Unmanned aerial vehicle power distribution network inspection image real-time recognition method based on artificial intelligence

Abstract

The application discloses an unmanned aerial vehicle power distribution network inspection image real-time identification method based on artificial intelligence, relates to the field of image recognition, and comprises the following steps: performing space-time calibration on a multi-modal data packet, extracting each modal feature, using a cross-modal attention mechanism to fuse the features, obtaining a multi-modal feature vector, constructing a heterogeneous graph through power distribution network topology data, setting nodes and edges, assigning the multi-modal feature vector to the nodes, using a graph convolutional neural network to aggregate neighbor device features, updating the node representation, outputting abnormal classification results and abnormal propagation path prediction results of the power distribution network device, and receiving the abnormal classification results and the abnormal propagation path prediction results by a bandwidth network center, continuously monitoring and training a bandwidth prediction model by using historical network data and environmental factors, and predicting the bandwidth change trend; and the application realizes the capture of complex dependency relationships between devices and the accurate prediction of abnormal propagation paths, and significantly improves the comprehensiveness of inspection and fault prevention capability.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image recognition, in particular to a real-time identification method for unmanned aerial vehicle power distribution network inspection images based on artificial intelligence. BACKGROUND

[0002] As an important part of power system operation and maintenance, power distribution network inspection has gradually evolved from traditional manual inspection to automation and intelligence in recent years with the rapid development of unmanned aerial vehicle technology and artificial intelligence technology. Early power distribution network inspection mainly relied on manual on-site inspection, which was inefficient and posed safety risks. With the advancement of unmanned aerial vehicle technology, unmanned aerial vehicles equipped with visible light cameras were introduced for image acquisition, combined with convolutional neural networks for equipment state analysis, significantly improving the coverage and efficiency of inspection. In addition, infrared thermal imaging and laser radar and other sensing technologies have been gradually applied to unmanned aerial vehicle inspection to capture equipment temperature and geometric information, enhancing the ability to detect abnormalities. The development of artificial intelligence has further promoted the intelligentization of inspection.

[0003] Although the existing technology has made certain progress in unmanned aerial vehicle power distribution network inspection, there are still deficiencies that can be improved, such as traditional image analysis methods focusing on single equipment, ignoring the electrical connection relationship between devices in the power distribution network, and failing to effectively predict the propagation path of abnormalities, such as the potential impact of transformer overheating on adjacent lines, limiting the comprehensiveness of inspection. In addition, existing real-time processing technology is limited by bandwidth fluctuations in weak network environments (such as remote mountainous areas), and the computing power of edge computing is insufficient to support complex models, resulting in high latency or decreased accuracy. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a real-time identification method for unmanned aerial vehicle power distribution network inspection images based on artificial intelligence to solve the problem of ignoring the electrical connection relationship between devices in the power distribution network and failing to effectively predict the propagation path of abnormalities.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] The present application provides a real-time identification method for unmanned aerial vehicle power distribution network inspection images based on artificial intelligence, which comprises,

[0008] The unmanned aerial vehicle generates an inspection path through a path planning algorithm and collects multi-modal data to generate a multi-modal data package;

[0009] The multi-modal data package is subjected to spatio-temporal calibration, and after extracting the features of each mode, the features are fused using a cross-modal attention mechanism to obtain a multi-modal feature vector;

[0010] A heterogeneous graph is constructed using distribution network topology data. Nodes and edges are set, multimodal feature vectors are assigned to nodes, and graph convolutional neural networks are used to aggregate features of neighboring devices, update node representations, and output anomaly classification results and anomaly propagation path prediction results for distribution network devices.

[0011] The bandwidth network center receives anomaly classification results and anomaly propagation path prediction results, and continuously monitors and utilizes historical network data and environmental factors to train a bandwidth prediction model, predict bandwidth change trends, proactively adjust the allocation of computing tasks, and run different graph convolutional neural networks in low and high bandwidth situations to obtain compressed data packets and enhanced inspection reports.

[0012] As a preferred embodiment of the AI-based UAV power distribution network inspection image real-time recognition method described in this invention, the UAV generates an inspection path and collects multimodal data through a path planning algorithm. Generating a multimodal data packet refers to obtaining a power distribution network topology map, determining equipment locations as waypoints based on the power distribution network topology map, generating an inspection path using the A* path planning algorithm, importing the inspection path into the UAV, collecting appearance images, temperature distribution images, 3D point cloud data, timestamps, and location information to obtain a multimodal data packet.

[0013] As a preferred embodiment of the real-time image recognition method for UAV power distribution network inspection based on artificial intelligence described in this invention, the method of generating inspection paths using the A* path planning algorithm refers to defining a cost function based on the Euclidean distance between waypoints and flight priority, and generating inspection paths by minimizing the total cost, starting from the UAV take-off and landing point.

[0014] As a preferred embodiment of the real-time image recognition method for UAV power distribution network inspection based on artificial intelligence described in this invention, the spatiotemporal calibration of multimodal data packets refers to aligning the appearance image, temperature distribution image, and three-dimensional point cloud data in time using a linear interpolation method based on the timestamp, and then projecting the spatial pose transformation matrix onto the imaging coordinate system of the appearance image.

[0015] As a preferred embodiment of the AI-based UAV power grid inspection image real-time recognition method of the present invention, wherein: the extraction of modal features specifically includes,

[0016] The appearance features of the spatiotemporally calibrated appearance image are extracted using a pre-trained ResNet network to generate an appearance feature map;

[0017] A pre-trained dedicated convolutional neural network is used to extract temperature distribution features from the temperature distribution image to generate a temperature distribution feature map;

[0018] Geometric features are extracted from 3D point cloud data using a pre-trained PointNet network to generate a geometric feature map.

[0019] As a preferred embodiment of the AI-based UAV power grid inspection image real-time recognition method of the present invention, wherein: the use of a cross-modal attention mechanism to fuse features to obtain a multimodal feature vector specifically involves,

[0020] The appearance feature map, temperature distribution feature map, and geometric structure feature map are input into the cloud server for cross-modal attention fusion. The feature dimensions are aligned by global average pooling to generate appearance feature vector, temperature distribution feature vector, and laser point cloud feature vector.

