A reverse identification method and device for typical defects of an overhead power transmission line

Abstract

The application discloses an overhead power transmission line typical defect reverse identification method and device, relates to the electric power inspection technical field, and includes the following steps: constructing a component identification model according to visual relationship information between power transmission components; constructing a component feature extraction network based on component region images; extracting component features by using the component feature extraction network based on normal component image data to construct a normal component feature library; extracting real-time component region images of unmanned aerial vehicle inspection images based on the component identification model, performing similarity calculation on the real-time component region images and the normal component feature library, and judging whether the line is abnormal or not. The application focuses on the normal feature and state identification of components, improves the identification precision, can adapt to various defect types and conditions, and improves the identification flexibility of diversified defects. The application overcomes the problems that the prior art is poor in image quality and lacks defect samples in complex forest areas or mountainous areas, and thus cannot accurately detect defects.

Description

Technical Field

[0001] This application relates to the field of power line inspection technology, specifically a method and device for reverse identification of typical defects in overhead transmission lines. Background Technology

[0002] Defect detection in key components of power transmission lines mainly includes the detection of defects in insulators, fittings, bolts, and other components that play a crucial role in power transmission lines. Compared with manual inspection, drone inspection is less costly, shorter in cycle, more mobile, and safer, and is gradually replacing manual inspection. Drones transmit a large number of images of power transmission lines, and it is impractical to manually check each image for defects. Therefore, an image processing system is needed to automatically detect defects in key parts of the power transmission lines from the large amount of image data transmitted by drones. Compared with traditional image processing methods, the biggest advantage of deep learning-based image processing technology is its higher defect detection accuracy and shorter processing time. Therefore, deep learning-based target detection models are the optimal choice for processing power transmission line inspection images. However, deep learning models require a large amount of sample data. For deep target detection models with good detection performance, a large amount of defect image data covering multiple types of scenes is usually needed as a training basis. Annotating aerial images captured by drones requires significant manpower, resources, and time. Furthermore, limitations in drone photography technology result in some aerial images being blurry, low-resolution, backlit, and limited shooting angles, all of which impact the accuracy of model detection. In actual production, image data for certain defect types is readily available in large quantities, but other defect types occur less frequently, making collection difficult, data volume small, and samples insufficient for various defect types, thus limiting the learning capabilities of deep learning models. Additionally, defects in critical components of certain power transmission lines often constitute a small percentage of aerial images, making detection challenging. Small target detection and achieving fine-grained recognition even with small visual differences between different types of defects in the dataset are pressing problems that need to be solved. Summary of the Invention

[0003] The purpose of this application is to address the problem that existing deep learning-based transmission line defect detection methods suffer from poor image quality and insufficient defect samples in complex forest or mountainous environments, leading to inaccurate defect detection. This application proposes a reverse identification method and device for typical defects in overhead transmission lines. By focusing on the normal features and state identification of components, the method reduces the impact of background complexity on identification accuracy and improves identification precision. At the same time, it can adapt to various defect types and conditions, improving the flexibility of identifying diverse defects.

[0004] To achieve the above objectives, the technical solutions adopted in the embodiments of this application are as follows: In a first aspect, embodiments of this application provide a method for reverse identification of typical defects in overhead transmission lines, the method comprising: A component recognition model is constructed based on the visual relationship information between power transmission components; Construct a component feature extraction network based on component region images; Based on normal component image data, a component feature extraction network is used to extract component features to construct a normal component feature library; Based on the component recognition model, real-time component area images are extracted from UAV inspection images, and similarity calculations are performed between these images and a normal component feature library to determine whether the line is abnormal.

[0005] In this scheme, the visual relationships between power transmission equipment are analyzed through a component recognition model, such as the relative positions, orientations, and interactions between components. This allows for more accurate identification and classification of various components on the power transmission line, improving the effectiveness of subsequent deep learning processes. Based on the identified core components of the power equipment, a feature extraction network is used to further refine the identification of the unique visual features of each component, such as shape, size, color, and texture. This ensures the identification of components at a fine-grained level, thereby obtaining accurate component difference features. This facilitates similarity analysis between the difference features of each component and their corresponding normal features, thus determining whether a component is abnormal. Compared to the method of learning abnormal features of each component to extract abnormal features from inspection images to identify defects, the reverse identification mode (i.e., instead of using the abnormal features of various defect samples as the main learning object, but first establishing a normal component feature library, and then comparing the features of the component to be detected with the corresponding normal features, and judging whether the component is abnormal based on the degree of similarity deviation) constructs a component-level full-scale line sample library and identification mechanism, which improves the accuracy and detection rate of current defect identification.

