Power transmission line icing detection method, system, device and medium

Abstract

The present application belongs to the technical field of transmission line icing detection, and specifically discloses a transmission line icing detection method, system, device and medium, which comprises the following steps: acquiring a target line image; conducting conductor detection on the target line image to generate a first conductor candidate area; wherein the conductor detection is based on at least the inherent geometric characteristics of the transmission conductor; conducting conductor detection on the target line image by using a pre-trained deep learning model to generate a second conductor candidate area; generating a final conductor candidate area based on the overlapping relationship between the first conductor candidate area and the second conductor candidate area; and determining an icing grade classification result according to the final conductor candidate area. The present application can improve the conductor positioning accuracy in a complex background and improve the robustness and generalization ability of the system in a complex environment.

Description

Technical Field

[0001] This invention belongs to the field of power transmission line icing detection technology, and particularly relates to a method, system, equipment and medium for power transmission line icing detection. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Icing on transmission lines is one of the main threats to the safe operation of the power grid in winter. It not only adds extra load to the transmission lines, but may also trigger a series of chain accidents such as conductor galloping, insulator flashover, line breakage and tower collapse, bringing huge risks and economic losses to the operation of the power system.

[0004] Currently, most methods for detecting icing on power transmission lines employ deep learning-based object detection and instance segmentation, directly learning and detecting icing areas from images of power transmission lines by constructing neural network models. However, when there is significant background interference (e.g., complex mountainous terrain, dense vegetation, heavy snowfall, fog, variable lighting, camera shake, etc.), the accuracy of recognition may be affected. For example, when the conductor overlaps with or is similar in color to background elements such as tree branches, hillsides, or buildings, existing technologies struggle to accurately distinguish conductor boundaries, easily misclassifying non-conductor areas as icing areas. Alternatively, when the texture and color of the icing on the conductor are highly similar to the background environment, the presence of ice may be completely undetectable, leading to misjudgments or missed detections of icing on power transmission lines. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method, system, device, and medium for detecting icing on transmission lines. It decouples the tasks of conductor positioning and icing type identification, constructs a cascaded network architecture for conductor positioning and icing detection, and employs a complementary approach of geometric prior and deep model for conductor positioning, thereby improving the accuracy of conductor positioning and icing type identification.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of the present invention provides a method for detecting icing on power transmission lines, comprising: Acquire the target line image; Conductor detection is performed on the target line image to generate a first conductor candidate region; wherein, the conductor detection is based at least on the inherent geometric characteristics of the transmission conductor; The target line image is used to detect conductors using a pre-trained deep learning model to generate a second conductor candidate region; Based on the overlap between the first and second candidate conductor regions, the final candidate conductor regions are generated. Based on the final candidate regions for the conductor, the icing level classification result is determined.

[0007] This invention utilizes the inherent geometric characteristics (long, straight, continuous, parallel) of power transmission lines to construct computable geometric constraints, enabling explicit segmentation of candidate regions of the transmission lines without relying on a large number of annotations. Then, a deep learning model is used to perform refined icing detection and classification on the segmented regions. Geometric priors provide structured and physically interpretable coarse localization, while deep learning handles difficult examples with blurred textures and low contrast. The two work together to improve the robustness and generalization ability of the system in complex environments.

[0008] This invention decouples wire positioning and icing classification into two cascaded independent models. The positioning model focuses on geometric boundary positioning, while the classification model focuses on texture and shape recognition. The two models do not interfere with each other and are optimized independently. They can specifically handle the detailed differences of various types of icing, effectively resolve the feature requirement conflicts between different tasks, and significantly improve the accuracy of icing recognition in complex backgrounds.

[0009] As a further approach, conductor inspection is performed based on the inherent geometric characteristics of the transmission conductors, specifically as follows: Extract linear features from the target line image; Based on length and / or angle features, linear features representing conductors are selected from the linear features to generate the first conductor candidate region.

[0010] As a further approach, linear features are extracted from the target line image, specifically: Based on the preprocessed image, linear features in the image are extracted. Specifically, edge feature maps are extracted using an edge detection algorithm, and then all line segments are extracted from the edge feature maps using cumulative probability Hough transform. The endpoint coordinates, length, and angle of each line segment are recorded.

[0011] As a further approach, based on length and / or angle features, linear features representing conductors are selected from the linear features to generate the first conductor candidate region, specifically: Remove straight line segments whose length is less than a preset first threshold; Calculate the angle histogram of all remaining line segments, select line segments within a preset angle range as candidate line segments, and construct a candidate line segment set; If, in the candidate line segment set, there exists a distance between the endpoints of two line segments that is less than a preset second threshold, and the angle between the extensions of the two line segments is less than a preset angle, then the two line segments are spliced ​​into one line segment. All the line segments ultimately formed in the candidate line segment set are expanded outward along their normal direction by a set pixel width to generate multiple rectangular wire candidate boxes, which serve as the first wire candidate region.