[0021] The correlation score between each pair of appearance feature vector, temperature distribution feature vector and laser point cloud feature vector is calculated by the dot product attention mechanism, and the attention weights are generated by applying the softmax function. Based on the attention weights, the feature vectors are weighted and fused to generate multimodal feature vectors.

[0022] As a preferred embodiment of the AI-based UAV-based real-time image recognition method for power distribution network inspection described in this invention, the method involves: constructing a heterogeneous graph using power distribution network topology data, setting nodes and edges, and assigning multimodal feature vectors to the nodes. Specifically...

[0023] Obtain distribution network topology data from the power company's maintenance database. The distribution network topology data includes equipment type, equipment number, and electrical connection relationships.

[0024] Each device is set as a node, and electrical connections are set as edges to construct a basic graph structure;

[0025] Based on the basic graph structure, multimodal feature vectors are assigned to the corresponding nodes by device number, and the edge weights are calculated based on electrical distance and then normalized to generate a heterogeneous graph.

[0026] As a preferred embodiment of the AI-based UAV power distribution network inspection image real-time recognition method of the present invention, the step of using a graph convolutional neural network to aggregate neighbor device features, update node representations, and output anomaly classification results and anomaly propagation path prediction results for power distribution network devices specifically involves:

[0027] The heterogeneous graph is input into a pre-trained graph convolutional neural network. The neighborhood information is aggregated according to the edge weights through multi-layer graph convolution operations, the node feature representation is updated, and the anomaly classification result is output after a fully connected layer and a softmax function.

[0028] The probability of anomaly propagation is calculated based on node and edge weights, and the shortest path algorithm is used to generate the anomaly propagation path prediction results.

[0029] As a preferred embodiment of the AI-based UAV power grid inspection image real-time recognition method of the present invention, the bandwidth network center receives anomaly classification results and anomaly propagation path prediction results, and continuously monitors and utilizes historical network data and environmental factors to train a bandwidth prediction model to predict bandwidth change trends. Specifically,

[0030] Receive the anomaly classification results and anomaly propagation path prediction results, start the bandwidth center to continuously monitor the real-time network status, and generate a unified feature vector by combining historical network data and environmental factors.

[0031] By analyzing the unified feature vector using a pre-trained bandwidth prediction model, the bandwidth change trend can be obtained.

[0032] As a preferred embodiment of the AI-based UAV power grid inspection image real-time recognition method of the present invention, the method of actively adjusting the allocation of computing tasks and running different graph convolutional neural networks under low and high bandwidth conditions to obtain compressed data packets and enhanced inspection reports specifically involves:

[0033] Based on bandwidth change trends and preset bandwidth thresholds, determine low-bandwidth and high-bandwidth scenarios;

[0034] In low-bandwidth scenarios, a subset of heterogeneous graphs is generated based on the heterogeneous graphs and anomaly classification results, and a lightweight graph convolutional network is run to process it, generating preliminary results of anomaly classification, and a compressed data packet is generated through principal component analysis.

[0035] In high-bandwidth scenarios, heterogeneous graph requirements are uploaded to a cloud database, a pre-trained graph convolutional network is run, anomaly classification results and anomaly propagation path prediction results are output, and dynamic risk assessment graphs and 3D visualization models are generated and integrated into an enhanced inspection report.

[0036] The beneficial effects of this invention are as follows: By constructing a heterogeneous graph, the equipment types, numbers, and electrical connection relationships in the distribution network topology data are modeled as nodes and edges, and multimodal feature vectors are assigned. A pre-trained graph convolutional neural network is used to aggregate neighborhood information, outputting anomaly classification results and anomaly propagation path prediction results. This achieves the capture of complex dependencies between equipment and accurate prediction of anomaly propagation paths, significantly improving the comprehensiveness of inspections and fault prevention capabilities, overcoming the limitations of traditional image analysis that only focuses on a single device. Simultaneously, by monitoring real-time network status and adaptively allocating computational tasks, efficient data processing that dynamically adapts to bandwidth fluctuations is achieved, significantly reducing latency in weak network environments and improving edge computing efficiency, thus enhancing the adaptability and real-time performance of inspections in complex environments. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Fig. 1 A flowchart of an AI-based method for real-time image recognition in power distribution network inspection using drones;

[0039] Fig. 2 This is a schematic diagram of cross-modal feature fusion.

[0040] Fig. 3 This is a schematic diagram of heterogeneous graph processing;

[0041] Fig. 4 A schematic diagram of bandwidth adaptive task allocation. Detailed Implementation

[0042] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0043] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0044] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0045] Reference Figs. 1-4 This is one embodiment of the present invention, which provides a real-time image recognition method for UAV power distribution network inspection based on artificial intelligence, including the following steps:

[0046] S1: The UAV generates an inspection path and collects multimodal data through a path planning algorithm, generating a multimodal data packet;

[0047] Specifically, the steps include the following:

[0048] S1.1: Obtain the power grid topology map provided by the power company. The topology map includes the geographical coordinates, equipment type, and electrical connections of all equipment to be inspected. Based on the topology map, determine the locations of all equipment to be inspected and use these locations as waypoints for route planning.

[0049] The A* path planning algorithm is used to optimize and sort waypoints. Running in a two-dimensional geographic space, the A* algorithm takes the UAV's take-off and landing point as the starting point and comprehensively considers the Euclidean distance between waypoints and flight priority, prioritizing coverage of equipment with high load or historical failure rates to generate a globally optimal and continuous flight path. Specifically:

[0050] For any two waypoints, define a cost function to represent the actual flight cost between them, expressed as:

[0051]

[0052] The A* algorithm searches for the optimal path by minimizing the total cost: Starting from the UAV's take-off and landing point, it initializes an open list (storing waypoints to be explored) and a closed list (storing explored waypoints). From the take-off and landing point, the A* algorithm iteratively performs the following operations: selecting the waypoint with the lowest total cost in the open list as the current node; calculating the cost from the current node to neighboring waypoints and updating the total cost; if a neighboring waypoint is a high-priority device (such as a high-load transformer), reducing the cost of the high-priority device through priority weights and adding it to the path first; moving the current node to the closed list, repeating this process until all waypoints have been visited, generating a continuous flight path.

