Power transmission tower foundation image anomaly detection method and device, computer device and program product

Abstract

This application relates to an anomaly detection method, apparatus, computer equipment, computer-readable storage medium, and computer program product for transmission tower base images. The method includes: performing target detection on the transmission tower base image to generate a graph node set containing component nodes and anomaly nodes; obtaining an initial node feature vector for each graph node; establishing directed edges between component nodes and anomaly nodes to construct a directed topology graph; performing graph convolution processing based on the directed topology graph and the initial node feature vector of each graph node to obtain a higher-order feature vector for each graph node; calculating a topology matching degree for each pair of component nodes and anomaly nodes connected by directed edges to filter anomaly nodes in the directed topology graph; and determining the effective anomaly detection result of the transmission tower base image based on the higher-order feature vectors of the remaining anomaly nodes after filtering, which can improve the accuracy of transmission tower base anomaly detection.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to an anomaly detection method, apparatus, computer equipment, computer-readable storage medium, and computer program product for images of power transmission tower bases. Background Technology

[0002] With the continuous expansion of the power grid, the safe and stable operation of transmission tower foundations is of paramount importance. In order to ensure the safety and stability of transmission tower foundations, images of transmission tower foundations can be collected through inspections, and anomalies in the transmission tower foundation scene can be identified and detected. Anomalies in the transmission tower foundation scene include, but are not limited to, soil erosion on the foundation surface, backfill subsidence, retaining wall collapse, protective soil landslide, water accumulation in the tower foundation, and vine entanglement.

[0003] Traditional technologies typically use YOLO series detectors for anomaly detection and identification, but these methods suffer from low detection accuracy. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, computer equipment, computer-readable storage medium, and computer program product for anomaly detection of transmission tower base images, which can improve the accuracy of anomaly detection results of transmission tower base images.

[0005] In a first aspect, this application provides an anomaly detection method for images of power transmission tower bases, including:

[0006] Acquire an image of the transmission tower base to be detected, perform target detection on the image of the transmission tower base, and obtain the detection results of the tower base components and the initial anomaly detection results corresponding to the image of the transmission tower base;

[0007] Component nodes are generated based on the tower base component detection results, and abnormal nodes are generated based on the tower base anomaly detection results, to form a graph node set;

[0008] Extract the local visual features corresponding to each graph node in the graph node set, and generate an initial node feature vector for each graph node based on the local visual features;

[0009] Based on the spatial distance between component nodes and abnormal nodes in the graph node set, directed edges are established between component nodes and abnormal nodes to construct a directed topology graph.

[0010] Graph convolution is performed based on the directed topological graph and the initial node feature vector of each graph node to obtain the high-order feature vector of each graph node.

[0011] For each pair of component nodes and abnormal nodes connected by directed edges in the directed topology graph, the topology matching degree is calculated based on the corresponding higher-order feature vector.

[0012] The abnormal nodes in the directed topology graph are filtered based on the calculated topology matching degree, and the effective anomaly detection results of the transmission tower base image are determined based on the high-order feature vectors of the remaining abnormal nodes after filtering.

[0013] In some exemplary embodiments, the step of performing graph convolution processing based on the directed topological graph and the initial node feature vector of each graph node to obtain the high-order feature vector of each graph node includes:

[0014] Based on the feature similarity and spatial distance between graph nodes connected by directed edges in the directed topological graph, the edge weights of the directed edges are calculated to construct an adjacency weight matrix.

[0015] Based on the adjacency weight matrix, the initial node features of the graph nodes in the directed topology graph are propagated in the directed topology graph to obtain the high-order feature vector of each graph node.

[0016] In some exemplary embodiments, the step of calculating the topology matching degree based on the corresponding higher-order feature vector for each pair of component nodes and abnormal nodes connected by directed edges in the directed topology graph includes:

[0017] For each pair of component nodes and abnormal nodes connected by directed edges in a directed topology graph, obtain the high-order feature vectors, position information, and bounding box information of the component nodes and abnormal nodes.

[0018] The feature similarity component is calculated based on the higher-order feature vector, the spatial distance attenuation component is calculated based on the location information, and the bounding box overlap component is calculated based on the bounding box information.

[0019] Based on the feature similarity component, the spatial distance attenuation component, and the spatial distance attenuation component, the topology matching degree between the component node and the abnormal node is determined.

[0020] In some exemplary embodiments, determining the effective anomaly detection result of the transmission tower base image based on the high-order feature vectors of the remaining abnormal nodes after filtering includes:

[0021] For the remaining abnormal nodes after filtering, the high-order feature vectors of the abnormal nodes are input into the variational information bottleneck encoder.

[0022] The gate signal value is calculated based on the distributed parameters output by the variational information bottleneck encoder;

[0023] When the gate signal value is greater than the preset routing threshold, the large language model is invoked to perform anomaly detection and obtain a valid anomaly detection result.

[0024] In some exemplary embodiments, the step of calling a large language model to perform anomaly detection on the transmission tower base image to obtain effective anomaly detection results includes:

[0025] Obtain the local image patch corresponding to the abnormal node, and based on the local image patch, perform inference through a large language model to output the original log probability vector of each category corresponding to the local image patch.

[0026] Each element in the original log-odds vector is converted into a non-negative evidence quantity. Based on the non-negative evidence quantity, the Dirichlet distribution parameters are constructed to calculate the total evidence quantity. Based on the total evidence quantity, the prediction probability of each category and the cognitive uncertainty measure that characterizes the cognitive ignorance of the model are calculated respectively.

[0027] Based on the cognitive uncertainty metric and the predicted probability, the effective anomaly detection result of the abnormal node is determined.

[0028] In some exemplary embodiments, the target detection of the transmission tower base image includes:

[0029] Spatial and frequency domain features are extracted from the image of the transmission tower base, and the spatial and frequency domain features are fused to obtain global fused features.

[0030] Target detection is performed based on the global fusion features to obtain the tower base component detection results and initial anomaly detection results corresponding to the transmission tower base image.

[0031] In some exemplary embodiments, the step of performing target detection based on the global fusion features to obtain the tower base component detection results and initial anomaly detection results corresponding to the transmission tower base image includes:

[0032] The global fusion features are expanded into one-dimensional sequence data in a preset direction;

[0033] Local context extraction and nonlinear mapping are performed on the one-dimensional sequence data to generate dynamic step size parameters. The dynamic step size parameters are then used to discretize the preset state space model and update the hidden state to obtain state transition features.

[0034] The one-dimensional sequence data is nonlinearly mapped to generate a gated signal;

[0035] The state transition features are fused with the gating signal through element-wise multiplication. The updated image features are output based on the fusion result. Based on the updated image features, the tower base component detection results and initial anomaly detection results corresponding to the transmission tower base image are obtained.

[0036] Secondly, this application also provides an anomaly detection device for images of transmission tower bases, comprising:

[0037] The target detection module is used to acquire the image of the transmission tower base to be detected, perform target detection on the image of the transmission tower base, and obtain the detection results of the tower base components and the initial anomaly detection results corresponding to the image of the transmission tower base.

[0038] The graph node set generation module is used to generate component nodes based on the tower base component detection results and to generate abnormal nodes based on the tower base anomaly detection results, so as to form a graph node set;

[0039] The initial node feature vector generation module is used to extract the local visual features corresponding to each graph node in the graph node set, and generate an initial node feature vector for each graph node based on the local visual features.

[0040] A directed topology graph construction module is used to establish directed edges between component nodes and abnormal nodes based on the spatial distance between component nodes and abnormal nodes in the graph node set, so as to construct a directed topology graph.

[0041] The high-order feature vector generation module is used to perform graph convolution processing based on the directed topological structure graph and the initial node feature vector of each graph node to obtain the high-order feature vector of each graph node.

[0042] The topology matching degree calculation module is used to calculate the topology matching degree based on the corresponding high-order feature vector for each pair of component nodes and abnormal nodes connected by directed edges in the directed topology graph.

[0043] An abnormal node filtering module is used to filter abnormal nodes in the directed topology graph based on the calculated topology matching degree, and to determine the effective anomaly detection result of the transmission tower base image based on the high-order feature vector of the remaining abnormal nodes after filtering.

[0044] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the anomaly detection method for transmission tower base images described in the above embodiments.

[0045] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the anomaly detection method for transmission tower base images described in the above embodiments.

[0046] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the anomaly detection method for transmission tower base images described in the above embodiments.

