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

Abstract

The invention discloses an unmanned aerial vehicle power distribution network inspection tour image real-time identification method based on artificial intelligence, and relates to the field of image identification, and the method comprises the steps: carrying out the space-time calibration of a multi-modal data packet, extracting the features of each modal, fusing the features through a cross-modal attention mechanism, obtaining a multi-modal feature vector, constructing a heterogeneous graph through the topological data of a power distribution network, and carrying out the recognition of the power distribution network inspection tour image. Setting nodes and edges, assigning the multi-modal feature vectors to the nodes, aggregating neighbor equipment features by using a graph convolutional neural network, updating node representation, outputting an anomaly classification result and an anomaly propagation path prediction result of power distribution network equipment, and receiving the anomaly classification result and the anomaly propagation path prediction result by a bandwidth network center. Historical network data and environmental factors are continuously monitored and utilized to train a bandwidth prediction model, and the bandwidth change trend is predicted; according to the invention, the capturing of the complex dependency relationship between equipment and the accurate prediction of the abnormal propagation path are realized, and the comprehensiveness and the fault prevention capability of the inspection tour are obviously improved.

Description

Technical Field

[0001] The present invention relates to the field of image recognition, and in particular to a method for real-time recognition of UAV distribution network inspection images based on artificial intelligence. Background Art

[0002] Distribution network inspection, a crucial component of power system operation and maintenance, has evolved from traditional manual inspections to automated and intelligent ones in recent years, driven by the rapid development of drone and artificial intelligence technologies. Early distribution network inspections relied primarily on manual field inspections, which were inefficient and posed safety risks. Advances in drone technology have led to the introduction of drones equipped with visible light cameras for image acquisition, combined with convolutional neural networks for equipment status analysis, significantly improving inspection coverage and efficiency. Furthermore, sensing technologies such as infrared thermal imaging and lidar are increasingly being incorporated into drone inspections to capture equipment temperature and geometry information, enhancing anomaly detection capabilities. Advances in artificial intelligence are further driving intelligent inspections.

[0003] While existing technologies have made some progress in drone-based distribution network inspections, there are still areas for improvement. For example, traditional image analysis methods focus on a single device, ignoring the electrical connections between devices in the distribution network. This fails to effectively predict the propagation paths of abnormalities, such as the potential impact of transformer overheating on adjacent lines, limiting the comprehensiveness of inspections. Furthermore, existing real-time processing technologies are limited by bandwidth fluctuations in weak network environments (such as remote mountainous areas), and edge computing power is insufficient to support complex models, resulting in high latency and reduced accuracy. Summary of the Invention

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

[0005] Therefore, the present invention provides an artificial intelligence-based UAV distribution network inspection image real-time recognition method to solve the problem of ignoring the electrical connection relationship between devices in the distribution network and failing to effectively predict the abnormal propagation path.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: The present invention provides a method for real-time recognition of UAV distribution network inspection images based on artificial intelligence, which includes: The drone generates an inspection route through a path planning algorithm and collects multimodal data to generate a multimodal data packet; Perform spatiotemporal alignment on multimodal data packets, extract features from each modality, and fuse them using a cross-modal attention mechanism to obtain a multimodal feature vector. 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 neighboring device features. Node representations are updated, and the anomaly classification results and anomaly propagation path prediction results for distribution network devices are output. The bandwidth network center receives anomaly classification results and anomaly propagation path prediction results, and continuously monitors and uses historical network data and environmental factors to train bandwidth prediction models, predict bandwidth change trends, actively adjust the allocation of computing tasks, and run different graph convolutional neural networks at low and high bandwidths to obtain compressed data packets and enhanced inspection reports.

[0007] As a preferred solution of the real-time recognition method of distribution network inspection images of a drone based on artificial intelligence described in the present invention, the drone generates an inspection path and collects multimodal data through a path planning algorithm. Generating a multimodal data packet refers to obtaining a distribution network topology map, determining the equipment location as a waypoint based on the distribution network topology map, using the A* path planning algorithm to generate an inspection path, importing the inspection path into the drone, collecting appearance images, temperature distribution images, three-dimensional point cloud data, timestamps and location information, and obtaining a multimodal data packet.

[0008] As a preferred solution of the real-time recognition method of UAV distribution network inspection images based on artificial intelligence described in the present invention, the use of the A* path planning algorithm to generate the inspection path refers to taking the UAV take-off and landing point as the starting point, defining the cost function through the Euclidean distance between waypoints and the flight priority, and generating the inspection path by minimizing the total cost search.