[0053] Path optimization considers UAV flight constraints, such as maximum range and obstacle avoidance requirements. Obstacles are detected using ultrasonic sensors, and the flight trajectory between waypoints is dynamically adjusted. The A* algorithm is then used to replan local paths, ensuring path continuity and safety. Priority is given to covering high-load or frequently faulty equipment (such as transformer T1 and insulator I2) to maximize inspection efficiency.

[0054] To further explain, the A* path planning algorithm, combined with power distribution network topology data and equipment priority weights, generates a globally optimal and continuous flight path. This prioritizes coverage of power distribution network equipment with high loads or historically high failure rates, ensuring that the UAV efficiently inspects critical equipment within a limited range. Compared to traditional path planning methods based solely on geographical distance, this method significantly improves the targeting of inspections, reduces redundant flight time, and optimizes energy consumption and mission duration through priority adjustment coefficients and weighting mechanisms.

[0055] S1.2: Import the generated flight path into the UAV flight control center. The UAV is equipped with a high-resolution visible light camera, an infrared thermal imaging camera, and a lidar, as well as an RTK-GPS center and an inertial measurement center. Before the mission begins, complete the synchronous calibration of each sensor with the positioning center to ensure consistency of time and space references during data acquisition.

[0056] The drone flies autonomously along the generated flight path. During flight, the RTK-GPS center provides high-precision position information in real time, and the inertial measurement center records the flight attitude to ensure flight stability. When the drone approaches each waypoint, it triggers simultaneous acquisition of multimodal data: a visible light camera captures images of the equipment's exterior, an infrared thermal imaging camera records the equipment's temperature distribution, and a lidar scans the equipment's structure and generates point cloud data.

[0057] Each frame of acquired appearance image, temperature distribution, and point cloud data is appended with a precise timestamp and current location information. The timestamp is generated uniformly by the airborne main control center, and the location information comes from the RTK-GPS center output. All appearance images, temperature distribution, and point cloud data are organized in time sequence in the airborne memory, forming data units containing appearance images, temperature distribution, point cloud data, and position and attitude information.

[0058] Finally, the collected appearance images, temperature distribution and point cloud data are preliminarily checked for quality. After confirming that there are no frame drops or sensor abnormalities, they are packaged and integrated to generate a multimodal data package with a unified structure, which is then stored in the airborne storage device.

[0059] S2: Perform spatiotemporal calibration on the multimodal data packets, extract the features of each modality, and then use a cross-modal attention mechanism to fuse the features to obtain a multimodal feature vector;

[0060] Specifically, the steps include the following:

[0061] S2.1: Receive a multimodal data packet containing an appearance image, a temperature distribution image, 3D point cloud data, a timestamp, and location information;

[0062] The appearance image, temperature distribution image, and 3D point cloud data are time-aligned based on the timestamps in the multimodal data packet, and linear interpolation is used to compensate for the time sequence differences in data acquisition between sensors. The spatial pose transformation matrix is ​​calculated using the position information provided by the RTK-GPS center and the flight attitude data recorded by the inertial measurement center. Specifically, the rotation matrix is ​​calculated from Euler angles, or converted from quaternions to a rotation matrix. The UAV pose transformation matrix is ​​constructed using the RTK-GPS position as the translation vector. Based on pre-calibrated sensor extrinsic parameters (such as the transformation matrix from the appearance camera to the LiDAR, and the transformation matrix from the infrared camera to the appearance camera), the transformation matrix from each sensor data to the appearance image coordinate system is calculated.

[0063] Temperature distribution images and 3D point cloud data are projected onto the imaging coordinate system of the appearance image through a spatial pose transformation matrix to achieve spatial alignment of multimodal data.

[0064] S2.2: Extract features from the appearance image, temperature distribution image, and 3D point cloud data respectively, specifically as follows:

[0065] (1) The ResNet network is used to extract appearance features from the spatially aligned appearance image to generate an appearance feature map. The process is as follows:

[0066] The ResNet network is used to extract appearance features from images and identify appearance defects (such as cracks, corrosion, etc.) in power distribution network equipment. The training process is as follows:

[0067] A dataset of appearance images of power distribution network equipment was collected, including various equipment types such as transformers, insulators, and switches. Each image was labeled with an appearance defect category (e.g., normal, cracks, and corrosion). The appearance image dataset was divided into training, validation, and test sets.

[0068] The ResNet network structure is initialized, containing multiple convolutional layers, pooling layers, and residual connections. Residual connections alleviate the vanishing gradient problem by skipping layers, improving the training performance of deep networks. The input appearance image is adjusted to a uniform resolution, and spatial features are extracted layer by layer to generate a high-dimensional feature representation.

[0069] A cross-entropy loss function is used to quantify the difference between the predicted class and the true label. A stochastic gradient descent optimizer is employed to update the network parameters through backpropagation, iteratively optimizing the loss function. During training, the ResNet network performance is periodically evaluated on the validation set, and the learning rate is adjusted to avoid overfitting. Finally, the parameters of the best-performing ResNet network are saved to obtain the pre-trained ResNet network.

[0070] The spatially aligned appearance image is input into a pre-trained ResNet network. The appearance image is first adjusted to a uniform resolution to ensure input consistency. The ResNet network extracts spatial features from the image through multiple convolutional operations, capturing the appearance details of the power distribution network equipment. After residual connections and pooling operations, a high-dimensional appearance feature map is output, containing semantic information about the equipment's appearance. The appearance feature map is represented in matrix form, with dimensions reflecting the depth and spatial resolution of the features.

[0071] (2) The temperature distribution features are extracted from the spatially aligned temperature distribution image using a dedicated convolutional neural network to generate a temperature distribution feature map. The process is as follows:

[0072] A dedicated convolutional neural network is used to extract temperature distribution features from infrared thermal imaging to detect thermal anomalies in power distribution network equipment. The training process is as follows:

[0073] A dataset of temperature distribution images of power distribution network equipment was collected, including images of normal and abnormal temperature distributions, with annotations indicating the temperature anomaly category (e.g., normal, slight overheating, and severe overheating). The temperature distribution image dataset was divided into a training set and a validation set.