[0047] The aforementioned anomaly detection method, apparatus, computer equipment, computer-readable storage medium, and computer program product for transmission tower base images, after acquiring the transmission tower base image to be detected, can perform target detection on the image, thereby generating component nodes and anomaly nodes, extracting local visual features and converting them into initial node feature vectors, thus obtaining a graph node representation that can characterize global information, enriching the underlying semantic features of anomaly discrimination, and further establishing directed edges based on the spatial distance between component nodes and anomaly nodes to construct a directed topological graph, which can effectively model the spatial constraints between components in the physical rigid structure of the tower. Based on the coupling relationship, graph convolution processing is introduced to fuse high-order spatial information. By leveraging the advantages of graph convolutional networks in mining deep logical relationships between nodes and dynamic topology weight allocation, global context enhancement of local anomaly features is achieved. Finally, the topology matching degree is calculated through high-order feature vectors to filter out abnormal nodes and output effective anomaly detection results. This effectively eliminates false alarm noise caused by complex background interference. Thus, based on the high-order feature vectors of the remaining abnormal nodes after filtering, the effective anomaly detection results of the transmission tower base image are determined, which can significantly improve the accuracy and robustness of transmission tower base anomaly detection in complex environments. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is an application environment diagram of the anomaly detection method for transmission tower base images in some embodiments;

[0050] Figure 2 This is a flowchart illustrating an anomaly detection method for transmission tower base images in some embodiments;

[0051] Figure 3 This is a flowchart illustrating the steps for calculating the topology matching degree in some embodiments;

[0052] Figure 4 This is a flowchart illustrating the steps for calling a large language model to perform anomaly detection in some embodiments.

[0053] Figure 5 This is a structural block diagram of an anomaly detection device for transmission tower base images in some embodiments;

[0054] Figure 6These are internal structural diagrams of the computer device in some embodiments;

[0055] Figure 7 This is an internal structural diagram of a computer device in some other embodiments. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0057] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0058] The anomaly detection method for transmission tower base images provided in this application embodiment can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or located in the cloud or on other network servers. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, drones, low-altitude aircraft, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, projection devices, etc. Portable wearable devices can include smartwatches, smart bracelets, head-mounted devices, etc. Head-mounted devices can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. Server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0059] The anomaly detection method for transmission tower base images in this application can be executed by terminal 102 or server 104. Taking server 104 as an example, server 104 can acquire the transmission tower base image to be detected, perform target detection on the transmission tower base image, obtain the tower base component detection results and initial anomaly detection results corresponding to the transmission tower base image, generate component nodes based on the tower base component detection results, and generate anomaly nodes based on the tower base anomaly detection results to form a graph node set. Local visual features corresponding to each graph node in the graph node set are extracted, and an initial node feature vector for each graph node is generated based on the local visual features. The spatial distance between component nodes and abnormal nodes in the set is used to establish directed edges between component nodes and abnormal nodes to construct a directed topology graph. Graph convolution is performed based on the directed topology graph and the initial node feature vector of each graph node to obtain the high-order feature vector of each graph node. For each pair of component nodes and abnormal nodes connected by directed edges in the directed topology graph, the topology matching degree is calculated based on the corresponding high-order feature vector. Abnormal nodes in the directed topology graph are filtered based on the calculated topology matching degree. The effective anomaly detection results of the transmission tower base image are determined based on the high-order feature vectors of the remaining abnormal nodes after filtering.

[0060] In one exemplary embodiment, such as Figure 2 As shown, an anomaly detection method for transmission tower base images is provided, which can be applied to... Figure 1 Taking server 104 as an example, the following steps are included:

[0061] Step 202: Obtain the image of the transmission tower base to be detected, perform target detection on the transmission tower base image, and obtain the detection results of the tower base components and the initial anomaly detection results corresponding to the transmission tower base image.

[0062] The images to be detected, specifically the images of the transmission tower bases, are those requiring anomaly detection. The images of the transmission towers are taken directly from the towers. The tower base component detection results can include information such as the location, type, size, and corresponding local visual features of the tower base components. Initial anomaly detection results can include information such as the location, type, size, and corresponding local visual features of the anomalies. Tower base component types can include, for example, guy wires, cement poles, tower poles, tower bases, tower feet, and drainage ditches. Anomaly types can include, but are not limited to, soil erosion on the base surface, backfill settlement, retaining wall collapse, protective soil landslides, water accumulation at the tower base, and vine entanglement.

[0063] Specifically, the server can obtain images of the transmission tower to be detected from the terminal or other computer devices. These images can be obtained, for example, through drone inspections. The server can further perform target detection on the transmission tower base images to obtain one or more tower base component detection results and one or more initial anomaly detection results corresponding to the transmission tower base images.

[0064] In practical implementation, the server can use a trained object detection model to perform anomaly detection on the image to be detected, obtaining the detection results of the tower base components and the initial anomaly detection results corresponding to the transmission tower base image. The trained object detection model can be trained using supervised training on image samples labeled with the location and category of tower base components, as well as the location and type of anomalies. The model architecture used can be, for example, the YOLO model, or other models capable of object detection; this application does not limit the model architecture.

[0065] Step 204: Generate component nodes based on the tower base component detection results, and generate abnormal nodes based on the tower base abnormality detection results, to form a graph node set.

[0066] Specifically, the server can map the location information of each tower base component in the tower base component detection results to a component node, and map the location information of each abnormality in the tower base anomaly detection results to an abnormality node. All component nodes and abnormality nodes form a graph node set, which is used as the data foundation for subsequent construction of a directed topology graph.

[0067] Step 206: Extract the local visual features corresponding to each graph node in the graph node set, and generate the initial node feature vector for each graph node based on the local visual features.

[0068] Among them, the local visual features corresponding to the graph nodes refer to the local visual features corresponding to the tower base components or anomalies represented by the graph nodes.

[0069] Specifically, the server can extract the local visual features corresponding to each graph node from the feature map of the transmission tower base image, and then generate the initial node feature vector of each graph node based on the local visual features of each graph node.

[0070] In the specific implementation, for each graph node, the local visual features, the position information and size information corresponding to the graph node can be concatenated to obtain the initial node feature vector of the graph node.

[0071] Step 208: Based on the spatial distance between component nodes and abnormal nodes in the graph node set, establish directed edges between component nodes and abnormal nodes to construct a directed topology graph.

[0072] Spatial distance refers to the geometric distance between the center points of two nodes in an image coordinate system or a real physical coordinate system. Specifically, spatial distance can include Euclidean distance or Manhattan distance; for example, calculating the straight-line pixel distance between the center coordinates of an anomaly node and the center coordinates of a component node. A directed edge is a line in a graph structure that connects two nodes with a clear direction, used to represent specific logical or dependent relationships between nodes. Specifically, a directed edge represents a unidirectional data transmission path from a source node to a target node; for example, establishing an edge from a component node to an anomaly node indicates that the anomaly is attached to that component. A directed topology graph is a mesh data structure composed of a set of graph nodes and directed edges between nodes. A directed topology graph can reflect the spatial location associations and interrelationships between transmission tower base components and anomalies.

[0073] For example, the server can connect node pairs with a spatial distance of less than a set threshold, starting from component nodes and ending with abnormal nodes, to form a complete network diagram, thereby obtaining a directed topology diagram.

[0074] Step 210: Perform graph convolution processing based on the directed topological structure graph and the initial node feature vector of each graph node to obtain the high-order feature vector of each graph node.

[0075] Among them, graph convolution processing is a graph neural network information transmission mechanism that integrates the features of a node itself with the features of its neighboring nodes.

[0076] Specifically, the server can perform graph convolution processing based on the directed topology graph and the initial node feature vector of each graph node to obtain the high-order feature vector of each graph node.

[0077] For example, the server can input a directed topology graph and the initial node feature vector of each graph node into a graph convolutional network model. In each layer of the network, the node's own feature vector is weighted and summed or nonlinearly transformed with the feature vectors of its neighboring nodes connected by directed edges.

[0078] Step 212: For each pair of component nodes and abnormal nodes connected by directed edges in the directed topology graph, calculate the topology matching degree based on the corresponding high-order feature vectors.

[0079] Step 214: Filter out abnormal nodes in the directed topology graph based on the calculated topology matching degree, and determine the effective anomaly detection results of the transmission tower base image based on the high-order feature vectors of the remaining abnormal nodes after filtering.

[0080] For each pair of component nodes and anomaly nodes connected by a directed edge in a directed topology graph, the server can calculate the topology matching degree based on the corresponding high-order feature vectors. The topology matching degree is a quantitative metric used to measure the degree of mutual fit or association between the nodes at both ends of a directed edge at the topological and semantic levels.

[0081] Based on the calculated topology matching degree, the server can filter out abnormal nodes in the directed topology graph. Filtering refers to the process of removing false or redundant data according to preset conditions. Specifically, filtering involves setting a threshold, classifying abnormal nodes with a topology matching degree below the threshold as false detections, and removing them from the graph. For example, if the matching degree threshold is set to 0.5, and the matching degree between an abnormal node and its connected component nodes is only 0.2, the server will filter it out as a false detection. The server can then determine the valid anomaly detection results for the transmission tower base image based on the high-order feature vectors of the remaining abnormal nodes after filtering.