[0012] As a further approach, a pre-trained deep learning model is used to detect conductors in the target line image, outputting conductor candidate boxes and corresponding confidence scores to generate a second conductor candidate region.

[0013] Alternatively, a pre-trained deep learning model can be used to detect conductors in the target line image, and output conductor candidate boxes, corresponding confidence scores, and icing level classification results for each conductor candidate box. Based on the confidence level of the candidate wire frames and the spatial overlap of the candidate wire frames under the same icing level category, the candidate wire frames are filtered to generate a second candidate wire region.

[0014] This invention performs redundancy processing on the output of the ice-covered guide wire positioning model to avoid the accumulation of repeated erroneous identification structures and their transmission to subsequent classification models, which would affect the classification accuracy of the classification models.

[0015] The deep learning model includes: a backbone network, a neck network, and a detection head connected in sequence; The backbone network includes at least two cross-stage dual convolutional fusion modules and a spatial pyramid pooling module connected in sequence. Each cross-stage dual convolutional fusion module progressively extracts details and semantic features at different scales from the input image and performs feature fusion through the spatial pyramid pooling module to obtain a feature map containing multi-scale information. The feature map is then fused by the neck network and enters the detection head to output the detection result.

[0016] As a further solution, the cross-stage dual convolutional fusion module includes a first convolutional layer, a second convolutional layer, and a residual layer connected in sequence; the input image is sequentially passed through the first convolutional layer and the second convolutional layer for depth feature extraction, and the extracted depth features are concatenated with the input image through the residual layer to obtain multi-scale fusion features that fuse details and semantics.

[0017] As a further solution, based on the overlap between the first and second candidate conductor regions, a final candidate conductor region is generated, specifically as follows: Calculate the spatial overlap between each wire candidate box in the first wire candidate region and each wire candidate box in the second wire candidate region. If the spatial overlap is greater than a set threshold, merge the two wire candidate boxes and retain the larger wire candidate box. If the first wire candidate region is empty, then output the wire candidate boxes in the second wire candidate region; If the second wire candidate region is empty, then output the wire candidate boxes in the first wire candidate region.

[0018] As a further solution, the icing level classification result is determined based on the final candidate conductor region. Specifically, the target line image and the final candidate conductor region are used as inputs, and a pre-trained icing classification model is used to determine the icing level classification result.

[0019] The icing classification model comprises a global-local fusion network, a backbone network, and a classification head connected in sequence. The global-local fusion network includes a global encoder and a local encoder configured in parallel. The global encoder is used to encode features in the preprocessed image, and the local encoder is used to encode features in each wire candidate box in the final wire candidate region. The outputs of the global encoder and the local encoder are fused by a feature fusion module. The fused features are input to the backbone network to extract multi-scale features. The multi-scale features are then processed by the classification head to output the icing level classification result for each wire candidate box.

[0020] The ice classification model of this invention adopts a global-local heterogeneous feature collaborative fusion method. Through a cross-scale semantic complementarity mechanism, it effectively suppresses the gradual accumulation and propagation of errors in the cascaded inference process while ensuring high classification accuracy, thereby reducing the risk of error propagation in the cascaded model.

[0021] As a further solution, after the icing classification model outputs the icing level classification result, it also includes: For candidate wire frames belonging to the same icing level category, calculate the spatial overlap between each candidate wire frame, merge candidate wire frames with a spatial overlap greater than a set threshold, and retain the larger candidate wire frames.

[0022] Based on the post-processing of the output results of the ice-covered conductor positioning model, this invention further adopts a multi-result joint ROI spatial consistency decision mechanism. For the category conflict redundancy caused by the overlap of detection boxes between mutually exclusive categories, the detection results are disambiguated and refined through joint decision of spatial location and category probability, effectively removing redundant overlap.

[0023] As a further embodiment, the spatial overlap is specifically defined as the crossover ratio between two candidate wireframes.

[0024] A second aspect of the present invention provides a transmission line icing detection system, comprising: The data acquisition module is configured to acquire images of the target line. The first conductor candidate region extraction module is configured to perform conductor detection on the target line image and generate a first conductor candidate region; wherein the conductor detection is based at least on the inherent geometric characteristics of the transmission conductor. The second conductor candidate region extraction module is configured to use a pre-trained deep learning model to perform conductor detection on the target line image and generate a second conductor candidate region. The conductor candidate region fusion module is configured to generate a final conductor candidate region based on the overlap relationship between the first conductor candidate region and the second conductor candidate region; The icing level classification module is configured to determine the icing level classification result based on the final candidate conductor region.