[0006] Optionally, the step of constructing a component recognition model based on the visual relationship information between power transmission components includes: Component type labeling is performed on drone inspection images to obtain power transmission component samples; Based on the target detection algorithm, a basic component detection model is trained using power transmission component samples to obtain the basic region features of the power transmission components; Based on the spatial and semantic relationships between components of power transmission equipment, a graph neural network is used to construct a visual relationship model of the components and obtain visual relationship information of the components. Semantic features are extracted based on visual relationship information of components. These semantic features are then fused with the basic regional features of power transmission components. A component recognition model is then constructed based on the fused features.

[0007] Optionally, the component types include at least poles, foundations, ancillary facilities, hardware, and insulators.

[0008] Optionally, the component feature extraction network constructed based on component region images includes: The component region image is segmented into non-overlapping image blocks, and each image block is transformed into a one-dimensional feature vector through linear transformation; The target image patch is obtained by masking and filtering a one-dimensional feature vector based on a random masking algorithm. Based on Transformer and positional encoding, the target image patch is transformed into deep image features containing positional mask markers; The deep image features are mapped back to the original pixel space by the decoder to reconstruct the original pixel values ​​of the filtered image blocks, and the difference between the original pixel values ​​and the real pixels of the part region image is used as a supervision signal to establish a loss function. A self-supervised model is trained based on the loss function to form a component feature extraction network.

[0009] Optionally, the step of training a self-supervised model based on a loss function to form a component feature extraction network includes: Based on the error between the original pixel values ​​of the reconstructed image patch and the actual pixels of the part region image, the parameters of the encoder and decoder are optimized. The encoder and decoder are trained based on the optimized parameters. When training is complete, the decoder used for image reconstruction is removed, and the encoder used for image feature extraction is retained as the final part feature extraction network.

[0010] Optionally, the step of extracting component features from normal component image data using a component feature extraction network to construct a normal component feature library includes: Acquire normal component images containing all component types of transmission lines, and filter valid component images based on the component recognition model; The component features of valid component images are extracted based on the component feature extraction network. The component ID is associated and bound with the corresponding feature to form a component ID and feature mapping dictionary containing the normal state features of various components, which is used as the normal component feature library.

[0011] Optionally, the step of extracting real-time component region images from UAV inspection images based on the component recognition model, calculating their similarity with a normal component feature library, and determining whether the line is abnormal includes: In response to real-time inspection images from the inspection drone, all components are located based on the component recognition model, and the target component area image is cropped. The target component region image is input into the component feature extraction network to generate a real-time component feature vector; Based on the component ID, the corresponding normal feature vector is retrieved from the normal component feature library, and the similarity between the real-time component feature vector and the normal feature vector is calculated. The similarity is compared with a preset similarity threshold. If the similarity is greater than or equal to the similarity threshold, the target component is determined to be normal; if the similarity is less than the similarity threshold, the target component is determined to be abnormal.

[0012] Secondly, embodiments of this application provide a reverse identification device for typical defects in overhead transmission lines, including: a component identification module, used to construct a component identification model based on visual relationship information between transmission components; The feature extraction module is used to construct a component feature extraction network based on component region images; The component feature library module is used to extract component features from normal component image data using a component feature extraction network to construct a normal component feature library. The defect identification module is used to extract real-time component area images from UAV inspection images based on the component identification model, calculate the similarity between these images and a normal component feature library, and determine whether the line is abnormal.

[0013] Thirdly, embodiments of this application provide a computer device, including: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory is used to store computer programs; and the processor is used to implement the steps of the method described in the first aspect above when executing the program stored in the memory.

[0014] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method described in the first aspect above.