[0082] In the above-mentioned anomaly detection method for transmission tower base images, after acquiring the transmission tower base image to be detected, target detection can be performed on the image to generate component nodes and anomaly nodes. Local visual features are extracted and converted into initial node feature vectors, thus obtaining a graph node representation that can characterize global information, enriching the underlying semantic features of anomaly discrimination. Furthermore, directed edges are established based on the spatial distance between component nodes and anomaly nodes to construct a directed topology graph, which can effectively model the spatial constraints and coupling relationships between components in the physical rigid structure of the tower. On this basis, graph convolution processing is introduced to fuse high-order spatial information. By utilizing the advantages of graph convolutional networks in mining deep logical relationships between nodes and dynamic topology weight allocation, global context enhancement of local anomaly features is achieved. Finally, the topology matching degree is calculated through high-order feature vectors to filter anomaly nodes and output effective anomaly detection results, effectively eliminating false alarm noise caused by complex background interference. Thus, the effective anomaly detection results of the transmission tower base image are determined based on the high-order feature vectors of the remaining anomaly nodes after filtering, which can significantly improve the accuracy and robustness of transmission tower base anomaly detection in complex environments.

[0083] In some exemplary embodiments, graph convolution is performed based on the directed topological graph and the initial node feature vector of each graph node to obtain the high-order feature vector of each graph node, including:

[0084] Based on the feature similarity and spatial distance between graph nodes connected by directed edges in a directed topological graph, the edge weights of the directed edges are calculated to construct an adjacency weight matrix.

[0085] Based on the adjacency weight matrix, the initial node features of the graph nodes in the directed topological structure graph are propagated in the directed topological structure graph to obtain the high-order feature vector of each graph node.

[0086] In a graph structure, edge weight is a numerical parameter used to control the strength of information transmission between nodes. Specifically, edge weight refers to a non-negative real value assigned to a directed edge in a directed topological graph, calculated by combining feature similarity and spatial distance. For example, if two nodes are more similar in features and closer in space, the server can calculate a higher edge weight, and vice versa.

[0087] An adjacency weight matrix is ​​a two-dimensional data structure used to represent the connectivity and strength between all nodes in a graph. Specifically, it's a square matrix where rows and columns represent nodes in the graph, and each element corresponds to the weight of the directed edge between a node in that row and a node in that column. For example, if there are N nodes in a graph, the server can generate an N x N matrix. The element in the i-th row and j-th column represents the weight of the directed edge from node i to node j; if the two nodes are not connected, the corresponding element is 0.

[0088] Specifically, for each pair of graph nodes connected by directed edges in a directed topological graph, the server can obtain the initial feature vectors of the two nodes and calculate the feature similarity between them. At the same time, the server can obtain the position information of the two nodes in the image coordinate system and calculate the spatial distance between them.

[0089] The server can further calculate the edge weights of directed edges based on the calculated feature similarity and spatial distance. Information transmission in graph neural networks should be prioritized; high feature similarity indicates a strong semantic connection, while close spatial distance indicates physical adjacency. Therefore, the server can use a similarity multiplied by a distance decay factor, or a pre-defined weight calculation function, to fuse these two dimensions into a single edge weight value. The server can assign the calculated edge weights to the corresponding directed edges. For all directed edges in the directed topological graph, the server can repeat the above calculation process to construct a complete adjacency weight matrix. This matrix comprehensively characterizes the strength of mutual influence between all nodes in the graph.

[0090] The server can process the features of graph nodes based on a pre-constructed adjacency weight matrix. Specifically, the server can distribute the features of a node to its neighboring nodes according to the edge weights recorded in the adjacency weight matrix, while the node also receives features from its neighbors. This process is equivalent to a feature aggregation operation, ensuring that the features of each node are no longer isolated but rather integrated with information from the surrounding topological environment. The server can implement this feature forward propagation through operations such as matrix multiplication. After at least one round of feature propagation and aggregation, a high-order feature vector for each graph node, incorporating neighborhood information, can be obtained.

[0091] To illustrate the above steps more clearly, a specific calculation example is provided below:

[0092] Suppose there exists a directed edge in a directed topological graph, pointing from component node A to anomaly node B. After obtaining the initial feature vectors of nodes A and B, the server calculates their cosine similarity to be 0.8, indicating that they are visually and semantically similar. The server also obtains that the Euclidean distance between the center coordinates of nodes A and B is 50 pixels. The server can calculate the edge weight using a preset function, such as multiplying the similarity by a distance-based exponential decay coefficient, assuming the calculated edge weight is 0.6. The server then fills the position corresponding to row A and column B in the adjacency weight matrix with 0.6.

[0093] Subsequently, the server performs feature propagation based on the adjacency weight matrix. The server can multiply the initial feature vector of node A by the edge weight of 0.6 and add the result to the feature vector of node B. If node B is connected to other component nodes, the server also passes the features of other nodes to node B according to their respective edge weights. Node B sums or averages all the incoming weighted features with its own initial features. Through this propagation process, the high-order feature vector of node B not only contains its own visual features, but also integrates the structural features of node A and other adjacent components. This high-order feature vector, which integrates contextual information, can more accurately reflect whether the abnormal node is truly attached to the normal tower base component, thus providing a solid data foundation for subsequent accurate filtering.

[0094] In some exemplary embodiments, such as Figure 3 As shown, for each pair of component nodes and abnormal nodes connected by directed edges in the directed topology graph, the topology matching degree is calculated based on the corresponding high-order feature vectors, including:

[0095] Step 302: For each pair of component nodes and abnormal nodes connected by directed edges in the directed topology graph, obtain the high-order feature vectors, position information, and bounding box information of the component nodes and abnormal nodes.

[0096] Step 304: Calculate the feature similarity component based on the high-order feature vector, calculate the spatial distance attenuation component based on the location information, and calculate the bounding box overlap component based on the bounding box information.

[0097] Step 306: Determine the topology matching degree between component nodes and abnormal nodes based on feature similarity components, spatial distance attenuation components, and spatial distance attenuation components.

[0098] Location information is a parameter reflecting the orientation of an object within a specific coordinate system. For example, location information can be the numerical values ​​of spatial coordinates in the image pixel coordinate system output by an object detection algorithm. In other words, location information could be the horizontal and vertical coordinates of the center point of the bounding box corresponding to a component node or anomaly node in an image of a power transmission tower base.

[0099] Bounding box information is a geometric description used to define the region where a target object is located in an image. Specifically, bounding box information refers to the size and coordinate parameters of the smallest rectangle output by the object detection model that can select a part or anomalous region in the image. For example, bounding box information may include the coordinates of the top-left and bottom-right corners of the rectangle, or the coordinates of the center point of the rectangle along with its width and height values.

[0100] Feature similarity components are quantitative indicators that measure the degree of consistency between two objects in a feature space. Specifically, a feature similarity component refers to a numerical value calculated by comparing the higher-order feature vectors of two graph nodes, reflecting the degree of semantic association between them. For example, a server can calculate the cosine similarity between the higher-order feature vectors of a component node and the higher-order feature vector of an anomaly node, and use this cosine value as the feature similarity component.

[0101] The spatial distance attenuation component is a numerical value that reflects the degree to which the positional distance between objects affects the correlation relationship. Specifically, the spatial distance attenuation component refers to a weight value that decreases as the distance increases, obtained by calculating the Euclidean distance between two graph nodes based on their positional information in the image and then mapping it through an exponential function. For example, after the server calculates the pixel distance between the center points of two nodes, it substitutes this value into a negative exponential function with distance as the independent variable; the result is the spatial distance attenuation component.

[0102] The bounding box coincidence component is a metric that evaluates the proportion of spatial overlap between two regions. Specifically, the bounding box coincidence component refers to the value obtained by calculating the intersection-union ratio or the percentage of overlapping area of ​​the bounding boxes of two graph nodes based on their bounding box information. For example, the server can calculate the ratio of the intersection area to the union area of ​​the bounding boxes of component nodes and exception nodes, and use this ratio as the bounding box coincidence component.

[0103] Topology matching score is a comprehensive evaluation of the degree of association and fit between two nodes in a graph across multiple dimensions. Specifically, topology matching score refers to a comprehensive index calculated through weighted fusion of multiple components, such as feature similarity, spatial distance decay, and bounding box overlap, to determine whether an anomaly node is truly attached to a component node. For example, the server uses the product or weighted sum of the three components as the topology matching score; a higher score indicates a greater likelihood of an anomaly actually existing.

[0104] The server can traverse all directed edges in the graph. For each directed edge connecting a component node and an abnormal node, the server can obtain the high-order feature vector, position information, and bounding box information of the component node. The server can also synchronously obtain the corresponding data of the abnormal node, and then use the obtained data as the basic input for subsequent calculations.