[0025] As a further option, it also includes: The first post-processing module is configured to filter and merge the candidate bounding boxes of the ice-covered conductor positioning model. The second post-processing module is configured to merge candidate wires belonging to the same icing level category output by the icing classification model.

[0026] A third aspect of the present invention provides a transmission line icing detection device, the device being configured to operate the above-described transmission line icing detection method.

[0027] A fourth aspect of the present invention provides a power transmission line icing detection chip, which integrates the aforementioned power transmission line icing detection device.

[0028] A fifth aspect of the present invention provides a computer-readable storage medium storing a plurality of instructions adapted to be loaded by a processor of a terminal device and executed by the above-described transmission line icing detection method.

[0029] A sixth aspect of the present invention provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the above-described method for detecting icing on power transmission lines.

[0030] Compared with the prior art, the beneficial effects of the present invention are: This invention utilizes both geometric prior constraints and a deep neural network model for wireline candidate box localization. Geometric prior constraints provide structured and physically interpretable coarse localization, while deep learning can handle difficult examples with blurred textures and low contrast. The two wireline detection branches operate independently and simultaneously, and then the results are combined through a fusion mechanism. This approach leverages the advantages of geometric priors—fast, label-free, and physically interpretable—while also utilizing the sensitivity of the deep learning model to textured and blurred regions. The two methods mutually validate and complement each other, improving wireline localization accuracy in complex backgrounds and enhancing the system's robustness and generalization ability in challenging environments.

[0031] This invention decouples the tasks of conductor detection and icing classification. The conductor detection model focuses on geometric boundary localization, while the icing classification model focuses on texture and shape recognition. The two do not interfere with each other. Compared with the end-to-end joint training method that directly outputs icing detection results, higher recognition accuracy can be obtained. At the same time, the independent cascading architecture of the two models has excellent scalability and low maintenance cost. When it is necessary to add new icing categories (such as sub-types of mixed frost), there is no need to retrain the conductor detection model. Only the classifier needs to be updated or added independently to quickly adapt to the new requirements, which greatly reduces the cost of model iteration and engineering deployment. Attached Figure Description

[0032] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0033] Figure 1 This is a flowchart of the transmission line icing detection method in an embodiment of the present invention; Figure 2 This is a schematic diagram of the network structure of the ice-covered conductor positioning model in an embodiment of the present invention; Figure 3 This is a schematic diagram of the network structure of the ice classification model in an embodiment of the present invention. Detailed Implementation

[0034] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0035] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0036] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0037] Example 1 In one or more embodiments, a method for detecting icing on transmission lines is disclosed, combined with... Figure 1 Specifically, it includes the following processes: S101: Acquire the target line image data and perform preprocessing.

[0038] As a specific example, image data of icing on power transmission line conductors is mainly obtained by image acquisition equipment (such as visible light cameras) installed on power transmission towers at a preset frequency, or by inspection using drones equipped with image acquisition equipment.

[0039] Preprocessing of the acquired transmission line image data mainly includes image denoising and enhancement, geometric correction and scale normalization. Image denoising and enhancement mainly solve the noise and blurring problems caused by heavy fog, rain, snow and rapid movement in the transmission line images. Geometric correction and scale normalization mainly solve the inconsistencies in the images caused by different shooting angles or changes in conductor scale.

[0040] S102: Based on the preprocessed image, extract the linear features in the image and use geometric constraints to select the first candidate region of the transmission line; wherein, the geometric constraints are constructed based on the inherent geometric characteristics of the transmission line.

[0041] Transmission lines appear as slender, continuous, and low-curvature curved structures in images, while natural backgrounds (trees, mountains, clouds, snow) do not possess such strict geometric constraints. Therefore, by extracting linear features from the image and utilizing geometric constraints such as collinearity, continuity, and parallelism, pixels or line segment clusters belonging to the transmission lines can be filtered out, thereby separating the transmission lines from the background and reducing the interference of the background on transmission line detection.

[0042] As a specific implementation method, this embodiment selects candidate regions for traverses based on geometric prior constraints. The specific process is as follows: S1021: After converting the input image to grayscale and performing Gaussian filtering, the Canny edge detection algorithm is used to extract the edge map; The edges of conductors are typically strong, continuous, and nearly straight, while natural backgrounds are usually short, cluttered, weak, or scattered. The goal of the Canny edge detection algorithm is to extract conductor edges from the original image while suppressing noise and background texture as much as possible.

[0043] The main process of the Canny edge detection algorithm includes: applying Gaussian filtering to the image to smooth it and suppress high-frequency noise; using the Sobel operator to obtain the gradient magnitude and direction of each pixel; retaining only the points with the largest local gradient along the gradient direction to make the edges thinner; setting high and low thresholds, marking strong and weak edges through double threshold detection, and obtaining a binary image with only edges (white) and background (black) through edge connection, thereby extracting the edge contour map of the transmission line.