[0009] As a preferred solution of the real-time recognition method of UAV distribution network inspection images based on artificial intelligence described in the present invention, the spatiotemporal calibration of multimodal data packets refers to temporal 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 spatial pose transformation matrix is ​​calculated and then projected onto the imaging coordinate system of the appearance image.

[0010] As a preferred solution of the method for real-time recognition of UAV distribution network inspection images based on artificial intelligence of the present invention, wherein: the extraction of each modal feature is specifically as follows: The pre-trained ResNet network is used to extract appearance features from the spatiotemporally aligned appearance image and generate an appearance feature map. 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; The pre-trained PointNet network is used to extract geometric structure features from 3D point cloud data and generate a geometric structure feature map.

[0011] As a preferred solution of the real-time recognition method of UAV distribution network inspection images based on artificial intelligence of the present invention, wherein: the cross-modal attention mechanism is used to fuse features to obtain a multimodal feature vector, specifically, The appearance feature map, temperature distribution feature map, and geometric structure feature map are input to the cloud server for cross-modal attention fusion. The feature dimensions are aligned through global average pooling to generate appearance feature vectors, temperature distribution feature vectors, and laser point cloud feature vectors. The dot product attention mechanism is used to calculate the correlation score between the appearance feature vector, temperature distribution feature vector and laser point cloud feature vector, and the softmax function is applied to generate the attention weight. Based on the attention weight, each feature vector is weightedly fused to generate a multimodal feature vector.

[0012] As a preferred solution of the real-time recognition method of UAV distribution network inspection images based on artificial intelligence described in the present invention, wherein: a heterogeneous graph is constructed through the distribution network topology data, nodes and edges are set, and multimodal feature vectors are assigned to the nodes, specifically, Obtaining distribution network topology data from a database maintained by a power company, wherein the distribution network topology data includes device type, device number, and electrical connection relationship; Set each device as a node and the electrical connection relationship as an edge to build a basic graph structure; Based on the basic graph structure, the multimodal feature vectors are assigned to the corresponding nodes through the device number, and the edge weights are calculated based on the electrical distance and then normalized to generate a heterogeneous graph.

[0013] As a preferred solution of the real-time recognition method of UAV distribution network inspection images based on artificial intelligence described in the present invention, wherein: the graph convolutional neural network is used to aggregate neighboring device features, update node representations, and output abnormal classification results and abnormal propagation path prediction results of distribution network equipment, specifically, The heterogeneous graph is input into the pre-trained graph convolutional neural network. Through multi-layer graph convolution operations, neighborhood information is aggregated according to edge weights, node feature representation is updated, and the anomaly classification results are output through the fully connected layer and softmax function. The anomaly propagation probability is calculated based on the node and edge weights, and the shortest path algorithm is used to generate the anomaly propagation path prediction results.

[0014] As a preferred solution of the method for real-time recognition of UAV distribution network inspection images based on artificial intelligence described in the present invention, the bandwidth network center receives the abnormal classification results and abnormal propagation path prediction results, and continuously monitors and uses historical network data and environmental factors to train the bandwidth prediction model to predict bandwidth change trends, specifically, Receive 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 based on historical network data and environmental factors; The unified feature vector is analyzed through the pre-trained bandwidth prediction model to obtain the bandwidth change trend.

[0015] As a preferred solution of the real-time recognition method of UAV distribution network inspection images based on artificial intelligence described in the present invention, wherein: the allocation of computing tasks is actively adjusted, and different graph convolutional neural networks are run at low bandwidth and high bandwidth to obtain compressed data packets and enhanced inspection reports, specifically, Determine low-bandwidth and high-bandwidth scenarios based on bandwidth change trends and preset bandwidth thresholds; In low-bandwidth scenarios, a heterogeneous graph subset is generated based on the heterogeneous graph and anomaly classification results. A lightweight graph convolutional network is run to process the subset, generating preliminary anomaly classification results. Principal component analysis is then used to generate compressed data packets. In high-bandwidth scenarios, heterogeneous graph requirements are uploaded to the cloud database, and the pre-trained graph convolutional network is run to output anomaly classification results and anomaly propagation path prediction results. Dynamic risk assessment maps and three-dimensional visualization models are also generated and integrated into enhanced inspection reports.

