. At the same time, it verifies the rationality of each parameter adjustment scheme through power grid flow iteration, eliminates invalid schemes that exceed limits, effectively solves the problems of weak targeting and poor adaptability of traditional optimization schemes, and avoids the risk of power grid flow exceeding limits. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the steps of a defect detection method for unmanned aerial vehicle (UAV) power transmission line inspection according to the present invention; Figure 2 This is a schematic diagram illustrating the model training and application steps of a defect detection method for unmanned aerial vehicle (UAV) power line inspection according to the present invention. Detailed Implementation

[0020] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example: Figures 1-2 As shown, the present invention provides a technical solution: like Figure 1 As shown, this application provides a defect detection method for unmanned aerial vehicle (UAV) power line inspection, including the following steps: Step S100: Use a drone to sample images of defective transmission lines and generate annotation files; perform data augmentation on the annotation files, input them into a backbone network with pre-trained weights, and extract multi-scale features; Specifically, the image size is adjusted, and random photometric distortion, random resizing, random IoU cropping, random horizontal flipping, and mosaic data enhancement are applied to the image. The image format is then converted to Tensor format. A backbone network with pre-trained dinov3 weights is used to extract multi-scale features from the image. The backbone network uses ConvNext v2.

[0022] In one specific embodiment, a drone equipped with a 20-megapixel visible light gimbal camera was used to collect a total of 12,000 images of a 500kV transmission line corridor inspection, covering four typical defects: insulator damage, conductor strand breakage, hardware corrosion, and bolt loosening. Corresponding VOC format annotation files were generated, including 4,500 images of difficult scenes with dense small defects and background occlusion, and 7,500 images of regular inspection scenes.

[0023] All images were resized to 1333×800 pixels. Random photometric distortion, random downsizing, random IoU cropping, random horizontal flipping, and mosaic data augmentation were performed on the training set images. The brightness and contrast perturbation range for photometric distortion was set to ±30%, the cross-union threshold for random IoU cropping was set to 0.3, and mosaic augmentation used a 4-image stitching mode. After augmentation, the images were converted to Tensor format and divided into training, validation, and test sets in an 8:1:1 ratio.

[0024] In the feature extraction stage, ConvNext v2-Base was used as the backbone network, loaded with dinov3 weights pre-trained on the ImageNet-22K dataset. The weights of the first two stages of the network were frozen for fine-tuning. The training batch size was set to 16, the optimizer was AdamW, the initial learning rate was set to 1e-4, the weight decay coefficient was set to 1e-5, and the training iterations were 30 rounds. After processing by the backbone network, multi-scale feature maps at four different depth stages were output, with corresponding channel dimensions of 128, 256, 512, and 1024, respectively. On the test set, the defect feature extraction recall rate of this backbone network reached 98.2%, and the accuracy of small defect feature extraction was improved by 12.7 percentage points compared with the network of the same architecture without pre-trained weights.

[0025] like Figure 2 As shown, step S200: input the multi-scale features into the first branch and the second branch respectively; construct a joint loss function and update the network parameters through error backpropagation; iterate training until the network converges to obtain the scene adaptive defect detection model; during the inspection process, obtain the geographical location identifier of the current location, determine the scene category of the current section, and output the defect location and defect category through the scene adaptive defect detection model; Specifically, the first branch is the difficult scenario branch, and the second branch is the normal scenario branch; The first branch constructs a multi-scale feature pyramid to complete multi-scale hybrid encoding, generates a pixel-level density map, calculates the batch adaptive query number, and outputs the defect detection result; the second branch constructs a lightweight multi-scale feature pyramid, inputs it into a lightweight decoder to complete decoding inference with a fixed number of queries, and outputs the defect detection result.