[0044] S1022: After obtaining the binary edge map of the transmission line using the Canny algorithm, all straight line segments are extracted from the edge map using the cumulative probabilistic Hough transform (PPHT), and the endpoint coordinates, length, and angle of each line segment are recorded.

[0045] PPHT accumulates random sampling of edge points to output the coordinates of the start and end points of the line segment; then it calculates the length and angle of the line segment (the angle between the line segment and the horizontal direction).

[0046] S1023: Construct a three-layer geometric prior constraint screening rule based on the inherent geometric characteristics (length, straightness, continuity, parallelism) of transmission lines, specifically including: Length constraint: Remove straight line segments whose length is less than a preset first threshold (such as short line segments of 1 / 10 of the image diagonal). These straight line segments may be noise or tree branches because of their short length.

[0047] Angle constraints: Calculate the angle histogram of all remaining line segments, select line segments within a preset angle range (e.g., peak angle ±10°) as candidate line segments, and construct a candidate line segment set; Continuity constraint: In the candidate line segment set, if there is a distance between the endpoints of two line segments that is less than a preset second threshold (e.g., 50 pixels) and the angle between the extensions of the two line segments is less than a preset angle (e.g., 5°), then the two line segments are spliced ​​into a long line segment to represent the complete conductor trajectory.

[0048] S1024: For all line segments finally formed in the candidate line segment set, expand outward along their normal direction by a certain pixel width (such as the pixel width corresponding to the actual diameter of the wire) to generate multiple rectangular wire candidate boxes (ROIs), which serve as the first wire candidate region for geometric prior.

[0049] This embodiment obtains the first candidate region of the conductor through geometric prior constraints, which has strong robustness to common interferences of transmission lines (such as trees, ground textures, tower materials, insulator strings, uneven lighting, shadows, etc.); it constructs computable geometric constraints by utilizing the inherent geometric characteristics of the transmission conductor, and can achieve explicit splitting of the candidate region of the conductor without relying on a large number of annotations.

[0050] S103: Input the preprocessed image into the pre-trained ice-covered guide line localization model to obtain the second guide line candidate region.

[0051] In this embodiment, the basic model of the ice-covered guide wire positioning model is yolov8s, which has a relatively small number of parameters and meets the requirements in terms of model performance and inference speed.

[0052] As a specific implementation method, combined with Figure 2The structure of the ice-covered guide wire positioning model includes a three-segment structure consisting of a backbone network, a neck network, and a detection head connected in sequence.

[0053] The backbone network, as the first part, is mainly responsible for extracting rich features from the original image. Its structure includes multiple cross-stage dual convolutional fusion modules and a spatial pyramid pooling module connected in sequence. Each cross-stage dual convolutional fusion module includes a first convolutional layer, a second convolutional layer, and a residual layer connected in sequence. The input image is processed through the first convolutional layer and the second convolutional layer for deep feature extraction. The extracted deep features are concatenated with the input image through the residual layer to obtain multi-scale fusion features that fuse details and semantics. By progressively extracting details and semantic features at different scales from the input image through multiple cross-stage dual convolutional fusion modules, this operation ensures a richer flow of gradient information while maintaining the model's lightweight nature. The processed feature input is then processed by a fast spatial pyramid pooling module. This module first reduces the dimensionality of the input using a 1x1 convolution. The dimensionality-reduced feature map is then processed three times consecutively through the same 5x5 max pooling layer. Finally, the original feature map and the four results from these three pooling operations are concatenated along the channel dimension (Concat), and then a new feature map is output by adjusting the number of channels to match the input using another 1x1 convolution. This achieves efficient fusion of features at different scales, significantly improving the model's receptive field and yielding a feature map containing multi-scale information.

[0054] Feature maps of different scales output by the backbone network are input into the neck network, which contains upsampling layers, feature splicing layers, and residual layers. The multi-scale feature maps output by the backbone network are processed through a path aggregation-feature pyramid network structure. Finally, the detector in the detection head receives the feature maps fused from the neck network and directly predicts the positions of multiple wire candidate boxes and their corresponding confidence scores.

[0055] In this embodiment, the cross-stage dual convolutional fusion module splits the input into two branches for processing. One branch enters the main branch for in-depth processing to extract complex features, while the other branch undergoes almost no in-depth processing, directly retaining the original information. Finally, the results of the two branches are merged, resulting in a new feature map with richer output gradient information and stronger feature representation capabilities. The size of the new feature map is consistent with the input. Connecting multiple cross-stage dual convolutional fusion modules sequentially increases the model's receptive field, resulting in richer information in the acquired features and better adaptability of the model to multi-scale targets.