[0015] The beneficial effects of this application are: By using a component recognition model and a component feature extraction network to understand and interpret the visual relationships between power equipment components, the morphology of normal components is learned in a structured manner. Whether a component is abnormal is determined by comparing the morphology of the component to be detected with that of a normal component. This reverse recognition method effectively overcomes the problem of low model recognition accuracy caused by the lack of samples in the current forward recognition, thereby improving the accuracy and reliability of recognition. Attached Figure Description

[0016] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0017] Figure 1 A flowchart illustrating a method for reverse identification of typical defects in overhead transmission lines, provided in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of the spatial relationship of components provided in an embodiment of this application.

[0019] Figure 3 This is a schematic diagram of a component identification result provided in an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely one preferred embodiment of this application and are only used to explain this application. They do not limit the scope of protection of this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] Example 1: As Figure 1 As shown, a method for reverse identification of typical defects in overhead transmission lines includes steps S1-S4, wherein: S1. Construct a component recognition model based on the visual relationship information between power transmission components.

[0022] In an optional embodiment, step S1 includes: Component type labeling is performed on drone inspection images to obtain power transmission component samples; Based on the target detection algorithm, a basic component detection model is trained using power transmission component samples to obtain the basic region features of the power transmission components; Based on the spatial and semantic relationships between components of power transmission equipment, a graph neural network is used to construct a visual relationship model of the components and obtain visual relationship information of the components. Semantic features are extracted based on visual relationship information of components. These semantic features are then fused with the basic regional features of power transmission components. A component recognition model is then constructed based on the fused features.

[0023] In one alternative embodiment, the component types include at least towers, foundations, ancillary facilities, hardware, and insulators.

[0024] In some embodiments, labeling component types in UAV inspection images to obtain power transmission component samples includes: labeling components in UAV inspection images, including towers, foundations, ancillary facilities such as signs, bird protection facilities, lightning protection facilities, etc.; hardware such as connecting hardware, suspension clamps, tension clamps, vibration dampers, spacers, etc.; and insulators such as composite insulators, ceramic insulators, glass insulators, ground wire insulators, etc. The labels are stored as XML files in PASCAL VOC format.

[0025] It should be noted that power transmission line components often have relatively fixed combination rules and structures. Since the camera line of sight is relatively parallel to the ground when the drone collects images, and the assembly and connection of components are structured and fixed, the different categories of components will show strong spatial visual relationships in the image, including vertical and horizontal visual relationships. When annotating the image, there are overlapping areas between the annotation boxes, which are regarded as overlapping visual relationships.

[0026] In this embodiment, based on the vertical, horizontal, and overlapping visual relationships of the image, rules for extracting visual relationships are formulated from the relative positions and spatial angles of the bounding boxes in the dataset: For a pair of parts appearing in the same image, the intersection-union ratio (IUR) between the two parts is first calculated. If the IUR is greater than a threshold (e.g., set to 0.3), then the two parts are considered to have a visual overlap relationship; if the two parts do not have a visual overlap relationship, the center coordinates of the parts are converted. The horizontal and vertical visual relationships between the two parts can be determined by the vector formed by the center coordinates of the two parts. Figure 2 As shown, with the weight at the center of the image as the object of observation, the yellow shaded area represents the horizontal visual relationship, the blue shaded area represents the vertical visual relationship, the two black arc angles represent the size of θ, the equalizing ring and the weight in the red box represent the vertical visual relationship, and the connecting plate and the weight in the green box represent the horizontal visual relationship.