[0074] A dedicated convolutional neural network was designed, containing multiple convolutional layers, normalization layers, and activation functions, optimized for processing single-channel thermal imaging data. The input temperature distribution image was resized to a uniform size, and spatial features of the temperature distribution were extracted layer by layer to generate a feature representation.

[0075] A mean squared error loss function is used to quantify the difference between the predicted temperature distribution and the actual labeled temperature. An adaptive moment estimation optimizer is used to update the parameters of the dedicated convolutional neural network (CNN) through backpropagation, iteratively optimizing the loss function. The accuracy of the CNN is evaluated on a validation set, and hyperparameters are dynamically adjusted to improve generalization ability. Finally, the parameters of the best-performing CNN are saved to obtain a pre-trained CNN.

[0076] The spatially aligned temperature distribution image is input into a pre-trained dedicated convolutional neural network (CNN). The temperature distribution image is resized to fit the input of the dedicated CNN. The dedicated CNN extracts the spatial features of the temperature distribution through multiple convolutions and batch normalization operations, capturing the thermal anomaly areas of the power distribution network equipment. After activation function processing, the output temperature distribution feature map contains the thermal feature information of the equipment. The temperature distribution feature map is represented in matrix form, with dimensions aligned with the appearance feature map to ensure consistency in fusion.

[0077] (3) The PointNet network is used to extract geometric features from the spatially aligned 3D point cloud data to generate a geometric feature map. The process is as follows:

[0078] The PointNet network is used to extract geometric features from laser-generated 3D point clouds to identify deformations or structural anomalies in power distribution network equipment. The training process is as follows:

[0079] A 3D point cloud dataset of power distribution network equipment was collected, containing the 3D point cloud data of the equipment and annotating its geometric structure status (such as normal, deformed, and missing). The 3D point cloud dataset was divided into training, validation, and test sets.

[0080] The PointNet network is initialized, comprising a point cloud input layer, a transformation network, a feature extraction layer, and a classification layer. A 3D point cloud is input, and a spatial transformation matrix is ​​applied to correct the point cloud pose. Local and global features of each point are extracted to generate a geometric representation.

[0081] A cross-entropy loss function is used to quantify the difference between the predicted geometric state and the ground truth annotation. A stochastic gradient descent optimizer is employed to update the PointNet network parameters through backpropagation, iteratively optimizing the loss function. The PointNet network performance is evaluated on a validation set, and the point cloud sampling density is adjusted to improve accuracy. The optimal PointNet network parameters are saved to obtain the pre-trained PointNet network.

[0082] The spatially aligned 3D point cloud is input into a pre-trained PointNet network. The 3D point cloud first undergoes stereofiltering to unify point cloud density and correct pose deviations. The PointNet network captures local and global geometric features of the point cloud through a spatial transformation matrix and feature extraction layers, identifying deformation or structural anomalies in power distribution network equipment. It outputs a geometric feature map, represented as a vector, containing the equipment's 3D structural information, with dimensions adapted to subsequent fusion requirements.

[0083] S2.3: Cross-modal attention fusion is performed on the appearance feature map, temperature distribution feature map, and geometric structure feature map. This fusion is executed on a cloud server. To ensure consistency, the three feature maps are preprocessed: the appearance and temperature distribution feature maps are adjusted to a unified dimension (vector form) through dimensionality reduction operations (such as global average pooling). The geometric structure feature map, already in vector form, is aligned using linear interpolation or padding to obtain preprocessed appearance feature vectors, temperature distribution feature vectors, and laser point cloud feature vectors.

[0084] Cross-modal attention fusion achieves adaptive weighted fusion by calculating the correlation weights between feature vectors of each modality (appearance, temperature distribution, and laser point cloud). The cross-modal attention mechanism dynamically allocates weights by evaluating the complementarity and correlation between modalities, thereby enhancing the contribution of key features. The implementation process is as follows:

[0085]

[0086] The softmax function is applied to the relevance score to generate attention weights with values ​​in the range [0,1], reflecting the contribution between modal features.

[0087] Based on attention weights, the feature vectors of each modality are weighted and fused to generate a multimodal feature vector.

[0088]

[0089] Furthermore, the cross-modal feature fusion method of this invention dynamically calculates the correlation weights between modes through an attention mechanism, achieving adaptive weighted fusion and significantly improving the accuracy and precision of anomaly detection in power distribution network equipment. Compared to existing methods that only process single-modal data, this method integrates appearance images, infrared thermal imaging, and laser point cloud features to comprehensively capture equipment status, overcoming the limitations of single information. The dynamic weight allocation of the attention mechanism is superior to traditional fixed-weight fusion, and can adaptively enhance key features according to equipment status.

[0090] S3: Construct a heterogeneous graph using distribution network topology data, set nodes and edges, assign multimodal feature vectors to nodes, use graph convolutional neural networks to aggregate neighbor device features, update node representations, and output the anomaly classification results and anomaly propagation path prediction results of distribution network devices;

[0091] Specifically, the steps include the following:

[0092] S3.1: Construct a heterogeneous graph using distribution network topology data, set up nodes and edges, and assign multimodal feature vectors to the nodes;

[0093] It receives multimodal feature vectors containing appearance, temperature, and geometric information as input data.

[0094] (1) Obtain distribution network topology data from the power company's maintenance database, which includes the following information:

[0095] Equipment type: This includes various types of power distribution network equipment such as transformers, insulators, switches, and conductors, reflecting the electrical functions and physical characteristics of the equipment.

[0096] Device ID: Assign a unique identifier to each device to associate multimodal feature vectors and topology.

[0097] Electrical connections: Describes the electrical connections between devices, such as the connection between transformers and conductors, and the connection between switches and insulators, reflecting the physical and electrical topology of the power distribution network.

[0098] (2) Based on the distribution network topology data, a graph structure model is constructed, specifically as follows:

[0099] Node setup: Each device (transformer, insulator, switch, and conductor, etc.) in the distribution network is set as a node in the graph structure model. Each node corresponds to a device, and the device type and number are marked. For example, node 1 represents the transformer numbered T1, and node 2 represents the insulator numbered I2.