[0056] In this embodiment, during the training process of the icy guideline localization model, the original prediction results are compared with the actual bounding boxes to calculate the localization loss and measure the accuracy of the predicted bounding boxes. Then, the optimizer adjusts the model's weight parameters layer by layer through the backpropagation algorithm until the optimization requirements are met.

[0057] This embodiment uses a deep learning model to perform refined detection of candidate regions for conductors, and has high recognition accuracy, especially for scenes with blurred textures and complex backgrounds (such as conductors covered by snow).

[0058] This embodiment utilizes geometric prior constraints and a deep learning model to identify candidate regions for conductors. Geometric priors are adept at handling strongly structured scenes, such as those with low contrast or changes in illumination, and can detect candidate regions as long as the structure exists. Meanwhile, the deep learning model excels at handling weak textures and complex backgrounds. The collaboration between the two can improve the system's robustness and generalization ability in complex environments, and prevent incorrect identification of conductor regions in complex backgrounds from affecting the processing accuracy of subsequent icing classification tasks.

[0059] As a further implementation, the icing guideline localization model, in addition to outputting guideline candidate boxes and their corresponding confidence scores, also outputs the icing level classification result for each guideline candidate box. Since the recognition accuracy of this classification result is not high, it is only used for post-processing of the guideline candidate box detection results; the specific post-processing process includes: Set a confidence threshold λ, compare the confidence score of each wire candidate box with the confidence threshold, and filter out all wire candidate boxes with confidence scores lower than the confidence threshold; All candidate frames belonging to the same icing level category are grouped together and sorted in descending order of confidence score. First, the overlap between the highest confidence score and each remaining candidate frame in the same group is calculated. If the overlap with a candidate frame is greater than the set overlap threshold k, the candidate frame is considered to overlap with the candidate frame with the highest confidence score and is removed. Then, the overlap between the second-highest confidence score in the group and each subsequent candidate frame is calculated, and overlapping candidate frames are removed in the same way. The overlap between each candidate frame and each subsequent candidate frame is calculated in turn, and overlapping candidate frames are removed in the same way. After all candidate frames have been calculated, all remaining candidate frames are used as the final second candidate area for traverse.

[0060] The overlap between candidate boxes is obtained by calculating the cross-union ratio between two candidate boxes.

[0061] This embodiment removes redundancy from the candidate wire bounding boxes output by the icing wire location model to prevent redundant candidate wire bounding boxes from affecting subsequent wire icing classification results.

[0062] S104: Merge the conductor regions by calculating the spatial overlap between the first and second conductor candidate regions to obtain the final conductor candidate regions.

[0063] This embodiment performs duct-level localization on the input image based on the geometric prior of the duct and a deep learning model, respectively, and outputs geometric candidate boxes and depth candidate boxes. Then, the candidate boxes obtained by the two methods are matched through parallel candidate box fusion, that is, the spatial overlap of each geometric candidate box and each depth candidate box is calculated, as shown in the following formula: ; Where A represents the geometric candidate box and B represents the depth candidate box.

[0064] If the spatial overlap between the geometric candidate box and the depth candidate box exceeds a set threshold, the two candidate boxes are merged, and the output with the larger candidate box is selected. If the deep learning model does not output candidate boxes, it outputs geometric candidate boxes. If the geometric prior method does not output geometric candidate boxes, for example, due to wire bending or occlusion causing no geometric prior results, then depth candidate boxes are output.

[0065] Based on the matching process described above, the final candidate regions for conductors are obtained. The final candidate regions for conductors contain multiple candidate boxes for conductors, which are all used as inputs to the subsequent icing classification model.

[0066] S105: Using the preprocessed image and the final candidate region of the conductor as input, and utilizing the pre-trained icing classification model, output the icing level classification result.

[0067] In this embodiment, the icing classification model differs slightly from the icing guideway localization model. Since the icing classification model does not require multi-scale feature fusion to locate the guideway, the neck network can be removed, and a backbone network and classification head structure is mainly adopted. In addition, to avoid the accumulation and propagation of errors in the cascaded model, a specific global and local image feature fusion network layer is constructed before the backbone network. This layer can simultaneously utilize global scene information and local target features to improve the classification accuracy of tasks such as fine-grained classification and weakly supervised localization.

[0068] Combination Figure 3 The icing classification model includes: a global-local fusion network, a backbone network, and a classification head connected in sequence; the global-local fusion network includes a global encoder and a local encoder set in parallel. The global encoder is used to encode features of the preprocessed image, and the local encoder is used to encode features of each wire candidate box in the final wire candidate region; the outputs of the global encoder and the local encoder are fused by a feature fusion module, and the specific fusion formula is as follows: ; in, α , β The fusion weights are generated by the attention module of the fusion layer. This represents the output of the global encoder. This represents the output of the local encoder.