[0027] Specifically, the component recognition model works as follows: It acquires object information from images using a Convolutional Neural Network (CNN) to transform pixels into feature representations for each object; it then uses object detection models (such as Faster R-CNN or YOLO) to identify different objects in the image, treating each detected transmission line component as a graph node and the visual relationships between different components as graph edges. The components corresponding to the graph nodes include one or more of the following: ceramic insulators, composite insulators, parallel groove clamps, vibration dampers, U-shaped hanging rings, hanging plates, connecting plates, suspension clamps, tension clamps, spacers, equalizing rings, shielding rings, counterweights, and adjusting plates. The node features of each graph node include component category, category confidence, bounding box coordinates, bounding box center point coordinates, bounding box width and height, and basic region features. The basic region features are output by the feature extraction layer or candidate region feature extraction layer of the basic component detection model and are used to characterize the appearance, texture, and local morphological information of the target component. Graph edges are used to characterize the visual relationships between two components. For any two detected components in the same UAV inspection image, calculate the intersection-over-union (IoU) ratio, center point distance, relative direction, relative angle, and normalized coordinate offset of their bounding boxes. If the IoU ratio of the two component bounding boxes is greater than a preset overlap threshold, an overlapping visual relationship edge is established; if the direction of the line connecting their center points is mainly distributed along the horizontal direction of the image, a horizontal visual relationship edge is established; if the direction of the line connecting their center points is mainly distributed along the vertical direction of the image, a vertical visual relationship edge is established. The edge features of the graph edge include the horizontal offset, vertical offset, center point distance, relative angle, IoU ratio, and visual relationship type of the two components. Visual relationship types include horizontal visual relationship, vertical visual relationship, overlapping visual relationship, and no significant relationship.

[0028] The node and edge features described above are input into a graph neural network for relation propagation. The graph neural network aggregates the node and edge features of adjacent components based on adjacency relationships, updates the contextual representation of each node, and thus obtains the semantic relation features corresponding to each target component. These semantic relation features are used to characterize the spatial positional relationships, assembly relationships, and contextual relationships between the target component and its adjacent components.

[0029] In this model, Faster R-CNN = RPN + Fast R-CNN, meaning the RPN used for extracting candidate regions is combined with Fast R-CNN to integrate the entire object detection process into a single neural network, Faster R-CNN. The RPN network is essentially a fully convolutional structure that calculates and outputs a batch of candidate regions from the input raw image. The RPN consists of convolutional layers and activation layers, omitting fully connected layers to improve model efficiency and reduce training time. As a fully convolutional structure, regardless of the input feature map size, it will ultimately output corresponding rectangular candidate boxes, such as... Figure 3 As shown.

[0030] Furthermore, the object detection model performs classification and bounding box regression, determining whether each candidate bounding box contains an object and its category. The classification process includes: generating a set of candidate bounding boxes using an RPN (Regressive Part Network), which represent potential part locations; for each candidate bounding box, extracting feature vectors from the input image for classification and bounding box regression (this is typically done by a CNN network); during classification, the feature vectors are passed to the classification head of a fully connected neural network. The class head maps feature vectors to a fixed number of classes. The final layer of the class head uses a softmax activation function to assign a class probability distribution to each candidate bounding box and uses cross-entropy loss to compare the model's predicted classes. The classification distribution is compared with the true label's category distribution, and the loss function is optimized through backpropagation and gradient descent to improve classification performance; finally, for each candidate bounding box, the category with the highest category probability is selected as the final classification result.

[0031] Specifically, by adding a relationship modeling module to the object detection model, the architecture of the object detection model is modified to consider the relationships between power components, thus capturing the semantic relationships between components more accurately. Specifically, a Graph Convolutional Network (GCN), a Relational Feature Network (RFN), or a Recurrent Neural Network (RNN) can be adaptively added to any layer of the original object detection model architecture.

[0032] In some examples, CNNs are used to extract feature vectors for each object. Each object's feature vector typically includes location information, appearance information, and contextual information. By combining the object features with the location information using a modified object detection model, a relational feature representation can be constructed, thereby obtaining a part recognition model.

[0033] In some examples, the loss function is expressed as: ,in, For component category classification loss, For bounding box regression loss, The loss is used for classifying visual relationships between components. , , These are the weight coefficients for each loss term, and the weight coefficients can be flexibly set based on the actual feature extraction situation during the actual detection process.

[0034] In this embodiment, the inspection data acquired by the inspection drone includes images, audio, and text. The modified target detection model can process various types of inspection data, fuse data from each input modality to generate high-level feature representations, and perform feature combination, which helps the model capture the correlation and semantic relationships between different modalities. By combining graph neural networks, the visual relationships between power components are effectively introduced into the component recognition model, providing an effective recognition mechanism for components in complex scenarios, thereby improving the accuracy of power component detection.

[0035] S2. Construct a component feature extraction network based on component region images.