[0016] The beneficial effects of the present invention are as follows: by constructing a heterogeneous graph, the device 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, and output abnormal classification results and abnormal propagation path prediction results, thereby capturing the complex dependencies between devices and accurately predicting abnormal propagation paths, significantly improving the comprehensiveness of inspections and fault prevention capabilities, and overcoming the limitations of traditional image analysis that only focuses on a single device. At the same time, by monitoring the real-time network status and adaptively allocating computing tasks, efficient data processing that dynamically adapts to bandwidth fluctuations is achieved, significantly reducing delays in weak network environments and improving edge computing efficiency, thereby enhancing the adaptability and real-time performance of inspections in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive efforts. Figure 1 This is a flow chart of the real-time recognition method for UAV distribution network inspection images based on artificial intelligence; Figure 2 Schematic diagram of cross-modal feature fusion; Figure 3 Schematic diagram of heterogeneous graph processing; Figure 4 Schematic diagram of bandwidth adaptation task allocation. DETAILED DESCRIPTION

[0018] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0019] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0020] Secondly, the term "one embodiment" or "embodiment" 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 various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0021] Reference Figures 1 to 4 , is an embodiment of the present invention, which provides a method for real-time recognition of UAV distribution network inspection images based on artificial intelligence, comprising the following steps: S1: The drone generates an inspection route through a path planning algorithm and collects multimodal data to generate a multimodal data packet; The specific steps include: S1.1: Obtain a distribution network topology map from the power company. This map contains the geographic coordinates, device types, and electrical connections of all devices to be inspected. Based on the map, determine the locations of all devices to be inspected and use these locations as waypoints for route planning.

[0022] The A* path planning algorithm is used to optimize the sorting of waypoints. The A* path planning algorithm operates in a two-dimensional geographic space, starting from the drone's take-off and landing points. It comprehensively considers the Euclidean distance between waypoints and flight priority, giving priority to covering equipment with high loads or historical failure rates, and generates a globally optimal and continuous flight path. Specifically: For any two waypoints, a cost function is defined to represent the actual flight cost between the two. The expression is: The A* algorithm searches for the optimal path by minimizing the total cost: Starting from the drone's takeoff and landing point, it initializes an open list (which stores waypoints to be explored) and a closed list (which stores waypoints that have already been explored). Starting from the takeoff and landing point, the A* algorithm iteratively performs the following operations: The waypoint with the lowest total cost in the open list is selected as the current node; the cost from the current node to the neighboring waypoints is calculated and the total cost is updated; if the neighboring waypoint is a high-priority device (such as a high-load transformer), the cost of the high-priority device is reduced by using a priority weight, and the high-priority device is added to the path first; the current node is moved to the closed list. This process repeats until all waypoints have been visited, generating a continuous flight path.

[0023] Path optimization considers drone flight constraints, such as maximum range and obstacle avoidance requirements. Ultrasonic sensors detect obstacles, dynamically adjust the flight trajectory between waypoints, and incorporate the A* algorithm to replan local paths to ensure path continuity and safety. Routes prioritize equipment with high loads or frequent failures (such as transformer T1 and insulator I2) to maximize inspection efficiency.

[0024] Furthermore, the A* path planning algorithm, combined with distribution network topology data and device priority weights, generates a globally optimal and continuous flight path, prioritizing distribution network equipment with high loads or historical failure rates, ensuring efficient drone inspections of critical equipment within a limited range. Compared to traditional path planning methods based solely on geographic distance, this method significantly improves the targeted nature of inspections, reduces redundant flight time, and optimizes energy consumption and mission duration through the use of priority adjustment coefficients and weighting mechanisms.

[0025] S1.2: Import the generated flight path into the drone's flight control center. The drone is equipped with a high-resolution visible light camera, an infrared thermal imaging camera, and a lidar. It is also equipped with an RTK-GPS center and an inertial measurement center. Before the mission begins, synchronize the sensors with the positioning center to ensure consistent temporal and spatial references during data collection.

[0026] The drone flies autonomously according to the generated flight path. During flight, the RTK-GPS center provides high-precision real-time position information, and the inertial measurement center records flight attitude to ensure flight stability. As the drone approaches each waypoint, multimodal data collection is triggered simultaneously: a visible light camera captures the device's exterior image, an infrared thermal imaging camera records the device's temperature distribution, and a lidar scans the device's structure to generate point cloud data.