[0069] The fused features are input into the backbone network to extract multi-scale features. The backbone network has the same structure as the backbone network in the aforementioned icy guideline localization model. It also adopts a cross-stage dual convolutional fusion module, which performs depth transformation on the input feature map through multiple convolutional layers. The transformed features and the original input features are concatenated through a residual layer. After multiple cross-stage dual convolutional fusion modules extract features in sequence, they are processed in parallel by a fast spatial pyramid pooling module with pooling kernels of different sizes to efficiently extract multi-scale features. The obtained multi-scale features are processed by the fully connected layer and normalization layer in the classification head to obtain the icing level classification result for each guideline candidate box.

[0070] The difference between the predicted probability and the true label is calculated using the classification loss L_cls: ; in, Indicates the first i The true label of the class, The model predicts the first i The probability that the class is true.

[0071] After calculating the loss value, the optimizer will fine-tune the network weight parameters layer by layer through the backpropagation algorithm until the optimization requirements are met.

[0072] This embodiment uses a global-local heterogeneous feature fusion method. The global image provides context (such as meteorological and environmental clues), while the local ROI image provides microscopic textures and edge details of the ice cover. This ensures high classification accuracy while solving the problem of error accumulation and propagation in cascaded models.

[0073] As a further implementation method, a multi-result joint ROI post-processing decision mechanism is designed to perform spatial consistency verification on the location reliability of the detection box and the classification results, and output robust icing detection and icing level classification results.

[0074] In step S103, the post-processing of the candidate wire bounding box detection results has been performed to eliminate redundant candidate wire bounding boxes of the same category and improve the detection accuracy of the candidate wire bounding boxes. However, due to the limitation of the classification accuracy of the ice-covered wire location model, the post-processed candidate wire bounding box detection results may still contain redundant boxes of different categories. When these redundant boxes are input into the ice-covered classification model, redundant classification results will appear.

[0075] Based on this, this embodiment, in addition to the aforementioned post-processing, aggregates category-level results based on the precise classification category of each candidate wire bounding box output by the icing classification model; for candidate wire bounding boxes belonging to the same classification category, their category confidence is sorted, and following the same post-processing as the candidate bounding boxes output by the aforementioned icing wire bounding box positioning model, candidate bounding boxes of the same category with an overlap greater than a set threshold are eliminated by calculating the intersection-union ratio, and the remaining candidate bounding boxes and their corresponding icing category detection results are output, thereby further eliminating redundancy and avoiding false detections.

[0076] This embodiment's dual-stream input and joint decision-making mechanism fully leverages the complementarity between global scene information and local details, effectively mitigating the context loss problem caused by single ROI extraction. At the same time, post-processing decision logic suppresses false detections and classification conflicts, further enhancing the system's robustness and reliability.

[0077] As an optional implementation, when the conductor is obscured due to special circumstances, such as heavy fog, and neither the depth detection model nor the geometric prior detects the conductor, the entire image is passed to the subsequent icing classification model. The icing classification model sets a fuzzy category for this situation to improve the model's generalization. If the candidate regions for the conductor detected by the detection model or the geometric prior contain errors, the image will be initially eliminated based on the classification results by the classification model. Then, the multi-result joint ROI spatial consistency decision mechanism will perform result disambiguation again based on the detection results and classification results to avoid the accumulation and transmission of errors in the cascaded structure from affecting the analysis results.

[0078] As a further optional implementation, when training the icing guideway localization model and the icing classification model in this embodiment, the first step is to construct datasets for guideway localization and icing classification, including processes such as icing data acquisition, detection data resampling and augmentation, and classification data enhancement.

[0079] Specifically, data on icing of transmission line conductors is primarily collected using equipment installed on them, including data from multiple perspectives, time periods, and icing conditions to increase scene diversity. All image data is preprocessed and labeled using annotation tools to form an initial conductor icing database. Although the initial conductor icing data collection yields a certain amount of image data, it suffers from class imbalance. Therefore, a resampling strategy is employed, such as using interpolation methods, like bilinear interpolation, to determine new pixel values ​​by calculating a weighted average of a predetermined number of nearest-neighbor pixels; and / or, resampling through color transformation to simulate different lighting conditions and adjust the image's HSV color space (including hue, saturation, and brightness).

[0080] The processed data, together with the initial icing data of the guide wire, constitutes the training and validation datasets for the icing guide wire localization model. On the other hand, the initial icing data of the guide wire is enhanced by random rotation, aspect ratio, and color jitter according to the annotation. Finally, the enhanced icing image data, together with the unprocessed icing data, constitutes the training and validation datasets for the icing classification model.