[0026] Furthermore, multi-scale feature maps are extracted from different depth stages of the first branch, and after unifying the channel dimensions through a convolutional module, a multi-scale feature fusion encoder is constructed for the first branch to perform multi-scale hybrid encoding. The multi-scale feature fusion encoder outputs a first feature pyramid containing five levels. Each layer of features has different channel dimensions and spatial resolution, among which, This is a shallow, native feature layer with medium to low resolution. This is a medium-to-high resolution shallow native feature layer. This is the highest resolution shallow native feature layer. This is a mesoscale deep fusion feature layer. For large-scale deep fusion feature layers; Based on the first feature pyramid, the features of each layer are mapped to a unified hidden dimension H through the input projection layer of the feature fusion encoder of the first branch, resulting in a projected feature sequence. Each projected feature in the projected feature sequence is flattened along the spatial dimension and stitched together to construct the encoder output sequence. Where R represents the real number field, B is the batch size, and i is the index of the first feature pyramid level. Let be the spatial height of the feature map of the i-th layer. Let be the spatial width of the feature map of the i-th layer; record the spatial shape of the features of each layer. In M The corresponding sequence is input to the density regression module, which uses an encoder and decoder structure to generate a pixel-level density map and obtains a continuous estimate of the number of image instances through integration. The continuous estimate is then converted into an adaptive query number through a nonlinear mapping function. Based on the continuous estimate, a regression head is used to map the density value to a continuous regression value, and the adaptive query quantity for each batch of samples is calculated. ,in This represents the number of queries from the first sample to the Bth sample in the current batch. The maximum number of queries in this batch is determined as the unified number of queries for the current batch. Based on the pixel-level density map, a multi-scale density feature is generated using a multi-scale feature extractor. The multi-scale density feature and M are then input into a multi-scale density feature enhancement module. This module includes a channel attention mechanism and a spatial gating unit, which adaptively weights feature channels at different scales according to the category distribution, and outputs the enhanced memory feature. ; Will Instead of M, input the Transformer decoder and detector head of the RT-DETR network, and set the number of queries generated by the decoder to the uniform number of queries for the current batch; generate the bounding box coordinate regression results and class regression results of the first branch through the decoder; Multi-scale feature maps are extracted from different depth stages of the backbone network. After unifying the channel dimensions through a convolutional module, a multi-scale feature fusion encoder is constructed for the second branch to perform multi-scale hybrid encoding, resulting in a second feature pyramid with four levels output by the encoder. ; The second feature pyramid is input into a lightweight decoder without feature map augmentation, and the number of queries generated by the decoder is set to a fixed value; the decoder generates the predicted bounding box coordinate regression results and category regression results of the second branch.

[0027] Furthermore, through the loss function Calculate the class regression loss for the first branch and the second branch separately, using the following formula: ; Where N is the number of samples, and C is the number of classification categories in the dataset. One-hot encoding of the real label. This refers to the results of category regression, specifically the predicted samples. The probability of belonging to a type C defect; Loss function based on coordinate regression Calculate the coordinate regression loss for each predicted bounding box in the first and second branches separately, using the following formula: ; in, Let x be the x-coordinate of the center point of the true bounding box of the k-th defect target. Let y be the center point of the true bounding box of the k-th defect target. Let be the width of the true bounding box of the k-th defect target. Let the height be the actual bounding box height of the k-th defect target. The x-coordinate of the center point of the k-th defect target prediction box output by the model. The ordinate of the center point of the k-th defect target prediction box output by the model. Let be the width of the predicted bounding box for the k-th defect target output by the model. The height of the predicted bounding box for the k-th defect target output by the model; Calculate the generalized intersection-union ratio (GUU) loss for the first and second branches respectively. The smaller the loss value, the higher the overlap between the predicted box and the true box, the closer the position, and the more accurate the defect localization. Calculate the count regression loss of the first branch. The formula is: ; Where m is the total number of image samples in the current batch. For the first The total number of real defects in the images. For the first Total number of images showing predicted defects; Calculate the two-branch consistency loss The formula is: ; in, The feature vector output by the first branch. This is the feature vector output by the second branch; The total loss is obtained by weighting and summing the category regression loss, coordinate regression loss, and generalized intersection-union loss of the two branches with the count regression loss and bi-branch consistency loss of the first branch. The total loss is then input into the optimizer, and the network parameters are optimized through error backpropagation.

[0028] Furthermore, a transmission line segment difficulty database is pre-established, including difficult scenario segment identifiers and normal scenario segment identifiers. The trained model weights are exported into the ONNX open neural network exchange format, and after operator fusion and inter-layer optimization using the TensorRT inference engine, the model is deployed on the UAV platform. Before inspection, the UAV loads the transmission line segment difficulty database and determines the segment type of the current shooting location in real time based on GPS positioning or preset waypoint information. If the current segment is marked as a difficult scenario, the first branch is activated for detection; if the current segment is marked as a normal scenario, the second branch is activated for detection.