[0100] Edge Setting: Edges are set for the graph structure model based on the electrical connections in the distribution network topology data. Electrical connections represent the physical or electrical dependencies between devices; for example, the connection between transformer T1 and conductor W1 forms an edge. Edges can be represented as undirected or directed, determined based on the actual electrical flow direction of the distribution network.

[0101] Basic graph structure: By setting nodes and edges, a basic graph structure is formed to represent the topological relationships of the distribution network. The graph structure is stored as an adjacency list or adjacency matrix to accommodate subsequent feature assignment and weight calculation.

[0102] (3) Assign the multimodal feature vector to the node with the corresponding device number as the initial feature representation of the node. Specifically:

[0103] Feature matching: Based on the device number, the multimodal feature vectors are mapped one-to-one with the nodes in the graph structure model. For example, the multimodal feature vector of the transformer numbered T1 is assigned to node 1.

[0104] Feature assignment: The multimodal feature vector is used as the initial feature of the node and stored in the node's attributes. The feature vector of each node has a uniform dimension and includes information on appearance, temperature distribution, and geometric structure, reflecting the overall status of the device.

[0105] Data integrity: Before assigning values, check the completeness of the matching between the multimodal feature vectors and the device numbers to ensure there are no missing or incorrect assignments. If a device is missing a feature vector (e.g., a drone was not detected), use a preset default vector (such as the zero vector) to fill it in, maintaining the integrity of the graph structure.

[0106] (4) Based on the distribution network topology data and equipment operating characteristics, assign weights to the edges in the graph structure to reflect the electrical connection strength and dependencies between devices. Specifically:

[0107] Electrical distance: Calculate the electrical distance (e.g., conductor length or electrical impedance) between devices based on the distribution network topology data. For edges, the electrical distance reflects the strength of the physical or electrical connection between node 1 and node 2; for example, a shorter conductor length between a transformer and a conductor corresponds to a smaller electrical distance.

[0108] Load relationships: Extracting equipment load data (e.g., transformer load current, conductor power flow) from power company maintenance records quantifies the electrical dependencies between equipment. Load relationships reflect the mutual influence between nodes in the distribution network operation; for example, a high-load transformer has a significant impact on conductor current.

[0109] (5) Considering the relationship between electrical distance and load, calculate the edge weight, the expression is:

[0110]

[0111] The edge weights are normalized using a min-max method to ensure that the weight values ​​are within the [0,1] range, thus enhancing computational stability. Nodes, edges, and edge weights are integrated to form a heterogeneous graph with weighted connectivity. The heterogeneous graph is stored in a graph database or adjacency matrix, with nodes containing multimodal feature vectors and edges containing weighted information on electrical distance and load relationships.

[0112] To further explain, this step integrates multimodal feature vectors and distribution network topology data to construct a heterogeneous graph with weighted connectivity, significantly improving the accuracy and comprehensiveness of distribution network anomaly analysis. Compared to traditional methods that rely solely on single-modal data (such as images), this method uses multimodal feature vectors to assign initial features to nodes, enabling a more comprehensive reflection of equipment status.

[0113] S3.2: Use a graph convolutional neural network to aggregate features of neighboring devices, update node representations, and output the anomaly classification results and anomaly propagation path prediction results for distribution network devices, specifically:

[0114] (1) Graph convolutional networks are used to process heterogeneous graphs, aggregating contextual information between nodes to generate feature representations that include neighborhood relationships. The structure of a graph convolutional network is as follows:

[0115] The input layer is used to receive the heterogeneous graph of S3.1;

[0116] The graph convolutional layer contains multiple layers of graph convolution operations. Each layer aggregates the neighborhood information of nodes, updates the feature representation, and generates the final node features. The convolution operation is based on the weighted adjacency matrix and edge weights.

[0117] The classification decoding layer's function is to input the final node features into the fully connected layer and generate anomaly classification results. The classification layer uses the softmax function to output the anomaly type (e.g., crack, overheating, deformation) and severity level (e.g., minor, severe) for each node.

[0118] The propagation path analysis layer calculates the probability of abnormal propagation and predicts the abnormal diffusion path based on the characteristics of the final node and the edge weights.

[0119] (2) The graph convolutional network is trained to optimize node feature representation and anomaly classification performance. The process is as follows:

[0120] A heterogeneous distribution network dataset was collected, including nodes (equipment number, type, multimodal feature vector), edges (electrical connections, weights), and annotations (anomaly type, severity, such as "transformer crack, minor"). The dataset was divided into training, validation, and test sets, with the proportions adjusted based on the actual data volume. To simulate anomaly propagation, annotated propagation path data (e.g., "transformer overheating spreads to conductors") was generated. This data was extracted from power company maintenance records and simulation analysis.

[0121] Initialize the graph convolutional network parameters, including convolutional layer weights and classification layer weights, using the Xavier initialization method to ensure initial stability. Set hyperparameters, such as the number of convolutional layers (usually 2-3 layers), hidden layer dimensions, learning rate, etc., and adjust them based on validation set performance.

[0122] (3) The cross-entropy loss function is used for the anomaly classification task to quantify the difference between the predicted anomaly type and severity and the actual label; the mean squared error loss function is used for the propagation path prediction task to quantify the difference between the predicted propagation path and the actual path.

[0123] Optimization is achieved using the Adam optimizer, which updates the graph convolutional network parameters through backpropagation and iteratively optimizes the total loss. Classification accuracy and path prediction error are evaluated on the validation set, and hyperparameters are tuned to improve performance. The optimal graph convolutional network parameters are saved to obtain the pre-trained graph convolutional network.

[0124] (4) Input the heterogeneous graph of S3.1 into the pre-trained graph convolutional network and perform graph convolution operation, specifically:

[0125] First, neighborhood information aggregation is performed: In each layer of graph convolution, the features of adjacent nodes are weighted and summed according to the edge weights to update the current node feature representation. The weighted summation takes into account the electrical connection strength (e.g., high-weight edges reflect strong dependencies) to enhance contextual association. Then, through multiple convolution operations, multi-level neighborhood information is gradually aggregated to generate node feature representations that contain global context, capturing complex dependencies between power distribution network equipment, such as the impact of transformer overheating on conductors.