[0105] The server can calculate the components of the three dimensions based on the acquired data. On the one hand, if an anomaly truly exists on a certain component, then the two should have a high degree of consistency in deep semantic features. The server can quantify this consistency by calculating the cosine similarity of the two high-order feature vectors. On the other hand, since anomalies in transmission tower bases are usually attached to nearby components, the probability of nodes that are too far apart constituting a real anomaly is extremely low. The server can calculate the Euclidean distance between two points and use an exponential decay function to convert it into a weight value that approaches zero as the distance increases. In addition, the server can also calculate the bounding box overlap component based on the bounding box information. Real anomalies often overlap with components in spatial distribution. Therefore, the server can reflect this degree of overlap by calculating the intersection-union ratio of two bounding boxes.

[0106] Based on the calculated three components, the server can determine the topological matching degree between component nodes and abnormal nodes. In specific implementation, the server can fuse the feature similarity component, spatial distance decay component, and bounding box overlap component. The fusion method can be to multiply the three components or to perform a weighted sum according to preset weight coefficients. Through this multi-dimensional comprehensive calculation, the server can output a quantified topological matching degree value, which will serve as the core criterion for subsequent filtering of false abnormal nodes. The data ultimately flows to the abnormal node filtering module to remove false detection anomalies with a matching degree lower than the threshold.

[0107] In the above embodiments, the topology matching degree is calculated by integrating three dimensions: high-order feature similarity, spatial distance decay, and bounding box overlap. This enables a multi-perspective, accurate quantitative assessment of the relationship between anomalies and components, effectively eliminating isolated false anomalies and significantly improving the accuracy of anomaly detection.

[0108] In some exemplary embodiments, determining the effective anomaly detection results of the transmission tower base image based on the high-order feature vectors of the remaining anomalous nodes after filtering includes:

[0109] For the remaining abnormal nodes after filtering, the high-order feature vectors of the abnormal nodes are input into the variational information bottleneck encoder; the gating signal value is calculated based on the distribution parameters output by the variational information bottleneck encoder; when the gating signal value is greater than the preset routing threshold, the large language model is called to perform anomaly detection and obtain effective anomaly detection results.

[0110] Among them, the variational information bottleneck encoder is a model structure for feature compression and information filtering in deep neural networks. Specifically, by introducing variational inference methods, the encoder maps the input data into a potential probability distribution containing mean and variance, thereby filtering out redundant information that is irrelevant to the current task.

[0111] Distribution parameters refer to the set of numerical values ​​output by the variational information bottleneck encoder that define the probability distribution of latent features. For example, distribution parameters can include the mean vector and standard deviation vector of a multidimensional Gaussian distribution.

[0112] The gating signal value is a scalar value calculated based on the distribution parameters and used to measure the uncertainty or information content of a sample. For example, the server can calculate the function value or divergence of the distribution parameters to obtain a probability value between 0 and 1, thereby determining whether the abnormal sample needs to be sent to a more powerful large speech model for further analysis.

[0113] Valid anomaly detection results refer to the anomaly detection results generated after filtering. For example, this result specifically includes data such as the accurate category of the anomaly, its precise location in the image, and the confidence score given by the model. For instance, this result could clearly indicate that an anomaly of a foundation column crack has occurred at a certain location in an image of a transmission tower foundation, along with the bounding box coordinates of the crack and a confidence score of 0.95.

[0114] The server can acquire the remaining anomalous node data after filtering in the preceding steps, and input the high-order feature vectors of these anomalous nodes into the variational information bottleneck encoder. Based on this, the variational information bottleneck encoder compresses and maps the input high-order feature vectors. During this process, by constraining the mutual information of potential representations, the variational information bottleneck can remove background noise from the features, retaining only the features most relevant to anomaly detection, and thus output the distribution parameters corresponding to the feature vector. The server can acquire these distribution parameters and calculate the gating signal value based on them. This calculation process aims to quantify the uncertainty or decision ambiguity of the anomalous sample.

[0115] The server can compare the calculated gating signal value with a preset routing threshold to dynamically allocate computing resources. When the gating signal value is greater than the preset routing threshold, it indicates that the abnormal sample is a difficult sample to determine, and the server can call a large language model for anomaly detection.

[0116] In the above embodiments, the gating signal is calculated by using variational information bottleneck, and the large language model is called only for difficult samples with high ambiguity, which effectively saves computing resources. At the same time, the powerful reasoning ability of the large model is used to overcome difficult anomalies, significantly improving the overall accuracy of anomaly detection.

[0117] In some exemplary embodiments, such as Figure 4 As shown, a large language model is used to perform anomaly detection on the transmission tower base image, and effective anomaly detection results are obtained, including:

[0118] Step 402: Obtain the local image patch corresponding to the abnormal node, and infer based on the local image patch and the large language model to output the original log probability vector of each category corresponding to the local image patch.

[0119] Step 404: Convert each element in the original log-odds vector into non-negative evidence quantity, calculate the total evidence quantity based on the non-negative evidence quantity, and calculate the prediction probability of each category and the cognitive uncertainty measure that represents the cognitive ignorance of the model based on the total evidence quantity.

[0120] Step 406: Based on the cognitive uncertainty metric and the predicted probability, determine the effective anomaly detection results of the abnormal nodes.

[0121] Specifically, the original log-odds vector refers to the vector composed of real numbers output by the large language model's classification head, reflecting the relative probability of each category; its numerical range is unrestricted. The non-negative evidence quantity refers to a non-zero value obtained by mapping the original log-odds vector through an activation function, representing the number of features extracted from the input image that support the existence of a certain type of anomaly. The total evidence quantity specifically refers to the sum of the Dirichlet distribution parameters of all categories, determining the concentration of the probability distribution. The predicted probability specifically refers to a value between 0 and 1 obtained by dividing the Dirichlet distribution parameter of a single category by the total evidence quantity.

[0122] Cognitive uncertainty measure is a quantitative indicator that reflects the ambiguity in decision-making caused by a lack of knowledge in the model. For example, the cognitive uncertainty measure can be the value obtained by dividing the total number of categories by the total amount of evidence, specifically used to distinguish between "ambiguity due to unknown factors" and "ambiguity inherent in the data itself." For instance, if there are 6 categories and the total amount of evidence is 12, the measure would be 0.5; a smaller value indicates that the model has more sufficient clues.

[0123] The server can crop the original image based on the bounding box coordinates to obtain the local image patch corresponding to the abnormal node. Then, the local image patch can be sent to the large language model in the cloud for inference analysis. After receiving the image data, the large language model can perform multimodal feature extraction and logical reasoning, and finally output the original log probability vector of each type of abnormality corresponding to the local image patch.

[0124] The server can transform the acquired raw log-odds vector. For example, the server can use activation functions such as Softplus to convert each element in the raw log-odds vector into non-negative evidence, and construct Dirichlet distribution parameters based on the non-negative evidence. Specifically, the server can add a constant 1 to the evidence for each category to construct a subjective logical framework. The server can further calculate the sum of all category parameters to obtain the total evidence.

[0125] The server can calculate the predicted probability of each category based on the total amount of evidence, as well as the cognitive uncertainty measure that represents the degree of cognitive ignorance of the model. In this way, the black box model that originally only provides a single category output can be transformed into a transparent probability model that can output both probability and uncertainty at the same time.

[0126] The server can determine the valid anomaly detection result for anomaly nodes based on the cognitive uncertainty metric and the predicted probability. For example, the server can set an uncertainty threshold to determine the reliability of the large language model's inference. If the cognitive uncertainty metric is below the threshold and a certain category has the highest predicted probability, the server can directly determine that highest-probability category as the valid anomaly detection result. If the cognitive uncertainty metric is above the threshold, it indicates that the model cannot accurately determine the anomaly due to a lack of evidence. In this case, the server can trigger other strategies, such as marking it as requiring manual review or returning an "unknown anomaly" prompt. Through this mechanism, the server not only outputs the anomaly category but also provides a quantitative assessment of the reliability of the detection result, ensuring high reliability of the final result.

[0127] In the above embodiments, the total amount of evidence is calculated by using the log odds output by the large language model, thereby achieving joint quantification of prediction probability and cognitive uncertainty. This can effectively identify fuzzy decisions caused by insufficient evidence in the model, and significantly improve the credibility and overall accuracy of anomaly detection results.

[0128] In some exemplary embodiments, target detection of transmission tower base images includes: extracting spatial and frequency domain features from the transmission tower base images, fusing the spatial and frequency domain features to obtain global fused features; and performing target detection based on the global fused features to obtain tower base component detection results and initial anomaly detection results corresponding to the transmission tower base images.

[0129] Spatial domain features are visual information that characterizes the local structural morphology and contextual relationships of an image in pixel space. Specifically, spatial domain features refer to feature maps extracted by performing sliding window convolution calculations on the image pixel plane using a convolutional neural network. These feature maps contain component edges, texture directions, and color gradients.