[0029] In one specific embodiment, the multi-scale feature maps output from the backbone network are input into a dual-branch parallel detection architecture. The first branch is adapted for difficult scenarios with dense small defects and background occlusion, while the second branch is adapted for conventional scenarios with open plains. For the first branch, multi-scale feature maps are extracted from the four depth stages of the backbone network. These maps are then processed by a 3×3 convolution module to unify the channel dimension to 256, constructing a multi-scale feature fusion encoder. The output is a first feature pyramid with five levels, each with a feature space resolution of 1 / 32, 1 / 16, 1 / 8, 1 / 16, and 1 / 32, corresponding to feature map sizes of 42×25, 84×50, 167×100, 84×50, and 42×25. The features at each level are mapped to a unified hidden dimension of 256 through an input projection layer. After flattening and stitching, an encoder output sequence with a dimension of 16×43750×256 is constructed, simultaneously recording the feature space shape of each layer. The sequence corresponding to the highest resolution feature layer in the encoder output sequence is input into the density regression module to generate a pixel-level density map. After integration, a continuous estimate of the number of defect instances is obtained. This estimate is then converted into an adaptive query number using the ReLU nonlinear mapping function. The regression head calculates the adaptive query number for a single image within a batch, ranging from 10 to 300. The maximum query number within the batch is taken as the unified query number. In this embodiment, the batch size is 16, and the maximum unified query number per batch is 286. Based on the pixel-level density map, a three-layer multi-scale convolutional extractor generates cross-scale density features. These features, along with the encoder output sequence, are input into a cross-scale density feature enhancement module with a built-in channel attention mechanism and spatial gating unit. After adaptive weighting of the feature channels, enhanced memory features are output, replacing the original sequence input into the Transformer decoder and detector head of the RT-DETR network. Inference is completed by matching the corresponding batch unified query number, and the defect prediction box coordinates and category regression results of the first branch are output. For the second branch, multi-scale feature maps are extracted from the backbone network, and after the channel dimension is unified to 256 by the convolution module, a multi-scale feature fusion encoder is constructed. The output is a second feature pyramid containing four levels. The input is a lightweight decoder without feature enhancement. The number of decoder queries is fixed at 100. After inference is completed, the detection result of the second branch is output.

[0030] During training, a multi-dimensional joint loss function was constructed to calculate the class regression loss, L1 coordinate regression loss, and generalized intersection-union (OCU) loss for each branch. The counting regression loss was calculated only for the first branch. Simultaneously, the L2 consistency loss of the output feature vectors of both branches was calculated. The weights of each loss term were set as follows: class loss 1.0, coordinate regression loss 5.0, OCU loss 2.0, counting regression loss 0.1, and consistency loss 0.5. The optimizer used was AdamW with an initial learning rate of 1e-4, accompanied by a cosine annealing learning rate decay strategy. The total training epochs were 30, with the first 5 epochs serving as a warm-up phase. During the warm-up, the learning rate linearly increased from 1e-6 to 1e-4. Training was terminated early when the validation set loss did not decrease for three consecutive epochs. The model finally converged on the 27th epoch.

[0031] A section difficulty database covering the entire 500kV transmission line was pre-established, marking 28 sections with challenging scenarios such as mountain crossings and dense towers, and 76 sections with routine inspections in plains areas. The converged model weights were exported in ONNX format, and after operator fusion and inter-layer optimization were completed using the TensorRT 8.6 inference engine, they were deployed to an UAV-based edge computing platform. In terms of inference time per image, the second branch in a routine scenario took only 8ms, and the first branch in a challenging scenario took 22ms. Compared with the single-branch RT-DETR model, the average inference speed was improved by 47%. On the independent test set, the scene-adaptive defect detection model of this embodiment achieved an average accuracy (mAP@0.5) of 97.6% for defect detection in normal scenes and 95.3% for dense small defect detection in difficult scenes. Compared with the traditional single fixed structure YOLOv8 model, the accuracy of small defect detection in difficult scenes was improved by 18.2 percentage points, and the false negative rate was reduced from 12.4% to 2.1%. Compared with the native RT-DETR model, the overall mAP@0.5 was improved by 6.8 percentage points, and the confidence fluctuation of the detection results during scene switching was reduced from 15.3% to 3.7%, demonstrating excellent scene generalization ability and detection stability.