[0126] (5) Classify and decode the final node feature representation to generate anomaly classification results. The process is as follows:

[0127] The node features are input into the fully connected layer and mapped to the anomaly category space. Then, the softmax function is applied to output the anomaly classification probability of each node. The anomaly classification result of each node includes the anomaly type (such as crack, overheating, deformation) and severity level (such as slight, severe). The results are organized by device number and stored in the onboard storage device.

[0128] (6) Based on the characteristics of the final node and the edge weights, analyze the anomaly propagation pattern and predict the diffusion path, specifically:

[0129] Based on node characteristics and edge weights, the probability of an anomaly propagating from a node to its neighboring nodes is calculated using the following expression:

[0130]

[0131] Based on the propagation probability, anomaly propagation paths are generated using a greedy algorithm or a shortest path algorithm. For example, if an overheating anomaly is detected at a transformer node, it is predicted to propagate to the conductor nodes; the anomaly propagation path is output as a node sequence, including propagation probability and path weight.

[0132] Preferably, this step aggregates neighborhood information from heterogeneous graphs using graph convolutional networks and updates node features with edge weights to generate high-precision anomaly classification and propagation path prediction. This method utilizes multimodal features and electrical connection weights to capture contextual relationships between devices (such as the impact of overheating on adjacent conductors), significantly improving the comprehensiveness and accuracy of anomaly detection. Dynamic calculation of edge weights, compared to traditional fixed-weight graphs, enhances the targeting of anomaly propagation prediction, while propagation path prediction provides a new perspective for fault prevention.

[0133] S4: The bandwidth network center receives anomaly classification results and anomaly propagation path prediction results, and continuously monitors and utilizes historical network data and environmental factors to train the bandwidth prediction model, predict bandwidth change trends, actively adjust the allocation of computing tasks, and run different graph convolutional neural networks in low and high bandwidth situations to obtain compressed data packets and enhanced inspection reports.

[0134] Specifically, the steps include the following:

[0135] S4.1: Receive the distribution network equipment anomaly classification results and anomaly propagation path prediction results output by S3.2, activate the bandwidth network center, continuously monitor the real-time network status of the current communication link, and collect the following data:

[0136] Throughput: The actual data transmission rate of a communication link, reflecting bandwidth capacity;

[0137] Latency: The time it takes for data packets to travel from the drone to the cloud server, reflecting network response speed;

[0138] Signal strength: The received power of a wireless signal, reflecting the stability of the communication link.

[0139] Supplementary data is collected from historical databases and environmental sensors, specifically including: historical network data, and network performance data (such as throughput, latency, and signal strength time series) retrieved from the power company's maintenance database over a past period, reflecting long-term trends in communication links.

[0140] Environmental factors information: Environmental data of the current flight area is collected by sensors on board the drone, including: weather conditions, such as rainfall and wind speed, which affect wireless signal propagation; terrain features, such as mountains and urban buildings, which affect signal obstruction; and electromagnetic interference levels, such as interference from nearby high-voltage lines or base stations, which affect communication quality. The collected data is integrated into a unified feature vector, which includes real-time network status, historical network data, and environmental factors. The dimensions are aligned through normalization and stored in the onboard storage device.

[0141] S4.2: Use a bandwidth prediction model to analyze network conditions and environmental factors to predict bandwidth trends over a future period. The bandwidth prediction model construction and training process is as follows:

[0142] (1) Constructing the structure of the bandwidth prediction model:

[0143] Input layer: Receives a uniform feature vector;

[0144] LSTM layer: Uses a Long Short-Term Memory (LSTM) network to process time-series data and capture the temporal dependencies of the network state. The number of LSTM units is set according to the data complexity;

[0145] Fully connected layer: Maps the LSTM output to the bandwidth prediction value, outputting the bandwidth change trend (continuous value) over a future period of time (e.g., 10 minutes).

[0146] Activation function: The ReLU activation function is used to enhance the nonlinear expression.

[0147] (2) Data set preparation: Collect network data for power distribution network inspection scenarios, including historical network performance (throughput, delay, signal strength) and environmental factors (weather, terrain, interference), and label them as bandwidth value sequences.

[0148] (3) The training process is as follows: The mean squared error loss function is used to quantify the difference between the predicted bandwidth and the actual bandwidth. The Adam optimizer is used to update the model parameters through backpropagation and iteratively optimize the loss function. The prediction error is evaluated on the validation set, and hyperparameters (such as the number of LSTM units and the learning rate) are adjusted. The parameters of the bandwidth prediction model with the best performance are saved to obtain the pre-trained bandwidth prediction model.

[0149] (4) Real-time network status, historical network data, and environmental factors are input into the pre-trained bandwidth prediction model. The bandwidth prediction model receives the data through the input layer. Two long short-term memory (LSTM) network layers capture the temporal dependence of network status and environmental factors. The ReLU activation function is used to enhance the nonlinear expression. The fully connected layer maps the LSM network output to the bandwidth prediction value, outputting the bandwidth change trend prediction result for a future period. The bandwidth change trend prediction result is a time series, representing the continuous change of bandwidth value.

[0150] S4.3: Based on the predicted bandwidth change trend, determine the current network conditions:

[0151] Based on the communication requirements of power distribution network inspection and the amount of multimodal feature vector data, a preset bandwidth threshold is used to determine network conditions.

[0152] If the predicted bandwidth is lower than the broadband threshold, it is determined to be a low-bandwidth scenario, and the edge computing mode is triggered.

[0153] If the predicted bandwidth is higher than the broadband threshold, it is determined to be a high-bandwidth scenario, and the cloud collaboration mode is activated.

[0154] S4.4: In low-bandwidth scenarios, based on the anomaly classification results output by S3.2 and the power company's maintenance records, a heterogeneous graph subset is generated, specifically:

[0155] Preset fault thresholds include anomaly probability thresholds and historical fault rate thresholds. The anomaly probability threshold is determined based on the probability distribution of the anomaly classification results in S3.2, and the historical fault rate threshold is determined based on the fault frequency statistics of the power company's maintenance records.