[0081] The final detection result of this embodiment is to classify the degree of icing on the transmission lines into levels, specifically: no icing, slight icing, and severe icing. In addition, to address other situations in the conductor icing detection scenario, categories such as image blur and vibration damper are set. Image blur is mainly because image acquisition equipment often faces severe weather conditions such as low temperature and snow. Lens icing and conductor galloping are two typical scenarios that cause blurred acquired images. This embodiment specifically considers categories such as image blur and vibration damper when annotating the training dataset to achieve precise discrimination of icing type.

[0082] As a validation example, all icing image data were split into training and validation sets in a 9:1 ratio to iteratively optimize each model. The method ultimately achieved an accuracy of 95% on a test set of 4687 images.

[0083] This embodiment decouples conductor localization and icing detection into two independent cascaded branches. This separation avoids feature conflicts; the localization model focuses on geometric boundary localization, while the classification model focuses on texture and morphology recognition. The two do not interfere with each other, achieving higher overall accuracy compared to end-to-end joint training. Furthermore, this independently cascaded model architecture offers excellent scalability and low maintenance costs. When new icing categories (such as subcategories of mixed frost) are needed, there is no need to retrain the detection model; simply updating or adding a classifier is sufficient to quickly adapt to new requirements, significantly reducing the cost of model iteration and engineering deployment. This provides a high-precision, easily scalable, and engineerable new path for intelligent monitoring of icing on transmission lines.

[0084] Example 2 In one or more embodiments, a power transmission line icing detection system is disclosed, comprising: The data acquisition module is configured to acquire images of the target line. The first conductor candidate region extraction module is configured to perform conductor detection on the target line image and generate a first conductor candidate region; wherein the conductor detection is based at least on the inherent geometric characteristics of the transmission conductor. The second conductor candidate region extraction module is configured to use a pre-trained deep learning model to perform conductor detection on the target line image and generate a second conductor candidate region. The conductor candidate region fusion module is configured to generate a final conductor candidate region based on the overlap relationship between the first conductor candidate region and the second conductor candidate region; The icing level classification module is configured to determine the icing level classification result based on the final candidate conductor region.

[0085] As a further implementation, it also includes: The first post-processing module is configured to filter and merge the candidate bounding boxes of the ice-covered conductor positioning model. The second post-processing module is configured to merge candidate wires belonging to the same icing level category output by the icing classification model.

[0086] It should be noted that the specific implementation methods of the above modules are the same as those in Example 1, and will not be described in detail again.

[0087] Example 3 In one or more embodiments, a transmission line icing detection device is disclosed, the device being configured to operate the transmission line icing detection method as described in Embodiment 1.

[0088] In other embodiments, a transmission line icing detection chip is disclosed, which integrates the aforementioned transmission line icing detection equipment.

[0089] Example 4 In one or more embodiments, a computer-readable storage medium is disclosed, wherein a plurality of instructions are stored, the instructions being adapted to be loaded by a processor of a terminal device and executed by the transmission line icing detection method of Embodiment 1.

[0090] In other embodiments, a computer program product is disclosed, including a computer program or instructions that, when executed by a processor, implement the transmission line icing detection method described in Embodiment 1.

[0091] In such an embodiment, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by the central processing unit, it performs the various functions defined in the steps of Embodiment 1. Alternatively, the computer program instructions can also be stored in a computer-readable storage medium capable of directing a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium implement the working process of Embodiment 1.

[0092] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting icing on power transmission lines, characterized in that, include: Acquire the target line image; Conductor detection is performed on the target line image to generate a first conductor candidate region; wherein, the conductor detection is based at least on the inherent geometric characteristics of the transmission conductor; A pre-trained deep learning model is used to detect conductors in the target line image, and outputs conductor candidate boxes, corresponding confidence scores, and icing level classification results for each conductor candidate box. Based on the confidence scores of the conductor candidate boxes and the spatial overlap of each conductor candidate box under the same icing level category, the conductor candidate boxes are filtered to generate a second conductor candidate region. Based on the overlap between the first and second candidate conductor regions, the final candidate conductor regions are generated. Using the target line image and the final candidate conductor region as input, the icing level classification result is determined using a pre-trained icing classification model; For candidate wire frames belonging to the same icing level category, calculate the spatial overlap between each candidate wire frame, merge candidate wire frames with a spatial overlap greater than a set threshold, and retain the larger candidate wire frames. Specifically, based on the overlap between the first and second candidate conductor regions, the final candidate conductor regions are generated as follows: Calculate the spatial overlap between each wire candidate box in the first wire candidate region and each wire candidate box in the second wire candidate region. If the spatial overlap is greater than a set threshold, merge the two wire candidate boxes and retain the larger wire candidate box. If the first wire candidate region is empty, output the wire candidate box in the second wire candidate region; if the second wire candidate region is empty, output the wire candidate box in the first wire candidate region. The icing classification model includes: a global-local fusion network, a backbone network, and a classification head connected in sequence; the global-local fusion network includes a global encoder and a local encoder set in parallel, the global encoder is used to encode features of the preprocessed image, and the local encoder is used to encode features of each wire candidate box in the final wire candidate region; the outputs of the global encoder and the local encoder are fused by a feature fusion module, the fused features are input into the backbone network to extract multi-scale features, and the multi-scale features are output by the classification head to output the icing level classification result of each wire candidate box.