[0027] Each frame of acquired appearance imagery, temperature distribution, and point cloud data is accompanied by a precise timestamp and current position information. The timestamp is generated uniformly by the onboard main control center, while the position information is output from the RTK-GPS center. All appearance imagery, temperature distribution, and point cloud data are organized in chronological order in the onboard memory, forming a data unit containing appearance imagery, temperature distribution, point cloud data, and position and attitude information.

[0028] Finally, a preliminary quality check is performed on the collected appearance images, temperature distribution, and point cloud data. After confirming that there is no frame loss or sensor anomaly, they are packaged and integrated to generate a multimodal data package with a unified structure and stored in the onboard storage device.

[0029] S2: Perform spatiotemporal alignment on the multimodal data packets, extract the features of each modality, and then use the cross-modal attention mechanism to fuse the features to obtain a multimodal feature vector; The specific steps include: S2.1: Receive a multimodal data packet containing appearance images, temperature distribution images, 3D point cloud data, timestamps, and location information; The appearance image, temperature distribution image, and 3D point cloud data are time-aligned based on the timestamps in the multimodal data packets, and linear interpolation is used to compensate for differences in sensor acquisition timing. 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 extrinsics (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.

[0030] The temperature distribution image and 3D point cloud data are projected into the imaging coordinate system of the appearance image through the spatial pose transformation matrix to achieve spatial alignment of multimodal data. S2.2: Extract features of the appearance image, temperature distribution image, and 3D point cloud data respectively, specifically: (1) The ResNet network is used to extract the appearance features of the spatially aligned appearance image and generate an appearance feature map. The process is as follows: The ResNet network is used to extract the appearance features of the appearance image and identify appearance defects (such as cracks and corrosion) of distribution network equipment. The training process is as follows: A dataset of appearance images of distribution network equipment was collected, including various equipment types such as transformers, insulators, and switches. Each image was annotated with the appearance defect category (such as normal, cracked, and corroded). The appearance image dataset was divided into training, validation, and test sets.

[0031] Initialize the ResNet network structure, which includes multiple convolutional layers, pooling layers, and residual connections. Residual connections alleviate the vanishing gradient problem by skipping across layers, improving the training performance of deep networks. Input appearance images are resized to a uniform resolution, and spatial features are extracted layer by layer to generate a high-dimensional feature representation.

[0032] A cross-entropy loss function is used to quantify the difference between the predicted categories and the true labels. A stochastic gradient descent optimizer is used to update network parameters via backpropagation, iteratively optimizing the loss function. During training, the ResNet network performance is regularly evaluated on the validation set, and the learning rate is adjusted to avoid overfitting. Finally, the best-performing ResNet network parameters are saved to obtain the pretrained ResNet network.

[0033] The spatially aligned appearance images are fed into a pre-trained ResNet network. The appearance images are first resized to a uniform resolution to ensure input consistency. The ResNet network extracts spatial features from the image through multiple layers of convolution, capturing the appearance details of distribution network equipment. After residual connections and pooling, it outputs a high-dimensional appearance feature map containing semantic information about the device's appearance. The appearance feature map is represented in matrix form, with the dimensions reflecting the depth and spatial resolution of the features.

[0034] (2) A dedicated convolutional neural network is used to extract the temperature distribution features of the spatially aligned temperature distribution image and generate a temperature distribution feature map. The process is as follows: A dedicated convolutional neural network extracts temperature distribution features from infrared thermal imaging to detect thermal anomalies in distribution network equipment. The training process is as follows: A temperature distribution image dataset of distribution network equipment was collected. This dataset includes images of normal and abnormal temperature distributions, and annotates temperature anomaly categories (e.g., normal, slightly overheated, and severely overheated). The temperature distribution image dataset is divided into a training set and a validation set.

[0035] A dedicated convolutional neural network is set up, consisting of multiple convolutional layers, normalization layers, and activation functions, optimized for processing single-channel thermal imaging data. The input temperature distribution image is resized to a uniform size, and the spatial features of the temperature distribution are extracted layer by layer to generate a feature representation.

[0036] A mean squared error loss function is used to quantify the difference between the predicted temperature distribution and the ground-truth annotations. An adaptive moment estimation optimizer is used to update the parameters of the dedicated convolutional neural network through backpropagation, iteratively optimizing the loss function. The accuracy of the dedicated convolutional neural network is evaluated on a validation set, and hyperparameters are dynamically adjusted to improve generalization. Finally, the parameters of the dedicated convolutional neural network with the best performance are saved to obtain the pre-trained dedicated convolutional neural network.