[0130] Frequency domain features are transform domain information that reflects the periodic variation patterns and fine-grained texture of an image's frequency distribution. Specifically, frequency domain features refer to the spectral energy characteristics extracted after performing a two-dimensional Fourier or wavelet transform on an image, representing high-frequency edge abrupt changes and low-frequency smooth regions.

[0131] Global fusion features are comprehensive feature representations that integrate multi-dimensional visual information to fully characterize image content. Specifically, global fusion features refer to feature vectors or feature maps that, after processing by a feature fusion network, contain both spatial structural semantics and frequency domain texture details, and possess a global receptive field.

[0132] For example, the server can process an image using two parallel feature extraction branches. One branch can be used to extract spatial domain features, and the other branch can be used to extract frequency domain features. During spatial feature extraction, the server can use a deep convolutional network to capture component structure and semantic information in the image. During frequency domain feature extraction, the server can use mathematical transformations to convert the image to the frequency domain to capture subtle fine-grained textures and edge abrupt changes that are difficult to detect in the spatial domain. These extracted raw feature data then flow to the fusion processing module.

[0133] For example, considering that simple spatial features are easily affected by lighting and background occlusion, while frequency domain features are sensitive to texture changes but lack overall semantics, the fusion process aims to combine the advantages of both to enhance the expressive power of the features. Therefore, the server can combine the two sets of features along the channel dimension through feature concatenation operations. The server can also introduce a gating mechanism to adaptively adjust the weight ratio of spatial and frequency domain features in the fusion process according to the importance of the features. Through this interactive and integrated processing, the server can obtain globally fused features. At this point, the data has been transformed from a single visual pixel into a high-dimensional feature representation containing rich multi-dimensional semantics.

[0134] For example, the server can input global fused features into the detection head of the object detection network. The detection head can perform classification and regression prediction at each location on the feature map. The server can generate the corresponding bounding box by decoding the predicted offset output by the network. Through this process, the server can simultaneously identify normal structures and suspected abnormal regions in the image. Finally, the server can output the detection results of the tower base components and the initial anomaly detection results corresponding to the transmission tower base image. These results will serve as an important data source for subsequent topology map construction and anomaly filtering.

[0135] In the above embodiments, by extracting and fusing spatial and frequency domain features, the model can simultaneously capture macroscopic structural semantics and microscopic texture details, thereby effectively enhancing the expressive power of features in complex backgrounds, significantly improving the accuracy of target detection, and laying a solid foundation for subsequent accurate anomaly identification.

[0136] In some exemplary embodiments, target detection is performed based on global fusion features to obtain tower base component detection results and initial anomaly detection results corresponding to the transmission tower base image, including:

[0137] The global fusion features are expanded into one-dimensional sequence data in a preset direction; local context extraction and nonlinear mapping are performed on the one-dimensional sequence data to generate dynamic step size parameters, and the preset state space model is discretized and the hidden state is updated using the dynamic step size parameters to obtain state transition features; nonlinear mapping is performed on the one-dimensional sequence data to generate a gating signal; the state transition features and the gating signal are fused element-wise by multiplication, and the updated image features are output based on the fusion result. Based on the updated image features, the tower base component detection results and initial anomaly detection results corresponding to the transmission tower base image are obtained.

[0138] In one specific embodiment, an anomaly detection method for transmission tower base images is provided, which adopts a collaborative architecture based on "UAV edge YOLO-SSM + cloud-based VLS large model". The specific implementation steps are as follows:

[0139] I. Construction of Single Visible Light Data and Spatial-Frequency Dual-Domain Preprocessing

[0140] 1.1 Data Acquisition and Mosaic Enhancement: RGB images of the tower base were acquired using a drone equipped with a high-definition visible light camera (4096×2160 resolution).

[0141] 1.2 Decoupling of Spatial and Frequency Domain Features: To suppress interference from outdoor shadows, the RGB image is divided into 8×8 blocks and subjected to Discrete Cosine Transform (DCT). The spatial branch extracts spatial features through 3×3 convolutions; the frequency branch extracts the full-band coefficient matrix (containing low-frequency and high-frequency components) after the DCT transformation, and performs learnable frequency domain feature reweighting and dimensionality reduction through 1×1 convolutions to obtain frequency domain features. During training, this 1×1 convolutional layer automatically learns the importance weights of each frequency component: for shadow regions, the corresponding low-frequency component weights are optimized to approach 0, thereby effectively suppressing shadow interference; for key structures such as tower material edges and crack textures, the high-frequency and corresponding low-frequency component weights are assigned high response values, thereby enhancing the expression of key features. The two can be fused to obtain fused features.

[0142] During the fusion process, a gated cross-attention mechanism can be used, with dimensional settings... The number of heads, h = 8.

[0143]

[0144] in, Indicates fusion characteristics, Indicates spatial characteristics The query matrix obtained through linear mapping; and They respectively represent the frequency domain features The key matrix and value matrix obtained through linear mapping; Let be the dimension of the key matrix; is the scaling factor; · represents matrix multiplication; ⊙ represents element-wise multiplication (Hadamard product); This represents a feature concatenation operation along the channel dimension; is a learnable gating weight matrix; Sigmoid is the gating activation function, which outputs gating weights in the range of 0 to 1; Softmax is the normalized exponential function used to calculate the attention weight distribution.

[0145] The gating mechanism leverages the complementarity of the frequency and spatial domains, using learnable gating weights. Adaptively adjust the fusion ratio of spatial and frequency domain features at different spatial locations to effectively suppress background noise (such as shadows and uneven lighting) in RGB images, while enhancing the response of key anomalous texture features.

[0146] II. YOLO-SSM Bi-model and Dynamic Graph Inference

[0147] Instead of the traditional Transformer, this embodiment uses the SSM module for processing, which introduces the Mamba Block to achieve O(N) complexity. In the Mamba Block, the input feature map is first normalized by RMSnorm, then divided into two branches: Branch 1 is mapped through a linear layer and input into a one-dimensional causal convolution (Kernel size=4), then passes through the SiLU activation function to obtain the dynamic step size parameter Δ of the state transition matrix, and the output of Branch 1, i.e., the state transition feature, is obtained through the discretized state space recursive formula; Branch 2 is mapped through a linear layer and activated by SiLU, serving as the input to the gating mechanism. The outputs of the two branches are multiplied element-wise (Hadamard product) and then projected through a linear layer. The discretized state space recursive formula is as follows:

[0148]

[0149]

[0150] in, , Δ is the step size parameter, which is dynamically calculated from the input features. This allows the model to adaptively focus on high-frequency areas such as tower base cracks, where A, B, and C are learnable parameters. This refers to state transition characteristics.

[0151] For the linear structure of the tower, the fused features are flattened into a sequence and then serialized by performing original order, horizontal flip, vertical flip, and diagonal flip respectively. These sequences are then input into four parallel Mamba Blocks. The output is then reshaped into a two-dimensional feature map and fused using a 1×1 convolution to obtain the final output features. The formula is:

[0152]

[0153] in, This indicates the output after the feature is serialized in its original order and processed by the SSM module. This represents the output after the features are serialized and then horizontally flipped and fed into the SSM. This represents the output after the features are vertically flipped and serialized, then fed into the SSM. The serialized features are diagonally flipped, flattened, and then fed into the corresponding output of SSM. These are learnable parameters.

[0154] Based on the above final output characteristics This allows us to obtain the tower base component detection result P and the initial anomaly detection result A. Each component bounding box in the detection result is mapped to a component node, and each anomaly bounding box is mapped to an anomaly node, resulting in a graph node set. Further, the RoI (Region of Interest) Align algorithm (pooling size 7×7, feature dimension 256) can be used to extract the features of each box (i.e., the local visual features mentioned above), and the center coordinates of the box (…) can be… , After normalizing the width and height (w, h), the vectors are concatenated with the extracted features of the bounding box to obtain the initial feature vector of each graph node in the graph node set. .

[0155] For the graph node set, more specifically, all detected component nodes and abnormal nodes can be used as vertices of the graph. A sparse graph structure can be constructed using the K-Nearest Neighbors (K-NN) strategy. That is, for each component node, the top-3 abnormal nodes with the closest spatial Euclidean distance are selected and a directed edge is established with them. No connection is established between the remaining nodes to reduce the computational complexity of graph convolution.

[0156] A graph convolutional neural network is used to update the initial feature vectors of nodes in a sparse graph structure. The update formula is as follows:

[0157]

[0158]

[0159] in, , and All are learnable weight matrices (with dimensions d×d, where d is the node feature dimension). and These are the feature vectors of node i and node j, respectively; and These are the normalized center coordinates of node i and node j in the image coordinate system, respectively; Hyperparameters for controlling spatial distance sensitivity; Represents the set of adjacent nodes of node i; The dot (·) indicates the transpose operation after matrix multiplication; the dot (·) indicates scalar multiplication. Let denote the Euclidean norm; ReLU is the activation function; the superscripts (l) and (l-1) denote the layer indices of the graph convolution. Let be the edge weight of the directed edge between node i and node j.