2. The method for detecting icing on transmission lines as described in claim 1, characterized in that, Conductor inspection based on the inherent geometric characteristics of transmission lines, specifically: Extract linear features from the target line image; Based on length and / or angle features, linear features representing conductors are selected from the linear features to generate the first conductor candidate region.

3. The method for detecting icing on transmission lines as described in claim 2, characterized in that, The linear features to be extracted from the target line image are as follows: edge feature map is extracted using an edge detection algorithm, and then all straight line segments are extracted from the edge feature map by cumulative probability Hough transform, and the endpoint coordinates, length and angle of each straight line segment are recorded.

4. A method for detecting icing on transmission lines as described in claim 2 or 3, characterized in that, Based on length and / or angle features, linear features representing conductors are selected from the linear features to generate the first conductor candidate region, specifically: Remove straight line segments whose length is less than a preset first threshold; Calculate the angle histogram of all remaining line segments, select line segments within a preset angle range as candidate line segments, and construct a candidate line segment set; If, in the candidate line segment set, there exists a distance between the endpoints of two line segments that is less than a preset second threshold, and the angle between the extensions of the two line segments is less than a preset angle, then the two line segments are spliced ​​into one line segment. All the line segments ultimately formed in the candidate line segment set are expanded outward along their normal direction by a set pixel width to generate multiple rectangular wire candidate boxes, which serve as the first wire candidate region.

5. The method for detecting icing on transmission lines as described in claim 1, characterized in that, The deep learning model includes: a backbone network, a neck network, and a detection head connected in sequence; The backbone network includes at least two cross-stage dual convolutional fusion modules and a spatial pyramid pooling module connected in sequence. Each cross-stage dual convolutional fusion module progressively extracts details and semantic features at different scales from the input image and performs feature fusion through the spatial pyramid pooling module to obtain a feature map containing multi-scale information. The feature map is then fused by the neck network and enters the detection head to output the detection result.

6. The method for detecting icing on transmission lines as described in claim 5, characterized in that, The cross-stage dual convolutional fusion module includes a first convolutional layer, a second convolutional layer, and a residual layer connected in sequence. The input image is processed through the first convolutional layer and the second convolutional layer for depth feature extraction. The extracted depth features are then concatenated with the input image through the residual layer to obtain multi-scale fusion features that fuse details and semantics.

7. The method for detecting icing on transmission lines as described in claim 1, characterized in that, The spatial overlap is specifically defined as the crossover ratio between two candidate wireframes.

8. A system for implementing the transmission line icing detection method of claim 1, characterized in that, include: The data acquisition module is configured to acquire images of the target line. The first conductor candidate region extraction module is configured to perform conductor detection on the target line image and generate a first conductor candidate region; wherein the conductor detection is based at least on the inherent geometric characteristics of the transmission conductor. The second conductor candidate region extraction module is configured to use a pre-trained deep learning model to perform conductor detection on the target line image and generate a second conductor candidate region. The conductor candidate region fusion module is configured to generate a final conductor candidate region based on the overlap relationship between the first conductor candidate region and the second conductor candidate region; The icing level classification module is configured to determine the icing level classification result based on the final candidate conductor region.

9. The system as described in claim 8, characterized in that, Also includes: The first post-processing module is configured to filter and merge the candidate bounding boxes of the ice-covered conductor positioning model. The second post-processing module is configured to merge candidate wires belonging to the same icing level category output by the icing classification model.

10. A transmission line icing detection device, characterized in that, The device is configured to run the transmission line icing detection method as described in any one of claims 1-7.

11. A power transmission line icing detection chip, characterized in that, The device integrates the power transmission line icing detection equipment as described in claim 10.

12. A computer-readable storage medium storing a plurality of instructions, characterized in that, The instructions are adapted to be loaded by the processor of the terminal device and executed as described in any one of claims 1-7, the method for detecting icing on transmission lines.

13. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by the processor, they implement the transmission line icing detection method according to any one of claims 1-7.

Images