[0032] Step S300: Based on the defect location, associate and map the defect with all cross-section transmission lines to generate a defect-associated candidate line set; construct a transmission line fault mechanism library, match the causal association dimension of the current defect category, and quantify the causal contribution. Specifically, the data of the transmission line ledger is obtained, and the three-dimensional spatial topology construction algorithm and the nearest neighbor matching algorithm are used to delineate the three-dimensional spatial ownership boundary of the core components of each transmission line in the crossing section, mark the line attributes, multi-line shared attributes of each component, as well as the physical adjacency relationship and electrical coupling relationship between the lines, restore the spatial layout and ownership boundary of all transmission lines in the crossing section, and generate a three-dimensional spatial topology model of the transmission lines in the crossing section. Based on the aforementioned three-dimensional spatial topology model, with the defect three-dimensional spatial coverage area as the core, a three-dimensional spatial distance matching algorithm and a spatial intersection verification algorithm are used to calculate the shortest spatial distance between the defect three-dimensional spatial coverage area and the three-dimensional spatial boundary of each transmission line core component in the topology model, and to verify the spatial intersection relationship; a preset spatial association judgment threshold is set to filter out transmission lines whose shortest spatial distance is less than the preset threshold or have spatial intersection. Based on the defect category, an ownership verification algorithm and an electrical coupling degree analysis algorithm are used. When a defect is physically attached to a component of a single transmission line, that line is marked as a core associated line. When a defect is physically attached to a common component of multiple transmission lines, all transmission lines associated with that common component are marked as core associated lines. When the defect is an electrical defect and is located within the electromagnetic coupling zone crossed by the transmission lines, all transmission lines forming that crossing zone and having electrical coupling associations are included in the core associated range. When the defect is a mechanical defect, only transmission lines with physical attachment associations are retained. A set of candidate lines for defect associations is generated.

[0033] Furthermore, the fault occurrence mechanisms of all types of defects in transmission lines are sorted out, and independent entries are divided according to defect categories. Each entry clarifies the corresponding defect's cause-related dimensions, the mapping rules between the cause and the line operating parameters, and the cause's influence weight benchmark, thus forming a transmission line fault mechanism library. Based on the current defect category, the fault mechanism database of transmission lines is retrieved, and the corresponding causal association dimension is obtained. Based on the causal association dimension, a weight allocation algorithm is used to determine the influence weight of each dimension. Combined with the standardized operating data of each line in the defect association candidate line set, the causal contribution of each line to the current defect is quantitatively calculated.

[0034] In one specific embodiment, standardized ledger data of four transmission lines within the crossing section is acquired, covering the three-dimensional spatial coordinates of towers and conductors in the WGS-84 coordinate system, the line ownership information of core components such as hardware or insulators, and the electrical coupling relationship data between lines. A three-dimensional spatial topology construction algorithm and a nearest neighbor matching algorithm are used to define a ±0.3m three-dimensional spatial ownership boundary for the core components of each line, marking the line-specific / shared attributes of the components, the physical adjacency and electrical coupling relationships between lines, reconstructing the spatial layout and ownership boundaries of the lines in the crossing section, and generating a three-dimensional spatial topology model. Using the detected defect three-dimensional spatial coverage area as the core, a three-dimensional spatial distance matching algorithm and a spatial intersection verification algorithm are used to calculate the shortest spatial distance between the defect area and the ownership boundary of each line component. A preset spatial association judgment threshold of 0.5m is set, and lines with distances less than the threshold or with spatial intersection are selected for initial screening. Verification was conducted based on defect categories. For the detected 500kV I-line conductor overheating defect, which is physically attached to the conductor components of the line, it was marked as a core associated line. For the corona discharge electrical defect at the crossing point, which is located within the electromagnetic coupling range of the 500kV and 220kV lines, both associated lines were included in the core associated range. For the mechanical hardware corrosion defect, only lines with physical attachment associations were retained. Finally, a set of candidate lines for defect association was generated. In this embodiment, the accuracy of defect-line association mapping reached 99.2%, and the false association rate was as low as 0.7%.