[0160] The dynamic topology graph convolution module consists of two stacked graph convolution layers. Through multi-hop aggregation of neighborhood information, the features of abnormal nodes can fully receive the topological features of their neighboring component nodes, and finally output the high-order feature vector of the graph node. This mechanism forcibly introduces a spatial physical distance penalty (exponential decay term) to ensure that abnormal nodes are necessarily adjacent components.

[0161] III. Topology Filtering

[0162] Calculate the topological matching degree by combining the graph embedding similarity output by DTGC:

[0163]

[0164] in, , These are the bounding boxes for component nodes and abnormal nodes, respectively. Represents intersection, union, and comparison. These represent the high-order feature vectors output after graph convolution for component nodes and abnormal nodes, respectively. and These represent the center coordinates of the bounding boxes of the component node and the exception node, respectively. , , These are preset parameters. Set the overall threshold. Only retain The abnormal nodes are identified, and the valid abnormal nodes are obtained.

[0165] IV. Adaptive Evidence Routing Based on Information Bottleneck (IB)

[0166] 4.1 Variational Information Bottleneck (VIB) Encoder Structure:

[0167] Design a 3-layer Multilayer Perceptron (MLP) as the encoder. The input is the high-order feature vector (dimension 512) of valid outlier nodes. The hidden layer has a dimension of 1024, and the output layer has a dimension of 128 (containing the 64-dimensional mean μ and the 64-dimensional standard deviation). ).

[0168] Employing reparameterization techniques ( Sampling compression characterization Z:

[0169]

[0170] in, It is obtained by random sampling from an independent Gaussian distribution with a mean of 0 and a variance of 1.

[0171] 4.2 Soft Router Gating Functions:

[0172] Calculate the KL divergence:

[0173]

[0174] Design a continuously differentiable soft routing gate function G(x):

[0175]

[0176] Wherein, the temperature coefficient T=0.1, and the routing budget threshold... y represents the predicted probability of the anomaly category corresponding to the anomaly node. When the sample is in the fuzzy decision domain, routing for the large cloud model is triggered. This function is equivalent to an information-theoretic bandpass filter, triggering routing only when the sample is in the fuzzy decision domain, reducing invalid calls by 90%.

[0177] V. Joint Reasoning in Large-Scale Models Based on Subjective Logic (EDL) and VSS

[0178] 5.1. VSS Joint Hint Construction:

[0179] The local crop of the RGB image (size 224×224) is input into the visual encoder of the multimodal large model to extract image feature vectors, which are then combined with spatial topological descriptions to construct a standard JSON format Prompt template.

[0180] {"image_features": "<vision_encoder_output> ", "text": "Spatial prior: Located at a distance of} meters to the lower left of {part name}; Frequency domain characteristics: Significant high-frequency texture; Please determine the anomaly category and confidence level."}

[0181] 5.2 Evidence Collection and Dirichlet Distribution Mapping:

[0182] Lightweight large models (such as Qwen-VL-7B) output Logits vectors, which are then transformed into non-negative evidence using the Softplus activation function.

[0183]

[0184] in, This represents the original output value of the k-th class in the Logits vector. Let be the amount of non-negative evidence for the k-th class.

[0185] Then it is converted into Dirichlet distribution parameters. = +1, and then calculate the total amount of evidence. Thus, the probability is predicted. Cognitive uncertainty (K is the total number of categories; in this example, K=6). When the large model "has never seen this type of feature," S is extremely small, U → 1, accurately quantifying the model's "degree of ignorance."

[0186] Decision logic: If U < 0.3 and If U > 0.6, output the final category; if U ≥ 0.3, it is judged as a sample with high cognitive uncertainty, and an early warning is triggered.

[0187] VI. Edge-to-Cloud Closed-Loop Evolution and System Deployment Based on Active Learning

[0188] 6.1 Hardware Deployment and Communication Architecture:

[0189] The edge deployment uses NVIDIA Jetson Orin modules (275 TOPS of computing power) to run a feature extraction model with two branches, YOLOv26-Mamba, dynamic topology convolutional networks, and variational and variational information bottleneck gating networks, with TensorRT used for INT8 quantization acceleration. The cloud deployment uses an NVIDIA A100 cluster to run a large VSS model. Data transmission between the edge and cloud is encrypted via MQTT protocol over a 5G network, with transmitted data consisting only of cropped images and prompt text, significantly saving bandwidth.

[0190] The training process for the above model is as follows:

[0191] 1. Edge detection and feature extraction;

[0192] The image of the transmission tower base containing component and anomaly annotations is input into the edge detection network to extract local visual features and combine them with the spatial location information of the nodes to generate the initial node feature vectors of component nodes and anomaly nodes.

[0193] 2. Forward propagation of dynamic topological graph convolutional networks;

[0194] Based on the C-KNN strategy, a sparse directed graph topology is constructed according to the spatial distance between nodes;

[0195] For each directed edge in a sparse directed graph, calculate the feature similarity between the connected component nodes and the abnormal nodes, as well as the spatial distance penalty term, dynamically calculate the edge weight of each directed edge, and construct a dynamic adjacency matrix.

[0196] The edge weights are normalized based on the dynamic adjacency matrix, and the features of neighboring nodes are weighted and aggregated according to the normalized weights. The feature update of the current layer graph nodes is completed by the graph convolution recursive formula.

[0197] The above feature update process is stacked and executed in a preset number of layers to achieve multi-hop topology information transmission and output a high-order node feature vector that incorporates physical space topology constraints.

[0198] Based on the high-order node feature vectors, the topology matching degree is calculated for each pair of component nodes and abnormal nodes connected by directed edges in a sparse directed graph.

[0199] The abnormal nodes are filtered based on the calculated topology matching degree. Abnormal nodes with a topology matching degree lower than a preset threshold are removed, and the higher-order node feature vectors of the filtered abnormal nodes are retained.

[0200] 3. Calculation of composite topology loss and information bottleneck loss;

[0201] Calculate the composite topology loss, which includes CIoU loss, Focal Loss loss, and topology penalty loss;

[0202] The topology penalty loss is calculated based on the directed edges of the "component-anomaly" truth values ​​constructed from manually labeled real physical dependencies. The specific calculation method is as follows:

[0203] For the i-th edge in the set of truth-valued directed edges, obtain the bounding boxes between the component nodes and the abnormal nodes connected by this edge, and extract the orientation angle of the component nodes. And the spatial azimuth angle of the abnormal node pointing to the component node ;

[0204] Calculate the cosine of the angle between the two, and base it on a preset angle threshold. Calculate the penalty term for this edge;

[0205] The average of the penalty terms for all truth-valued directed edges is calculated using the following formula:

[0206]

[0207] Where N is the total number of true directed edges, Set to 0.866 (corresponding to a 30-degree angle threshold), this loss forces outlier boxes to actively move closer to the center of the component during gradient descent;

[0208] The high-order node feature vectors of the retained outlier nodes after filtering are input into the variational information bottleneck gating network, which outputs the mean and variance that conform to a Gaussian distribution. The gating signal is calculated based on reparameterized sampling, and the information bottleneck loss, which is a weighted sum of classification cross-entropy loss and KL divergence loss, is calculated as follows:

[0209]

[0210] Where y is the ground truth value of the anomaly category corresponding to the retained anomaly node after filtering, and z is the compressed representation obtained based on reparameterized sampling.

[0211] 4. Joint gradient update of network parameters at the edge;

[0212] The total loss at the edge is constructed by weighting and summing the composite topology loss and the information bottleneck loss.

[0213] Based on the total loss at the edge, the parameters of the edge detection network, the dynamic topology graph convolutional network, and the variational information bottleneck gated network are jointly updated through the backpropagation algorithm until convergence.

[0214] 5. Forward propagation and independent optimization of large multimodal models in the cloud;

[0215] Obtain a large multimodal model from the cloud, and construct multimodal cue words containing visual feature vectors of local image patches and spatial topological text descriptions based on difficult anomaly samples uploaded from the edge routing.

[0216] Input multimodal cue words into a cloud-based multimodal large model, output anomaly categories and confidence text in an autoregressive manner, and calculate the cross-entropy loss of the autoregressive language model;

[0217] Obtain the raw log-odds vector output from the classification head of the large cloud-based model, and convert it into non-negative evidence using the Softplus activation function. And calculate the Dirichlet distribution parameters. And the total amount of evidence, S;

[0218] The formula for calculating Bayesian risk evidence loss is as follows:

[0219]

[0220] Where i is the sample index and y is the category index. One-hot encoding of the real label. For the Digamma function, It is a gamma function;

[0221] Based on the weighted sum of the cross-entropy loss of the autoregressive language model and the risk evidence loss of Bayes, the parameters of the cloud-based multimodal large model are independently updated through the backpropagation algorithm until convergence.