[0036] In an optional embodiment, step S2 includes: The component region image is segmented into non-overlapping image blocks, and each image block is transformed into a one-dimensional feature vector through linear transformation; The target image patch is obtained by masking and filtering a one-dimensional feature vector based on a random masking algorithm. Based on Transformer and positional encoding, the target image patch is transformed into deep image features containing positional mask markers; The deep image features are mapped back to the original pixel space by the decoder to reconstruct the original pixel values ​​of the filtered image blocks, and the difference between the original pixel values ​​and the actual pixels of the part region image is used as a supervision signal to establish a loss function. A self-supervised model is trained based on the loss function to form a component feature extraction network.

[0037] In an optional embodiment, the step of training a self-supervised model based on a loss function to form a component feature extraction network includes: Based on the error between the original pixel values ​​of the reconstructed image patch and the actual pixels of the part region image, the parameters of the encoder and decoder are optimized. The encoder and decoder are trained based on the optimized parameters. When training is complete, the decoder used for image reconstruction is removed, and the encoder used for image feature extraction is retained as the final part feature extraction network.

[0038] In this embodiment, the purpose of constructing a component feature extraction network is to separate the component feature space. A self-supervised model is used for training. Compared to supervised learning, it does not require pre-labeling a large amount of data. By automatically generating labels or targets, the task is transformed into a supervised learning problem. This means that from unlabeled data, a training task similar to that in supervised learning can be created, such as image reconstruction, text occlusion, or sequence prediction. These generated labels do not require manual intervention but are automatically generated based on the original data. Therefore, it reduces the reliance on a large number of defect-labeled samples and enhances the component feature extraction network's ability to represent the morphological, textural, and structural features of normal components under different shooting conditions, providing a feature basis for subsequent abnormal deviation judgment based on a normal feature library.

[0039] Specifically, self-supervised learning constructs an asymmetric encoder-decoder structure, defines a self-supervised objective, and uses unsupervised training to learn useful feature representations. The encoder operates only on visible, unmasked image patches, while the decoder reconstructs the original image from the latent representation and the masked target image patches. For masked target image patches, a self-supervised task is generated. Training based on this asymmetric encoder-decoder approach can accelerate training speed and improve accuracy. Here, the masking ratio is set to r, for example, 75% of the total image patches. Masked image patches do not participate in encoder feature extraction but are involved in image reconstruction during the decoding stage through mask markings.

[0040] In some examples, the input image is segmented into a series of overlapping or non-overlapping image patches, each a small image fragment. These patches are typically stretched into one-dimensional vectors before being input into the Transformer. Specifically, the part region image is normalized, uniformly scaling it to a 224×224 pixel three-channel image. Let the normalized part region image size be H×W×C, where H is the image height, W is the image width, C is the number of channels, and P is the size of the image patch. Then, the part region image is segmented into... There are three non-overlapping image patches, each of which has a dimension of [dimensionality] after being flattened. In the Transformer, the input image is transformed into a feature representation by an encoder to minimize the difference between the input and target image patches, establishing a self-supervised learning objective function for model training. During training, the loss between the feature representation acquired by the encoder and the target image patch is calculated, and the backpropagation algorithm is used to optimize the model parameters to reduce the loss. This process is repeated until the model converges or reaches a predetermined number of training epochs. After self-supervised training, the model can be fine-tuned on a specific task, such as image classification or object detection. Fine-tuning can be performed using labeled data to further improve the model's performance. For example, the encoder has 12 layers, each including a multi-head self-attention module, a feedforward neural network module, residual connections, and a normalization layer; the number of attention heads is set to 12, and the embedding dimension is 768. The decoder is used to reconstruct the masked image patch, and its number of layers is set to 4, with a decoding dimension that can be set to 512.

[0041] Understandably, considering that in the Transformer architecture, the input is a two-dimensional matrix with a shape that can be represented as (N, D), where N is the length of the matrix and D is the dimension of each vector in the matrix, the self-supervised model training process transforms the three-dimensional image into a two-dimensional input sequence of (N, D). This sequence can be viewed as a series of flattened image patches, that is, the image is divided into small pieces and then flattened. The sequence contains a total of... There are 10 image patches, and the dimension of each image patch is 1. However, the actual required vector dimension is not... Therefore, a linear transformation is needed to compress the image patch to the required dimension.