[0156] The anomaly classification results from S3.2 are read from the onboard storage device. The anomaly classification probability and severity of each node are extracted. Combined with the heterogeneous graph from S3.1, device nodes with anomaly classification probabilities higher than the anomaly probability threshold, severe severity, or historical failure rates higher than the historical failure rate threshold are selected to form a critical device subset. The electrical connection relationships between the nodes in the critical device subset are preserved. Based on the electrical distance and load relationships in the distribution network topology data, the edge weight calculation method from S3.1 is reused to calculate the edge weights and apply min-max normalization. According to the device number, the multimodal feature vectors generated in S2.3 are assigned to the subset nodes. If a node lacks a feature vector, a preset default vector is used to fill it. The subset nodes, edges, and edge weights are integrated to form a heterogeneous graph subset. A lightweight graph convolutional network is run by the UAV onboard computing center to process the heterogeneous graph subset and generate preliminary results. The process is as follows:

[0157] (1) Define the structure of the lightweight graph convolutional network:

[0158] Input layer: Receives a subset of the heterogeneous graph from S3.1 (local device nodes and edges);

[0159] Lightweight convolutional layers: Use simplified graph convolution operations (reduce the number of layers and hidden dimensions, e.g., 1-2 layers, and reduce dimensionality);

[0160] Classification layer: The fully connected layer outputs the anomaly classification probability, and the dimension is adapted to the key anomaly type (such as high temperature, defect).

[0161] (2) A subset of heterogeneous graphs was used as the dataset, annotated with anomaly types and severity, and a portion of it was set as a validation set. During the training of the lightweight graph convolutional network, the cross-entropy loss function was used to quantify the difference between the predicted labels and the true labels. The number of nodes in the heterogeneous graph subset represents the number of devices in the local heterogeneous graph, and the number of key anomaly categories covers the main anomaly types and severity. The Adam optimizer was used to update the parameters of the lightweight graph convolutional network through backpropagation, optimize the loss function, and evaluate the classification accuracy on the validation set. The performance was optimized by adjusting the number of layers and feature dimensions of the lightweight graph convolutional network, and the parameters with the best performance were saved as the lightweight graph convolutional network. To adapt to the resource constraints of the UAV onboard computing center, the lightweight graph convolutional network was compressed using weight pruning and 8-bit integer quantization to reduce the computational load and ensure the efficiency of edge computing.

[0162] (3) In low-bandwidth scenarios, a subset of key equipment is selected and input into a lightweight graph convolutional network to generate preliminary results containing key anomaly classifications, such as "Transformer T1, high temperature, severe". Then, semantic compression is performed on the preliminary results, prioritizing the retention of high-value information, including infrared features of high-temperature areas (such as overheating values ​​in temperature distribution feature vectors), visible light features of defective areas (such as crack information in appearance feature vectors), and geometric anomaly features (such as deformation data in laser point cloud feature vectors). Compressed data packets are generated through entropy coding or principal component analysis to reduce the amount of data, and the compressed data packets are stored in the onboard storage device.

[0163] S4.5: In high-bandwidth scenarios, the complete heterogeneous graph generated in S3.1 (including nodes, edges, multimodal feature vectors, and edge weights) and the anomaly analysis requirements in S3.2 are uploaded to the cloud database via the UAV communication center for refined inference. The cloud server runs a full-scale graph convolutional network consistent with S3.2, employing a 3-4 layer convolutional structure and high dimensionality, performing graph convolution, classification decoding, and propagation path prediction, outputting anomaly classification results and anomaly propagation path prediction results, which are then stored in the cloud database.

[0164] Based on cloud or edge computing results, combined with distribution network topology and anomaly classification results, a dynamic risk assessment map is generated. Nodes are labeled with anomaly type and severity (e.g., red indicates severe overheating), and edges are labeled with propagation probability (e.g., high-probability paths are bolded). The risk distribution is updated using a weighted adjacency matrix and anomaly probabilities to reflect the overall risk status of the distribution network. Simultaneously, the heterogeneous map is mapped to 3D space, and a 3D visualization model is generated using visualization methods. Nodes represent equipment, edges represent electrical connections, color and size represent anomaly severity, and arrows represent propagation paths. Finally, the anomaly classification results, anomaly propagation path prediction results, risk assessment map, and 3D visualization model are integrated to generate an enhanced inspection report, output in a structured format, and uploaded to a cloud database or stored on onboard storage.

[0165] Preferably, this step optimizes the real-time performance and accuracy of distribution network inspection through a bandwidth prediction model and adaptive computation strategy. Compared to methods that rely solely on cloud processing, this method uses a lightweight graph convolutional network in low-bandwidth scenarios, reducing communication dependencies and adapting to complex environments (such as areas with strong electromagnetic interference). Semantic compression prioritizes the preservation of high-value features (such as high temperature and defects), outperforming traditional data compression and reducing bandwidth requirements. The cloud-based collaborative mode utilizes a complete graph convolutional network to provide refined analysis, dynamic risk assessment graphs, and 3D visualization models, enhancing the intuitiveness and accuracy of maintenance decisions.

[0166] This embodiment also provides a computer device applicable to the real-time image recognition method for UAV power distribution network inspection based on artificial intelligence, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the real-time image recognition method for UAV power distribution network inspection based on artificial intelligence as proposed in the above embodiment.

[0167] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. 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 stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0168] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the real-time image recognition method for UAV power grid inspection based on artificial intelligence, as proposed in the above embodiment. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0169] In summary, this invention constructs a heterogeneous graph, modeling equipment types, serial numbers, and electrical connections in distribution network topology data as nodes and edges, and assigning them multimodal feature vectors. A pre-trained graph convolutional neural network aggregates neighborhood information, outputting anomaly classification results and anomaly propagation path prediction results. This achieves the capture of complex dependencies between equipment and accurate prediction of anomaly propagation paths, significantly improving the comprehensiveness of inspections and fault prevention capabilities, overcoming the limitations of traditional image analysis that focuses only on a single device. Simultaneously, by monitoring real-time network status and adaptively allocating computational tasks, it achieves efficient data processing that dynamically adapts to bandwidth fluctuations, significantly reducing latency in weak network environments and improving edge computing efficiency, thus enhancing the adaptability and real-time performance of inspections in complex environments.