[0035] The fault mechanisms of four typical defects in transmission lines were pre-analyzed and divided into independent entries according to defect category. Each entry clearly defines the corresponding causal correlation dimensions, the mapping rules between the cause and line operating parameters, and the causal influence weight benchmark, thus constructing a standardized transmission line fault mechanism library. The causal correlation dimensions for conductor overheating defects are line load rate, inter-line circulating current, and contact resistance; for insulator damage defects, voltage deviation, lightning overvoltage, and surface contamination; and for hardware corrosion defects, environmental humidity, line leakage current, and service life. Based on the currently detected conductor overheating defect categories, the fault mechanism library was searched to obtain the corresponding three causal correlation dimensions. The analytic hierarchy process (AHP) was used to assign weights of 0.6, 0.3, and 0.1 to each dimension. Combining real-time and historical operating data of each line in the defect correlation candidate line set, the comprehensive causal contribution of the 500kV I loop to this overheating defect was calculated to be 92.3%, and the collaborative causal contribution of adjacent loops on the same tower was 7.7%.

[0036] Step S400: Using a weighted summation algorithm, calculate the comprehensive causal contribution of each line in the defect-associated candidate line set to the defect, and generate a set of transmission lines to be adjusted; combine the defect categories to construct a multi-objective optimization objective function, construct a constrained multi-objective optimization equation set, verify the rationality of the power flow distribution under different parameter adjustment schemes, solve the constraint equation set, and obtain the transmission line adjustment scheme.

[0037] Specifically, a weighted summation algorithm is used to weight and fuse the causal contribution values ​​of individual lines according to preset weight coefficients, and calculate the comprehensive causal contribution of each line in the candidate line set to the current defect row by row; a preset threshold for the comprehensive causal contribution is set, and lines whose comprehensive causal contribution exceeds the threshold are selected and identified as defect-causing lines; combining the comprehensive causal contribution and the defect type, the lines are divided into core causal lines and co-causal lines, and organized according to the rule of core priority and co-supplementation to generate a set of transmission lines to be adjusted.

[0038] Furthermore, for heat-related defects, the optimization sub-objective is to minimize the line load rate and inter-line circulating current to eliminate the risk of line overload and heat generation; for insulation discharge defects, the optimization sub-objective is to minimize the line voltage deviation and inter-line voltage difference to eliminate the risk of line insulation breakdown and abnormal discharge; for mechanical defects, the optimization sub-objective is to minimize the line load fluctuation amplitude to stabilize the line sag and inter-line safe distance; differentiated weight coefficients are assigned to each sub-objective according to the defect category, the sum of the weight coefficients is 1, and the weight of the sub-objective corresponding to the core causative line is amplified to construct a multi-objective optimization objective function; A constraint system is constructed with power flow balance as the mandatory equality constraint and safe operation of lines and completion of transmission tasks as the core inequality constraints. The equality constraints follow the power flow balance rules of active and reactive power of power grid nodes, requiring that the power output difference between any node and the load in the power grid be completely matched with the power transmission between nodes. The inequality constraints include line current safety constraints, node voltage safety constraints, transmission task constraints, and transmission task constraints. Among them, the line current safety constraint means that the operating current of all lines to be adjusted must not exceed the rated current of the line, and must not be lower than zero. The node voltage safety constraint means that the operating voltage of the corresponding nodes of all lines to be adjusted must be controlled within the preset voltage allowable deviation range. The transmission task constraint means that the total active power transmission of the lines to be adjusted must match the predetermined transmission task target power, and the deviation must not exceed the preset power allowable threshold. The line loss constraint means that the operating power loss of the lines to be adjusted must not exceed the preset line loss limit. For each group of candidate transmission line adjustment schemes, the Newton-Raphson power flow calculation algorithm is used to perform power flow iterative verification. During the iteration process, the node voltage magnitude and phase angle are continuously corrected until the imbalance of active and reactive power in the power grid is less than the preset convergence threshold, and the power flow iteration is determined to be converged. When the iteration is converged and all parameters do not exceed the aforementioned safety constraint boundary, the power flow distribution of the scheme is determined to be reasonable and it is a valid scheme; otherwise, the scheme is eliminated. Initialize the particle swarm, with each particle corresponding to a set of candidate operating parameters for the line to be adjusted; calculate the optimization objective value corresponding to the effective scheme, and update the individual optimal scheme and the global optimal Pareto solution set; continuously iterate and update the operating parameters of the particles until the preset maximum number of iterations is reached, or the optimization objective converges to a stable state; from the Pareto optimal solution set, select the transmission line adjustment scheme that takes into account both the defect elimination effect and the completion of the transmission task.