[0042] Furthermore, to incorporate spatial relationship information between image patches, positional encoding is required, which embeds the positional information of each image patch into its feature representation. Positional encoding can employ sine or cosine functions. The self-attention mechanism is implemented by weighted summing of the feature representations of an image patch with those of itself and other image patches. In self-attention computation, each image patch performs attention calculations with other image patches to capture global relationships, and the result is a new representation for each image patch.

[0043] In this embodiment, by minimizing image reconstruction error and combining it with a self-supervised learning mechanism, the component feature extraction network's ability to learn useful features on unlabeled data is further enhanced. This enables the network to extract and learn key features of components from complex backgrounds more accurately and efficiently, especially to identify subtle abnormal areas. This allows for accurate and rapid response to component identification problems in transmission line images, improving the accuracy and efficiency of transmission line inspection.

[0044] S3. Based on normal component image data, a component feature extraction network is used to extract component features to construct a normal component feature library.

[0045] In an optional embodiment, step S3 includes: Acquire normal component images containing all component types of transmission lines, and filter valid component images based on the component recognition model; The component features of valid component images are extracted based on the component feature extraction network. The component ID is associated and bound with the corresponding feature to form a component ID and feature mapping dictionary containing the normal state features of various components, which is used as the normal component feature library.

[0046] In this embodiment, component ID assignment is the process of associating a unique identifier with each detected component. Component IDs are typically used to track the status, historical data, and maintenance records of components.

[0047] Specifically, image data along the inspection route is collected, and the location information for each image is recorded to ensure that images and location information are paired for association. Based on a component recognition model, the location of components in each inspection image is detected. The detected location of each component is matched with the location information at the time of image capture, and each component is assigned a unique ID, which can be an integer, alphanumeric combination, or any unique identifier. IDs can be generated according to certain rules, such as sequential allocation or using a specific encoding rule, to ensure that the ID is unique throughout the inspection route. Finally, the unique ID of each component, location information, component detection location, image data, and other information are recorded in a database or stored in a file system.

[0048] In some examples, preprocessing is performed on the acquired component regions to improve the performance and quality of the component feature extraction network. The preprocessed component regions are then input into the component feature extraction network. After processing through each layer, the network outputs a feature vector representing the feature representation of that component image extracted by the network. The feature vectors output from the component feature extraction network need to be formatted and transformed to meet database storage requirements. Simultaneously, the component ID, corresponding image filename, and acquisition time are integrated with the feature vectors to form a complete data record, which is then inserted into the database. Based on the component ID database, features of normal components are extracted to build a normal component feature database.

[0049] In this embodiment, by constructing a normal component feature library, detailed information on any specific component can be quickly accessed, simplifying the inspection process and improving the accuracy of maintenance work. Based on this normal component feature library, potential problems can be promptly identified and resolved by tracking component status in real time, reducing the risk of system failures and power outages.

[0050] S4. Based on the component recognition model, extract real-time component area images from the UAV inspection images, calculate the similarity between them and the normal component feature library, and determine whether the line is abnormal.

[0051] In an optional embodiment, step S4 includes: In response to real-time inspection images from the inspection drone, all components are located based on the component recognition model, and the target component area image is cropped. The target component region image is input into the component feature extraction network to generate a real-time component feature vector; Based on the component ID, the corresponding normal feature vector is retrieved from the normal component feature library, and the similarity between the real-time component feature vector and the normal feature vector is calculated. The similarity is compared with a preset similarity threshold. If the similarity is greater than or equal to the similarity threshold, the target component is determined to be normal; if the similarity is less than the similarity threshold, the target component is determined to be abnormal.

[0052] In this embodiment, the similarity between the real-time component feature vector and the normal feature vector can be calculated using Euclidean distance, correlation coefficient, cosine similarity, etc. By measuring the similarity between the real-time component image feature vector and the corresponding normal image feature vector, problems such as damaged components and foreign objects can be identified in reverse, improving the detection efficiency and accuracy of transmission lines, reducing false alarms and missed alarms, and ensuring the safe and stable operation of transmission lines.