[0037] The spatially aligned temperature distribution images are fed into a pre-trained, dedicated convolutional neural network. The temperature distribution images are resized to a uniform size to accommodate the input of the dedicated convolutional neural network. Through multi-layer convolution and batch normalization, the dedicated convolutional neural network extracts the spatial characteristics of the temperature distribution and identifies thermal anomalies in distribution network equipment. After activation function processing, the output is a temperature distribution feature map containing the thermal signature information of the equipment. The temperature distribution feature map is represented in matrix form, with its dimensions aligned with the appearance feature map to ensure fusion consistency.

[0038] (3) The PointNet network is used to extract geometric structure features from the spatially aligned 3D point cloud data and generate a geometric structure feature map. The process is as follows; The PointNet network is used to extract the geometric structural features of laser 3D point clouds and identify deformation or structural anomalies of distribution network equipment. The training process is as follows: Collect a 3D point cloud dataset of distribution network equipment, including 3D point cloud data of distribution network equipment, annotating the geometric structure status (such as normal, deformed, and missing). The 3D point cloud dataset is divided into training, validation, and test sets.

[0039] Initialize the PointNet network, which includes a point cloud input layer, a transformation network, a feature extraction layer, and a classification layer. Input a 3D point cloud, apply a spatial transformation matrix to correct the point cloud pose, extract local and global features for each point, and generate a geometric structure representation.

[0040] 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 used to update the PointNet network parameters via backpropagation, iteratively optimizing the loss function. The PointNet network performance is evaluated on the validation set, and the point cloud sampling density is adjusted to improve accuracy. The best-performing PointNet network parameters are saved to obtain the pretrained PointNet network.

[0041] The spatially aligned 3D point cloud is fed into a pre-trained PointNet network. The 3D point cloud is first subjected to voxel filtering to unify the point cloud density and correct for pose deviations. Using a spatial transformation matrix and feature extraction layers, the PointNet network captures the local and global geometric features of the point cloud, identifies deformations or structural anomalies in distribution network equipment, and outputs a geometric feature map, represented as a vector. This map contains the 3D structural information of the equipment and has dimensions suitable for subsequent fusion requirements.

[0042] S2.3: Perform cross-modal attention fusion on the appearance feature map, temperature distribution feature map, and geometric structure feature map. Cross-modal attention fusion is performed on the cloud server. To ensure fusion consistency, the three feature maps are preprocessed: the appearance feature map and temperature distribution feature map are resized to a uniform dimension and vector form through dimensionality reduction operations (such as global average pooling). The geometric structure feature map is already in vector form, and the preprocessed appearance feature vector, temperature distribution feature vector, and laser point cloud feature vector are obtained through linear interpolation or padding to align the dimensions.

[0043] In cross-modal attention fusion, adaptive weighted fusion is achieved by calculating the correlation weights between the feature vectors of each modality (appearance, temperature distribution, and laser point cloud). The cross-modal attention mechanism dynamically assigns weights by evaluating the complementarity and correlation between modalities, thereby enhancing the contribution of key features. The implementation process is as follows: A softmax function is applied to the correlation score to generate attention weights with values ​​in the range [0, 1] that reflect the contribution between modal features.

[0044] Based on the attention weight, the feature vectors of each modality are weighted and fused to generate a multimodal feature vector. Further explanation: The cross-modal feature fusion method of the present invention dynamically calculates inter-modal correlation weights through an attention mechanism, achieving adaptive weighted fusion and significantly improving the precision and accuracy of anomaly detection in 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 device status, breaking through the limitations of single-modal information. The dynamic weight allocation of the attention mechanism is superior to traditional fixed-weight fusion, and can adaptively enhance key features based on device status.

[0045] S3: Construct a heterogeneous graph using distribution network topology data, set nodes and edges, assign multimodal feature vectors to nodes, aggregate neighboring device features using a graph convolutional neural network, update node representations, and output anomaly classification results and anomaly propagation path prediction results for distribution network devices. The specific steps include: S3.1: Construct a heterogeneous graph using the distribution network topology data, set nodes and edges, and assign multimodal feature vectors to nodes; It receives a multimodal feature vector containing appearance, temperature, and geometry information as input data.

[0046] (1) Obtain distribution network topology data from the power company's maintenance database, including the following information: Equipment type: includes the types of distribution network equipment such as transformers, insulators, switches, and conductors, reflecting the electrical functions and physical characteristics of the equipment.