[0039] In one specific embodiment, a weighted summation algorithm is used to weight and fuse the independent and collaborative causal contribution values ​​of each line in the defect-associated candidate line set with preset weighting coefficients of 0.8 and 0.2, respectively. The comprehensive causal contribution of the 500kV I loop to the heating defect is calculated line by line, resulting in a contribution of 93.1% and 6.9% for adjacent II loops on the same tower. A preset threshold of 10% for the comprehensive causal contribution is set. 500kV I loops with contributions exceeding the threshold are selected as core causal lines and organized according to the core priority rule to generate a set of transmission lines to be adjusted containing only these lines.

[0040] To address the conductor overheating defect, a multi-objective optimization function was constructed. The optimization sub-objectives were minimizing the line load rate and inter-line circulating current. Weight coefficients of 0.7 and 0.3 were assigned to the two sub-objectives, with a sum of 1. Simultaneously, the weight of the sub-objective corresponding to the core causative line was increased to 1.2 times. An inequality constraint system was constructed, using the active and reactive power flow balance of the grid nodes as a mandatory equality constraint, and the safe operation of the lines and the completion of transmission tasks as the core. The rated current of the line was 1250A, the allowable deviation range of the node voltage was ±5%, the allowable deviation threshold of the transmission power was 2%, and the line power loss limit was 120kW / km.

[0041] For each set of candidate line operation parameter adjustment schemes, the Newton-Raphson power flow calculation algorithm is used to perform iterative verification of power flow. A power flow convergence threshold of 1e-6 is set. Schemes that converge iteratively and whose parameters are unconstrained are deemed valid; otherwise, they are directly eliminated. A multi-objective particle swarm optimization algorithm is used to solve the constraint equations. The particle swarm size is set to 50, and the maximum number of iterations is 100. After iteration until the optimization objective converges, the optimal scheme that balances defect elimination and transmission task completion is selected from the Pareto optimal solution set.

[0042] Experimental verification shows that after implementing the line adjustment scheme obtained in this embodiment, the operating load rate of the 500kV I loop line decreased from 87% before optimization to 62%, the inter-line circulating current decreased from 112A to 18A, and the conductor operating temperature decreased from 78℃ to 42℃, completely eliminating the risk of overload heating defects. After optimization, the power flow converged after 3 iterations, the transmission power deviation was only 0.8%, meeting the established transmission task requirements. The line power loss was reduced by 18.7% compared to before optimization, and there were no current or voltage over-limit situations. Compared with the traditional single-criterion optimization scheme, the defect cause elimination efficiency was improved by 62.5%, and the stability of power grid operation was significantly improved.

[0043] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A defect detection method for unmanned aerial vehicle (UAV) power transmission line inspection, characterized in that, Includes the following steps: UAVs are used to sample images of defective power transmission lines and generate annotation files; Data augmentation is performed on the labeled files, which are then input into a backbone network with pre-trained weights to extract multi-scale features; The multi-scale features are input into the first branch and the second branch respectively; a joint loss function is constructed and the network parameters are updated through error backpropagation; the network is iteratively trained until it converges to obtain a scene-adaptive defect detection model; during the inspection process, the geographic location identifier of the current location is obtained, the scene category of the current section is determined, and the defect location and defect category are output through the scene-adaptive defect detection model. For the first branch, multi-scale feature maps are extracted from the backbone network. After unifying the channel dimensions, a multi-scale feature fusion encoder is constructed, outputting a first feature pyramid with five levels. The features of each level are mapped to a unified hidden dimension through the input projection layer, and after flattening and stitching, the encoder output sequence is obtained. The highest resolution feature layer in the encoder output sequence is input into the density regression module to generate a pixel-level density map and calculate the batch adaptive unified query number. Cross-scale density features are extracted based on the density map and input into the feature enhancement module along with the encoder output sequence. After adaptive weighting of the feature channels by the channel attention and spatial gating unit, the enhanced memory features are output. The enhanced memory features are input into the Transformer decoder and detector head of the RT-DETR network, and inference is completed by matching adaptive query numbers. The defect prediction box coordinates and category regression results are output. For the second branch, multi-scale feature maps are extracted from the backbone network, and a multi-scale feature fusion encoder is constructed after unifying the channel dimensions to output a second feature pyramid with four levels. Input it into a lightweight decoder without feature enhancement, complete the inference with a fixed number of queries, and output the defect prediction box coordinates and category regression results; Based on the location of the defect, the defect is associated with all the transmission lines in the crossing section to generate a set of candidate lines associated with the defect; a transmission line fault mechanism library is constructed to match the causal association dimension of the current defect category and quantify the causal contribution. A weighted summation algorithm is used to calculate the comprehensive causal contribution of each line in the defect-associated candidate line set to the defect, thereby generating a set of transmission lines to be adjusted. By combining the defect categories, a multi-objective optimization objective function is constructed, and a constrained multi-objective optimization equation system is built. The rationality of the power flow distribution under different parameter adjustment schemes is verified, and the constrained equation system is solved to obtain the transmission line adjustment scheme.