[0170] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An unmanned aerial vehicle power distribution network inspection image real-time identification method based on artificial intelligence, characterized in that: The unmanned aerial vehicle generates an inspection path through a path planning algorithm and collects multi-modal data to generate a multi-modal data package; The multi-modal data package is subjected to space-time calibration, and after extracting features of each mode, a cross-modal attention mechanism is used to fuse the features to obtain a multi-modal feature vector; An heterogeneous graph is constructed using power grid topology data, nodes and edges are set, the multi-modal feature vector is assigned to the nodes, a graph convolutional neural network is used to aggregate neighbor device features, the node representation is updated, and the abnormal classification result and abnormal propagation path prediction result of the power grid device are output; The bandwidth network center receives the abnormal classification result and abnormal propagation path prediction result, and continuously monitors and trains a bandwidth prediction model using historical network data and environmental factors to predict the bandwidth trend and actively adjust the allocation of computing tasks, running different graph convolutional neural networks in low and high bandwidth scenarios to obtain compressed data packages and enhanced inspection reports; The bandwidth network center receives the abnormal classification result and abnormal propagation path prediction result, and continuously monitors and trains a bandwidth prediction model using historical network data and environmental factors to predict the bandwidth trend, specifically, Receive the abnormal classification result and abnormal propagation path prediction result, start the bandwidth center to continuously monitor the real-time network state, combine historical network data and environmental factors to generate a unified feature vector; Analyze the unified feature vector through the pre-trained bandwidth prediction model to obtain the bandwidth trend; The active adjustment of the allocation of computing tasks, the running of different graph convolutional neural networks in low and high bandwidth scenarios, and the obtaining of compressed data packages and enhanced inspection reports are specifically, According to the bandwidth trend and the preset bandwidth threshold, determine the low bandwidth and high bandwidth scenarios; In the low bandwidth scenario, generate a heterogeneous graph subset based on the heterogeneous graph and the abnormal classification result, and run a lightweight graph convolutional network for processing to generate a preliminary result of abnormal classification, and generate a compressed data package through principal component analysis; In the high bandwidth scenario, upload the heterogeneous graph demand to the cloud database, run the pre-trained graph convolutional network to output the abnormal classification result and abnormal propagation path prediction result, and generate a dynamic risk assessment graph and a three-dimensional visualization model, which are integrated into an enhanced inspection report. The space-time calibration of the multi-modal data package refers to the time alignment of the appearance image, temperature distribution image and three-dimensional point cloud data using a linear interpolation method according to the timestamp, and the projection to the imaging coordinate system of the appearance image after calculating the spatial pose transformation matrix. 2.The AI-based real-time image recognition method for UAV power distribution network inspection according to claim 1, characterized in that: The unmanned aerial vehicle generates an inspection path through a path planning algorithm and collects multi-modal data, and generating a multi-modal data package refers to obtaining a power distribution network topology map, determining device positions as waypoints based on the power distribution network topology map, using A The path planning algorithm generates an inspection path, the inspection path is imported into the unmanned aerial vehicle, appearance images, temperature distribution images, three-dimensional point cloud data, time stamps and position information are collected, and a multi-modal data package is obtained. 3.The AI-based real-time image recognition method for UAV power distribution network inspection according to claim 2, characterized in that: The adoption of A The path planning algorithm generates the inspection path, which is defined as the starting point of the unmanned aerial vehicle landing point, the Euclidean distance between the waypoints and the flight priority, the cost function, and the total cost search to generate the inspection path. 4.The method of claim 3, wherein the method further comprises: The extraction of features of each mode is specifically, 5.The AI-based real-time image recognition method for UAV power distribution network inspection according to claim 4, characterized in that: The appearance image subjected to space-time calibration is subjected to appearance feature extraction using a pre-trained ResNet network to generate an appearance feature map; The temperature distribution image is subjected to temperature distribution feature extraction using a pre-trained special convolutional neural network to generate a temperature distribution feature map; The three-dimensional point cloud data is subjected to geometric structure feature extraction using a pre-trained PointNet network to generate a geometric structure feature map. The use of a cross-modal attention mechanism to fuse features to obtain a multi-modal feature vector is specifically, 6.The method of claim 5, wherein the method further comprises: determining a location of the UAV based on the image; and determining a location of the power distribution network based on the image. ​ The appearance feature map, the temperature distribution feature map and the geometric structure feature map are input to a cloud server for cross-modal attention fusion, the feature dimensions are aligned by global average pooling, and appearance feature vectors, temperature distribution feature vectors and laser point cloud feature vectors are generated; The relevance scores of each two of the appearance feature vectors, the temperature distribution feature vectors and the laser point cloud feature vectors are calculated by a dot product attention mechanism, and the attention weights are generated by applying a softmax function, the feature vectors are weighted and fused based on the attention weights, and a multi-modal feature vector is generated. 7.The real-time image recognition method for UAV power distribution network inspection based on artificial intelligence according to claim 6, characterized in that: An heterogeneous graph is constructed by power distribution network topology data, nodes and edges are set, and the multi-modal feature vector is assigned to the nodes, specifically, The power distribution network topology data is obtained from the power company maintenance database, and the network topology data includes device type, device number and electrical connection relationship; Each device is set as a node, and the electrical connection relationship is set as an edge to construct a basic graph structure; Based on the basic graph structure, the multi-modal feature vector is assigned to the corresponding node through the device number, and after the edge weight is calculated based on the electrical distance, normalization processing is performed to generate the heterogeneous graph. 8.The AI-based real-time power distribution grid inspection image recognition method for UAVs according to claim 7, characterized in that: The neighbor device features are aggregated using the graph convolutional neural network, the node representation is updated, and the abnormal classification result and the abnormal propagation path prediction result of the power distribution network device are output, specifically, The heterogeneous graph is input to the pre-trained graph convolutional neural network, the neighbor information is aggregated according to the edge weight through multi-layer graph convolution operation, the node feature representation is updated, and the abnormal classification result is output through the full connection layer and the softmax function; The abnormal propagation probability is calculated based on the node and edge weight, and the shortest path algorithm is used to generate the abnormal propagation path prediction result.

Images