[0053] In some embodiments, the similarity is compared with a preset similarity threshold, and an anomaly level is output based on the degree of similarity deviation. Specifically, let S be the similarity between the real-time component feature vector and the corresponding normal component feature vector, where S ranges from 0 to 1. A larger S indicates that the component to be detected is closer to the normal state. A first similarity threshold T1, a second similarity threshold T2, and a third similarity threshold T3 are set, where T1 is greater than T2, and T2 is greater than T3. When S is greater than or equal to T1, the target component is determined to be normal; when S is less than T1 but greater than or equal to T2, the target component is determined to be suspected of being abnormal; when S is less than T2 but greater than or equal to T3, the target component is determined to be abnormal; and when S is less than T3, the target component is determined to be severely abnormal.

[0054] Understandably, with the passage of time and the accumulation of more data, the feature library of normal components can be updated and optimized to improve recognition accuracy. Simultaneously, the similarity threshold can be adjusted based on actual operational conditions to achieve optimal detection results. Compared to normal defect detection methods, this reverse recognition method avoids the problems of insufficient overall defect samples for some categories or insufficient subcategories within defect categories with sufficient sample sizes. Because it does not rely on training with a large number of defect images, even with a small number of defects, it can effectively identify defects by focusing on the features of normal components, even when faced with new defect morphologies, thus improving data utilization. Furthermore, reverse recognition is effective for defects of different types and degrees. The enhanced adaptability and generalization ability improve the adaptability and flexibility of power equipment defect detection.

[0055] Based on the same inventive concept, this application also provides a reverse identification device for typical defects of overhead transmission lines, corresponding to a reverse identification method for typical defects of overhead transmission lines. The device includes: The component identification module is used to construct a component identification model based on the visual relationship information between power transmission components. The feature extraction module is used to construct a component feature extraction network based on component region images; The component feature library module is used to extract component features from normal component image data using a component feature extraction network to construct a normal component feature library. The defect identification module is used to extract real-time component area images from UAV inspection images based on the component identification model, calculate the similarity between these images and a normal component feature library, and determine whether the line is abnormal.

[0056] In this embodiment, by analyzing the visual relationships between power transmission equipment through a component identification model, such as the relative positions, orientations, and interactions between components, various components on the power transmission line can be identified and classified more accurately, thereby improving the effect of subsequent deep learning processes and enhancing the accuracy and reliability of component feature recognition. Based on the identified core components of the power equipment, the unique visual features of each component, such as shape, size, color, and texture, are further refined using a feature extraction network to ensure the identification of component differences. This allows for the acquisition of accurate component difference features, facilitating similarity analysis between the difference features of each component and their corresponding normal features to determine whether a component is abnormal. Compared to the method of learning abnormal features of each component to extract abnormal features from inspection images to identify defects, the construction of a component-level full-scale line sample library and identification mechanism based on the reverse identification mode improves the accuracy and detection rate of current defect identification.

[0057] This application also provides a computer device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; The processor, when executing a program stored in memory, implements a reverse identification method for typical defects in overhead transmission lines.

[0058] The communication bus mentioned in the above computer equipment can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not indicate that there is only one bus or one type of bus.

[0059] The communication interface is used for communication between the aforementioned computer equipment and other devices.

[0060] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0061] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0062] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements a method for reverse identification of typical defects in overhead transmission lines.

[0063] The above-described embodiments are preferred embodiments of this application and are not intended to limit the specific scope of this application. The scope of this application includes but is not limited to the specific embodiments described above. All equivalent changes made in accordance with the shape, structure, and method of this application are within the protection scope of this application.

Claims

1. A method for reverse identification of typical defects in overhead transmission lines, characterized in that: Includes the following steps: A component recognition model is constructed based on the visual relationship information between power transmission components; Construct a component feature extraction network based on component region images; Based on normal component image data, a component feature extraction network is used to extract component features to construct a normal component feature library; Based on the component recognition model, real-time component area images are extracted from UAV inspection images, and similarity calculations are performed between these images and a normal component feature library to determine whether the line is abnormal.