2. The defect detection method for UAV power transmission line inspection according to claim 1, characterized in that, The process of data augmentation of the labeled files involves inputting them into a backbone network equipped with pre-trained weights and extracting multi-scale features, including: The image size is adjusted, and random photometric distortion, random resizing, random IoU cropping, random horizontal flipping, and mosaic data enhancement are applied to the image. The image format is then converted to Tensor format. Multi-scale features are extracted from the image using a backbone network with pre-trained dinov3 weights. The backbone network uses ConvNext v2.

3. The defect detection method for UAV power transmission line inspection according to claim 1, characterized in that, The step of inputting the multi-scale features into the first branch and the second branch respectively includes: The first branch is the difficult scenario branch, and the second branch is the normal scenario branch; The first branch constructs a multi-scale feature pyramid to complete multi-scale hybrid encoding, generates a pixel-level density map, calculates the batch adaptive query number, and outputs the defect detection result; the second branch constructs a lightweight multi-scale feature pyramid, inputs it into a lightweight decoder to complete decoding inference with a fixed number of queries, and outputs the defect detection result.

4. A defect detection method for UAV power transmission line inspection according to claim 1, characterized in that, The construction of the joint loss function and the updating of network parameters through error backpropagation include: The class regression loss, coordinate regression loss, and generalized intersection-union (CIU) loss for each branch are calculated separately. The class regression loss is calculated based on the one-hot encoding of the defect's true label and the predicted class probability. The coordinate regression loss is calculated based on the center point coordinates and width and height parameters of the defect's true and predicted bounding boxes. The CIU loss is used to constrain the overlap between the predicted and true bounding boxes, improving defect localization accuracy. The count regression loss is calculated only for the first branch, based on the total number of true defects in the image and the total number predicted by the model. The consistency loss of the output feature vectors of both branches is calculated to constrain the alignment of the feature spaces of the two branches. The total loss is obtained by weighted summing of all loss terms and input into the optimizer, where it is then used to optimize the network parameters through error backpropagation.

5. A defect detection method for unmanned aerial vehicle (UAV) power line inspection according to claim 1, characterized in that, During the inspection process, the geographical location identifier of the current location is obtained, the scene category of the current section is determined, and the defect location and defect category are output through the scene adaptive defect detection model, including: A transmission line segment difficulty database is pre-established, including difficult scenario segment identifiers and normal scenario segment identifiers. The trained model weights are exported into the ONNX open neural network exchange format. After operator fusion and inter-layer optimization using the TensorRT inference engine, the model is deployed on a UAV platform. Before inspection, the UAV loads the transmission line segment difficulty database and determines the segment type of the current shooting location in real time based on GPS positioning or preset waypoint information. If the current segment is marked as a difficult scenario, the first branch is activated for detection; if the current segment is marked as a normal scenario, the second branch is activated for detection.

6. A defect detection method for unmanned aerial vehicle (UAV) power line inspection according to claim 1, characterized in that, The process of associating defects with all crossing sections of transmission lines based on their locations to generate a set of candidate lines for defect association includes: The data of the transmission line ledger is obtained, and the three-dimensional spatial topology construction algorithm and the nearest neighbor matching algorithm are used to delineate the three-dimensional spatial ownership boundary of the core components of each transmission line in the crossing section, and generate a three-dimensional spatial topology model of the transmission line in the crossing section. Based on the aforementioned three-dimensional spatial topology model, with the defect three-dimensional spatial coverage area as the core, a three-dimensional spatial distance matching algorithm and a spatial intersection verification algorithm are used to calculate the shortest spatial distance between the defect three-dimensional spatial coverage area and the three-dimensional spatial boundary of each transmission line core component in the topology model, and to verify the spatial intersection relationship; a preset spatial association judgment threshold is set to filter out transmission lines whose shortest spatial distance is less than the preset threshold or have spatial intersection. Based on the defect category, an ownership verification algorithm and an electrical coupling degree analysis algorithm are used. When a defect is physically attached to a component of a single transmission line, that line is marked as a core associated line. When a defect is physically attached to a common component of multiple transmission lines, all transmission lines associated with that common component are marked as core associated lines. When the defect is an electrical defect and is located within the electromagnetic coupling zone crossed by the transmission lines, all transmission lines forming that crossing zone and having electrical coupling associations are included in the core associated range. When the defect is a mechanical defect, only transmission lines with physical attachment associations are retained. A set of candidate lines for defect associations is generated.