[0222] 6.2 Active Learning Fine-Tuning Strategy:

[0223] A hard sample pool is constructed using samples with high cognitive uncertainty (U ≥ 0.3). The AdamW optimizer (initial learning rate 1e-4, weight decay 1e-4) is used in conjunction with a cosine annealing learning rate strategy to incrementally fine-tune the model at the edge.

[0224] The following examples illustrate this in specific scenarios:

[0225] Assume the following physical scenario: at midday in summer, a drone is conducting a visible light inspection at a height of 30 meters above the tower base. Strong direct sunlight creates a large area of ​​high-contrast shadow on the south side of the tower base. Traditional YOLO models are highly likely to misdetect the dark textures at the edges of the shadow as "retaining wall collapse" or "water accumulation."

[0226] The processing procedure for this solution is as follows:

[0227] Frequency domain suppression: After image input, the 8×8 block DCT frequency domain branch found that the shadow area belongs to the low-frequency smooth signal, while the leaf texture of the vine belongs to the high-frequency signal. Through the spatial frequency gated fusion in step S3, the frequency domain weight of the shadow area is automatically suppressed to below 0.1 by the Sigmoid gate function, while the high-frequency weight of the vine edge is amplified.

[0228] Topology filtering: Model 2 detected a "suspected water accumulation" bounding box (confidence 0.65) and a "suspected vine" bounding box (confidence 0.58) near the tower base. After DTGC graph convolution calculation, it was found that the center point of the "suspected water accumulation" bounding box was located in the tower shadow area 2 meters above the tower base, and its spatial physical distance penalty term approached 0, resulting in a low topology matching score. It was directly removed.

[0229] Large-scale model routing and verification: The "suspected vine" bounding box is in the fuzzy decision domain, triggering the VIB routing gateway. The cloud-based large-scale model receives the Crop image and Prompt: "Spatial prior: located 0.2 meters below and to the left of the tower base; frequency domain features: significant high-frequency texture." The large-scale model combines the Dirichlet distribution to output evidence quantity and calculate cognitive uncertainty. The final diagnosis was "entwined vines," with a confidence level of 0.92.

[0230] The technical effects of the above embodiments are as follows:

[0231] Single-mode anti-interference and improved positioning accuracy: By introducing 8×8 block DCT frequency domain decoupling and Vision-SSM four-way scanning mechanism, interference from complex lighting and shadows in the field is effectively suppressed when using only a single RGB image. Compared with traditional YOLOv8, the positioning error (CIoU index) of slender parts such as tower feet and guy wires is reduced by an average of 18.5%, providing more accurate regions of interest (ROI) for large models.

[0232] Topological robustness and false positive suppression: By employing DTGC dynamic graph convolution and composite topological loss, "abnormalities must be adjacent components" is enforced from both geometric and semantic perspectives. In a test set with complex terrain containing a large amount of irrelevant vegetation and shadows, the false positive rate (FPR) plummeted from 18.5% in the baseline scheme to 3.2%, a reduction of approximately 82%.

[0233] Dynamic allocation of computing power and bandwidth saving based on information theory: A soft routing gating mechanism based on variational information bottleneck (VIB) enables "fast judgment of small models for simple samples and precise judgment of large models for difficult samples". Compared with the full input large model solution, the cloud large model call rate is reduced from 100% to 14.5%, the edge-cloud communication bandwidth usage is reduced by 85%, and the overall system inference latency is compressed from 2850ms to 380ms.

[0234] Cognitive Uncertainty Quantification and Closed-Loop Evolution: By introducing Dirichlet distribution and Softplus evidence mapping, the cognitive uncertainty (U) of large models is accurately quantified. The system not only outputs the results but also the "level of confidence the model has in the result," automatically triggering manual verification for high-uncertainty samples. Through active learning and closed-loop processing, the model's recall rate for rare anomalies (such as special geological landslides) improved by 24% after 6 months of deployment.

[0235] Edge-end adaptability for engineering deployment: Through TensorRT INT8 quantization and Mamba's linear complexity design, complex topology sensing models can be successfully deployed on the edge of drones with limited computing power (Jetson Orin), with a stable frame rate of over 25 FPS, fully meeting the real-time engineering requirements of power line inspection.

[0236] Multi-module physical scenario synergy: The solution in this application addresses two major physical pain points of power transmission tower bases: "interference from strong outdoor light shadows" and "abnormal spatial attachments," through deep customization and synergy. Specifically: 8×8 block DCT frequency domain decoupling filters out low-frequency shadow noise from the bottom layer, providing clean initial node characteristics for DTGC; the spatial distance penalty mechanism of DTGC eliminates more than 80% of false alarms from detached components, significantly reducing the redundancy pressure of subsequent VIB routing; VIB accurately filters out difficult samples in the fuzzy decision domain, avoiding the waste of computational power and illusions in the EDL large model; finally, the EDL large model combined with spatial topology Prompt realizes the quantification of cognitive uncertainty for rare anomalies. Each module is interconnected and indispensable, together forming a closed-loop high-reliability architecture suitable for the inspection of transmission tower base infrastructure.

[0237] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0238] Based on the same inventive concept, this application also provides an anomaly detection device for transmission tower base images to implement the anomaly detection method for transmission tower base images described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the anomaly detection device for transmission tower base images provided below can be found in the limitations of the anomaly detection method for transmission tower base images described above, and will not be repeated here.

[0239] In one exemplary embodiment, such as Figure 5 As shown, an anomaly detection device 500 for images of power transmission tower bases is provided, comprising:

[0240] The target detection module 502 is used to acquire the image of the transmission tower base to be detected, perform target detection on the image of the transmission tower base, and obtain the detection results of the tower base components and the initial anomaly detection results corresponding to the image of the transmission tower base.

[0241] The graph node set generation module 504 is used to generate component nodes based on the tower base component detection results and anomaly nodes based on the tower base anomaly detection results, so as to form a graph node set.

[0242] The initial node feature vector generation module 506 is used to extract the local visual features corresponding to each graph node in the graph node set, and generate the initial node feature vector of each graph node based on the local visual features.

[0243] The directed topology graph construction module 508 is used to establish directed edges between component nodes and abnormal nodes based on the spatial distance between component nodes and abnormal nodes in the graph node set, so as to construct a directed topology graph.

[0244] The high-order feature vector generation module 510 is used to perform graph convolution processing based on the directed topological structure graph and the initial node feature vector of each graph node to obtain the high-order feature vector of each graph node.

[0245] The topology matching degree calculation module 512 is used to calculate the topology matching degree based on the corresponding high-order feature vector for each pair of component nodes and abnormal nodes connected by directed edges in the directed topology graph.

[0246] The abnormal node filtering module 514 is used to filter abnormal nodes in the directed topology graph based on the calculated topology matching degree, and to determine the effective abnormal detection results of the transmission tower base image based on the high-order feature vectors of the remaining abnormal nodes after filtering.

[0247] The aforementioned anomaly detection device for transmission tower base images, after acquiring the image of the transmission tower base to be detected, can perform target detection on the image, thereby generating component nodes and anomaly nodes. It extracts local visual features and transforms them into initial node feature vectors, thus obtaining a graph node representation that can characterize global information, enriching the underlying semantic features of anomaly discrimination. Furthermore, it establishes directed edges based on the spatial distance between component nodes and anomaly nodes to construct a directed topology graph, effectively modeling the spatial constraints and coupling relationships between components in the rigid physical structure of the tower. Based on this, graph convolution processing is introduced to fuse higher-order spatial information. Utilizing the advantages of graph convolutional networks in mining deep logical relationships between nodes and dynamic topology weight allocation, it achieves global context enhancement of local anomaly features. Finally, it calculates the topology matching degree through higher-order feature vectors, filters anomaly nodes, and outputs valid anomaly detection results, effectively eliminating false alarm noise caused by complex background interference. Therefore, based on the higher-order feature vectors of the remaining anomaly nodes after filtering, it determines the valid anomaly detection results for the transmission tower base image, significantly improving the accuracy and robustness of anomaly detection in complex environments.

[0248] In some exemplary embodiments, the higher-order feature vector generation module is further configured to: calculate the edge weights of directed edges based on the feature similarity and spatial distance between graph nodes connected by directed edges in the directed topological graph, so as to construct an adjacency weight matrix; and based on the adjacency weight matrix, propagate the initial node features of the graph nodes in the directed topological graph to obtain the higher-order feature vector of each graph node.