2. The method for reverse identification of typical defects in overhead transmission lines according to claim 1, characterized in that: The step of constructing a component recognition model based on visual relationship information between power transmission components includes: Component type labeling is performed on drone inspection images to obtain power transmission component samples; Based on the target detection algorithm, a basic component detection model is trained using power transmission component samples to obtain the basic region features of the power transmission components; Based on the spatial and semantic relationships between components of power transmission equipment, a graph neural network is used to construct a visual relationship model of the components and obtain visual relationship information of the components. Semantic features are extracted based on visual relationship information of components. These semantic features are then fused with the basic regional features of power transmission components. A component recognition model is then constructed based on the fused features.

3. The method for reverse identification of typical defects in overhead transmission lines according to claim 2, characterized in that: The types of components include at least poles, foundations, ancillary facilities, hardware, and insulators.

4. The method for reverse identification of typical defects in overhead transmission lines according to claim 2, characterized in that: The component feature extraction network constructed based on component region images includes: The component region image is segmented into non-overlapping image blocks, and each image block is transformed into a one-dimensional feature vector through linear transformation; The target image patch is obtained by masking and filtering a one-dimensional feature vector based on a random masking algorithm. Based on Transformer and positional encoding, the target image patch is transformed into deep image features containing positional mask markers; The deep image features are mapped back to the original pixel space by the decoder to reconstruct the original pixel values ​​of the filtered image blocks, and the difference between the original pixel values ​​and the real pixels of the part region image is used as a supervision signal to establish a loss function. A self-supervised model is trained based on the loss function to form a component feature extraction network.

5. The method for reverse identification of typical defects in overhead transmission lines according to claim 4, characterized in that: The step of training a self-supervised model based on a loss function to form a component feature extraction network includes: Based on the error between the original pixel values ​​of the reconstructed image patch and the actual pixels of the part region image, the parameters of the encoder and decoder are optimized. The encoder and decoder are trained based on the optimized parameters. When training is complete, the decoder used for image reconstruction is removed, and the encoder used for image feature extraction is retained as the final part feature extraction network.

6. The method for reverse identification of typical defects in overhead transmission lines according to claim 4, characterized in that: The method of extracting component features from normal component image data using a component feature extraction network to construct a normal component feature library includes: Acquire normal component images containing all component types of transmission lines, and filter valid component images based on the component recognition model; The component features of valid component images are extracted based on the component feature extraction network. The component ID is associated and bound with the corresponding feature to form a component ID and feature mapping dictionary containing the normal state features of various components, which is used as the normal component feature library.

7. The method for reverse identification of typical defects in overhead transmission lines according to claim 6, characterized in that: The process of extracting real-time component region images from UAV inspection images based on a component recognition model, calculating their similarity with a normal component feature library, and determining whether the line is abnormal includes: In response to real-time inspection images from the inspection drone, all components are located based on the component recognition model, and the target component area image is cropped. The target component region image is input into the component feature extraction network to generate a real-time component feature vector; Based on the component ID, the corresponding normal feature vector is retrieved from the normal component feature library, and the similarity between the real-time component feature vector and the normal feature vector is calculated. The similarity is compared with a preset similarity threshold. If the similarity is greater than or equal to the similarity threshold, the target component is determined to be normal; if the similarity is less than the similarity threshold, the target component is determined to be abnormal.

8. A reverse identification device for typical defects of overhead transmission lines, applicable to the reverse identification method for typical defects of overhead transmission lines as described in any one of claims 1-7, characterized in that: include: The component identification module is used to construct a component identification model based on the visual relationship information between power transmission components. The feature extraction module is used to construct a component feature extraction network based on component region images; The component feature library module is used to extract component features from normal component image data using a component feature extraction network to construct a normal component feature library. The defect identification module is used to extract real-time component area images from UAV inspection images based on the component identification model, calculate the similarity between these images and a normal component feature library, and determine whether the line is abnormal.

9. A computer device, characterized in that: include: The system includes a processor, a communication interface, a memory, and a communication bus. The processor, communication interface, and memory communicate with each other via the communication bus. The memory stores computer programs. When the processor executes the program stored in the memory, it implements the steps of the reverse identification method for typical defects of overhead transmission lines as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a reverse identification method for typical defects in overhead transmission lines as described in any one of claims 1-7.

Images