7. A defect detection method for unmanned aerial vehicle (UAV) power line inspection according to claim 1, characterized in that, The construction of the transmission line fault mechanism library, matching the causal correlation dimensions of the current defect category, and quantifying the causal contribution include: Organize the fault occurrence mechanisms of all types of defects in transmission lines, divide them into independent entries according to defect categories, and form a fault mechanism library for transmission lines; Based on the current defect category, the fault mechanism database of transmission lines is retrieved, and the corresponding causal association dimension is obtained. Based on the causal association dimension, a weight allocation algorithm is used to determine the influence weight of each dimension. Combined with the standardized operating data of each line in the defect association candidate line set, the causal contribution of each line to the current defect is quantitatively calculated.

8. A defect detection method for UAV power transmission line inspection according to claim 1, characterized in that, The weighted summation algorithm is used to calculate the comprehensive causal contribution of each line in the defect-associated candidate line set to the defect, generating a set of transmission lines to be adjusted, including: A weighted summation algorithm is used to weight and fuse the causal contribution values ​​of individual lines according to preset weight coefficients, and calculate the comprehensive causal contribution of each line in the candidate line set to the current defect row by row. A preset threshold for the comprehensive causal contribution is set, and lines whose comprehensive causal contribution exceeds the threshold are identified as defect-causing lines. Combining the comprehensive causal contribution with the defect type, the lines are divided into core causal lines and co-causal lines, and organized according to the rule of core priority and co-contribution supplementation to generate a set of transmission lines to be adjusted.

9. A defect detection method for unmanned aerial vehicle (UAV) power line inspection according to claim 1, characterized in that, The objective function for multi-objective optimization is constructed by combining defect categories, a system of constrained multi-objective optimization equations is constructed, the rationality of power flow distribution under different parameter adjustment schemes is verified, the constrained equations are solved, and the transmission line adjustment scheme is obtained, including: For heat-related defects, the optimization sub-objective is to minimize the line load rate and inter-line circulating current; for insulation discharge defects, the optimization sub-objective is to minimize the line voltage deviation and inter-line voltage difference; for mechanical defects, the optimization sub-objective is to minimize the line load fluctuation amplitude. Differentiated weight coefficients are assigned to each sub-objective according to the defect category, and the sum of the weight coefficients is 1. The weight of the sub-objective corresponding to the core causative line is amplified to construct a multi-objective optimization objective function. A constraint system is constructed with power flow balance as the mandatory equality constraint and safe operation of lines and completion of transmission tasks as the core inequality constraints. The equality constraints follow the power flow balance rules of active and reactive power at power grid nodes, requiring that the power output difference between any node and the load in the power grid be perfectly matched with the power transmission between nodes. The inequality constraints include line current safety constraints, node voltage safety constraints, transmission task constraints, and transmission task constraints. For each group of candidate transmission line adjustment schemes, the Newton-Raphson power flow calculation algorithm is used to perform power flow iterative verification. During the iteration process, the node voltage magnitude and phase angle are continuously corrected until the imbalance of active and reactive power in the power grid is less than the preset convergence threshold, and the power flow iteration is determined to be converged. When the iteration is converged and all parameters do not exceed the aforementioned safety constraint boundary, the power flow distribution of the scheme is determined to be reasonable and it is a valid scheme; otherwise, the scheme is eliminated. Initialize the particle swarm, with each particle corresponding to a set of effective solutions; calculate the optimization objective value corresponding to the effective solutions, and update the individual optimal solutions and the optimal Pareto solution set; continuously iterate and update the running parameters of the particles until the preset maximum number of iterations is reached, or the optimization objective converges to a stable state; from the optimal Pareto solution set, select effective solutions that take into account both the defect elimination effect and the completion of the power transmission task as the power transmission line adjustment scheme.

Images