[0249] In some exemplary embodiments, the topology matching degree calculation module is further configured to, for each pair of component nodes and abnormal nodes connected by directed edges in the directed topology graph, obtain the high-order feature vectors, position information, and bounding box information of the component nodes and abnormal nodes; calculate the feature similarity component based on the high-order feature vectors, calculate the spatial distance attenuation component based on the position information, and calculate the bounding box overlap component based on the bounding box information; and determine the topology matching degree between the component nodes and abnormal nodes based on the feature similarity component, the spatial distance attenuation component, and the spatial distance attenuation component.

[0250] In some exemplary embodiments, the abnormal node filtering module is used to: input the high-order feature vector of the remaining abnormal nodes into the variational information bottleneck encoder; calculate the gating signal value based on the distribution parameters output by the variational information bottleneck encoder; and when the gating signal value is greater than a preset routing threshold, call the large language model to perform anomaly detection to obtain an effective anomaly detection result.

[0251] In some exemplary embodiments, the abnormal node filtering module is further configured to: obtain local image patches corresponding to abnormal nodes; perform inference based on the local image patches and a large language model to output the original log-odds vectors of each category corresponding to the local image patches; convert each element in the original log-odds vectors into non-negative evidence quantities; construct Dirichlet distribution parameters based on the non-negative evidence quantities to calculate the total evidence quantity; and calculate the predicted probability of each category and the cognitive uncertainty measure representing the degree of cognitive ignorance of the model based on the total evidence quantity; and determine the effective abnormal detection result of the abnormal node based on the cognitive uncertainty measure and the predicted probability.

[0252] In some exemplary embodiments, the target detection module is further configured to: extract spatial and frequency domain features from the transmission tower base image, and fuse the spatial and frequency domain features to obtain global fused features; perform target detection based on the global fused features to obtain the tower base component detection results and initial anomaly detection results corresponding to the transmission tower base image.

[0253] In some exemplary embodiments, the target detection module is further configured to: expand the global fusion features into one-dimensional sequence data in a preset direction; perform local context extraction and nonlinear mapping on the one-dimensional sequence data to generate dynamic step size parameters, and use the dynamic step size parameters to discretize the preset state space model and update the hidden state to obtain state transition features; perform nonlinear mapping on the one-dimensional sequence data to generate a gating signal; perform element-wise multiplication and fusion of the state transition features and the gating signal, output updated image features based on the fusion result, and output the tower base component detection result and initial anomaly detection result corresponding to the transmission tower base image based on the updated image features.

[0254] Each module in the aforementioned anomaly detection device for transmission tower base images can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0255] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 6As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements an anomaly detection method for images of power transmission tower bases.

[0256] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When the computer program is executed by the processor, it implements an anomaly detection method for images of power transmission tower bases. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0257] Those skilled in the art will understand that Figure 6 , Figure 7The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0258] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the anomaly detection method for transmission tower base images in any of the above embodiments.

[0259] In some embodiments, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the anomaly detection method for transmission tower base images in any of the above embodiments.

[0260] In some embodiments, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the anomaly detection method for transmission tower base images in any of the above embodiments.

[0261] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0262] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0263] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0264] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for anomaly detection in images of power transmission tower bases, characterized in that, The method includes: Acquire an image of the transmission tower base to be detected, perform target detection on the image of the transmission tower base, and obtain the detection results of the tower base components and the initial anomaly detection results corresponding to the image of the transmission tower base; Component nodes are generated based on the tower base component detection results, and abnormal nodes are generated based on the initial abnormality detection results, to form a graph node set; Extract the local visual features corresponding to each graph node in the graph node set, and generate an initial node feature vector for each graph node based on the local visual features; Based on the spatial distance between component nodes and abnormal nodes in the graph node set, directed edges are established between component nodes and abnormal nodes to construct a directed topology graph. Graph convolution is performed based on the directed topological graph and the initial node feature vector of each graph node to obtain the high-order feature vector of each graph node. For each pair of component nodes and abnormal nodes connected by directed edges in the directed topology graph, the topology matching degree is calculated based on the corresponding higher-order feature vector. Abnormal nodes in the directed topology graph are filtered based on the calculated topology matching degree. For the remaining abnormal nodes after filtering, the high-order feature vectors of the abnormal nodes are input into the variational information bottleneck encoder. The gate signal value is calculated based on the distributed parameters output by the variational information bottleneck encoder; When the gate signal value is greater than the preset routing threshold, the large language model is invoked to perform anomaly detection and obtain a valid anomaly detection result.

2. The method according to claim 1, characterized in that, The step of performing graph convolution processing based on the directed topological graph and the initial node feature vector of each graph node to obtain the high-order feature vector of each graph node includes: Based on the feature similarity and spatial distance between graph nodes connected by directed edges in the directed topological graph, the edge weights of the directed edges are calculated to construct an adjacency weight matrix. Based on the adjacency weight matrix, the initial node feature vectors of the graph nodes in the directed topology graph are propagated in the directed topology graph to obtain the higher-order feature vectors of each graph node.

3. The method according to claim 1, characterized in that, For each pair of component nodes and abnormal nodes connected by directed edges in the directed topology graph, the topology matching degree is calculated based on the corresponding higher-order feature vector, including: For each pair of component nodes and abnormal nodes connected by directed edges in a directed topology graph, obtain the high-order feature vectors, position information, and bounding box information of the component nodes and abnormal nodes. The feature similarity component is calculated based on the higher-order feature vector, the spatial distance attenuation component is calculated based on the location information, and the bounding box overlap component is calculated based on the bounding box information. Based on the feature similarity component, the spatial distance attenuation component, and the bounding box overlap component, the topological matching degree between the component node and the abnormal node is determined.

4. The method according to claim 1, characterized in that, The process of calling a large language model to perform anomaly detection on the transmission tower base image, and obtaining effective anomaly detection results, includes: Obtain the local image patch corresponding to the abnormal node, and based on the local image patch, perform inference through a large language model to output the original log probability vector of each category corresponding to the local image patch. Each element in the original log-odds vector is converted into a non-negative evidence quantity. The total evidence quantity is calculated based on the non-negative evidence quantity. The predicted probability of each category and the cognitive uncertainty measure that characterizes the cognitive ignorance of the model are calculated based on the total evidence quantity. Based on the cognitive uncertainty metric and the predicted probability, the effective anomaly detection result of the abnormal node is determined.

5. The method according to claim 1, characterized in that, The target detection of the transmission tower base image includes: Spatial and frequency domain features are extracted from the image of the transmission tower base, and the spatial and frequency domain features are fused to obtain global fused features. Target detection is performed based on the global fusion features to obtain the tower base component detection results and initial anomaly detection results corresponding to the transmission tower base image.

6. The method according to claim 5, characterized in that, The target detection based on the global fusion features, to obtain the tower base component detection results and initial anomaly detection results corresponding to the transmission tower base image, includes: The global fusion features are expanded into one-dimensional sequence data in a preset direction; Local context extraction and nonlinear mapping are performed on the one-dimensional sequence data to generate dynamic step size parameters. The dynamic step size parameters are then used to discretize the preset state space model and update the hidden state to obtain state transition features. The one-dimensional sequence data is nonlinearly mapped to generate a gated signal; The state transition features are fused with the gating signal through element-wise multiplication. The updated image features are output based on the fusion result. Based on the updated image features, the tower base component detection results and initial anomaly detection results corresponding to the transmission tower base image are obtained.

7. An anomaly detection device for images of transmission tower bases, characterized in that, The device includes: The target detection module is used to acquire the image of the transmission tower base to be detected, perform target detection on the image of the transmission tower base, and obtain the detection results of the tower base components and the initial anomaly detection results corresponding to the image of the transmission tower base. The graph node set generation module is used to generate component nodes based on the tower base component detection results and to generate abnormal nodes based on the initial abnormality detection results, so as to form a graph node set. The initial node feature vector generation module is used to extract the local visual features corresponding to each graph node in the graph node set, and generate an initial node feature vector for each graph node based on the local visual features. A directed topology graph construction module is used to establish directed edges between component nodes and abnormal nodes based on the spatial distance between component nodes and abnormal nodes in the graph node set, so as to construct a directed topology graph. The high-order feature vector generation module is used to perform graph convolution processing based on the directed topological structure graph and the initial node feature vector of each graph node to obtain the high-order feature vector of each graph node. The topology matching degree calculation module is used to calculate the topology matching degree based on the corresponding high-order feature vector for each pair of component nodes and abnormal nodes connected by directed edges in the directed topology graph. An abnormal node filtering module is used to filter abnormal nodes in the directed topology graph based on the calculated topology matching degree. For the remaining abnormal nodes after filtering, the high-order feature vector of the abnormal node is input into the variational information bottleneck encoder. The gating signal value is calculated based on the distribution parameters output by the variational information bottleneck encoder. When the gating signal value is greater than the preset routing threshold, the large language model is called to perform anomaly detection to obtain an effective anomaly detection result.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images