[0047] Device ID: Assigns a unique identifier to each device, which is used to associate the multimodal feature vector and topology structure.

[0048] Electrical connection relationship: describes the electrical connection 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 distribution network.

[0049] (2) Based on the distribution network topology data, a graph structure model is constructed, specifically: Node setup: Set each device in the distribution network (transformers, insulators, switches, conductors, etc.) as a node in the graph model. Each node corresponds to a device and is labeled with the device type and number. For example, node 1 represents a transformer numbered T1, and node 2 represents an insulator numbered I2.

[0050] Edge configuration: Based on the electrical connections in the distribution network topology data, edges are configured for the graph model. Electrical connections represent 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, depending on the actual electrical flow direction in the distribution network.

[0051] Basic graph structure: By setting nodes and edges, a basic graph structure is formed to represent the topological relationship of the distribution network. The graph structure is stored as an adjacency list or adjacency matrix to adapt to subsequent feature assignment and weight calculation.

[0052] (3) Assign the multimodal feature vector to the node corresponding to the device number as the initial feature representation of the node. Specifically: Feature matching: Based on the device number, the multimodal feature vectors are mapped one-to-one to the nodes in the graph structure model. For example, the multimodal feature vector of the transformer numbered T1 is assigned to node 1.

[0053] Feature Assignment: Multimodal feature vectors are used as the initial features of a node and stored in the node's attributes. The feature vector of each node has a uniform dimension and contains information about appearance, temperature distribution, and geometric structure, reflecting the comprehensive status of the device.

[0054] Data integrity: Before assigning values, check the integrity of the matching between multimodal feature vectors and device numbers to ensure there are no missing or incorrect assignments. If a device is missing a feature vector (for example, if the drone has not inspected it), a default vector (such as a zero vector) is used to fill it in to maintain the integrity of the graph structure.

[0055] (4) According to the distribution network topology data and equipment operation characteristics, weights are set for the edges in the graph structure to reflect the electrical connection strength and dependency relationship between devices. Specifically: Electrical Distance: Calculates the electrical distance (e.g., conductor length or electrical impedance) between devices based on distribution network topology data. For edges, electrical distance reflects the strength of the physical or electrical connection between nodes 1 and 2. For example, a shorter conductor length between a transformer and a conductor corresponds to a smaller electrical distance.

[0056] Load relationships: Extracting device load data (e.g., transformer load current, conductor power flow) from utility maintenance records to quantify the electrical dependencies between devices. Load relationships reflect the mutual influence between nodes in the distribution network operation. For example, a heavily loaded transformer has a greater impact on conductor current.

[0057] (5) Considering the relationship between electrical distance and load, the edge weight is calculated as follows: Apply min-max normalization to edge weights to ensure they are within the [0, 1] range, enhancing computational stability. Nodes, edges, and edge weights are integrated to form a heterogeneous graph with weighted connectivity. This heterogeneous graph is stored in a graph database or adjacency matrix. Nodes contain multimodal feature vectors, and edges contain weighted information about electrical distances and load relationships.

[0058] Further explanation: 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, providing a more comprehensive reflection of device status.

[0059] S3.2: Use a graph convolutional neural network to aggregate neighboring device features, update node representations, and output anomaly classification results and anomaly propagation path prediction results for distribution network devices. Specifically: (1) Graph convolutional networks are used to process heterogeneous graphs, aggregate contextual information between nodes, and generate feature representations that include neighborhood associations. The graph convolutional network structure is as follows: The input layer is used to receive the heterogeneous graph of S3.1; The graph convolution layer contains multiple layers of graph convolution operations. Each layer aggregates node neighborhood information, updates feature representation, and generates final node features. The convolution operation is based on the weighted adjacency matrix and edge weights. The classification decoding layer feeds the final node features into the fully connected layer to generate anomaly classification results. The classification layer uses a softmax function to output the anomaly type (e.g., crack, overheating, deformation) and severity level (e.g., minor, severe) for each node.

[0060] The function of the propagation path analysis layer is to calculate the anomaly propagation probability and predict the anomaly diffusion path based on the final node features and edge weights.

[0061] (2) Training the graph convolutional network to optimize node feature representation and anomaly classification performance. The process is as follows: A heterogeneous distribution network graph dataset was collected, consisting of nodes (device number, type, and multimodal feature vectors), edges (electrical connections, weights), and annotations (anomaly type and severity, such as "transformer crack, minor"). This 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 propagates to conductors") was generated. This annotated propagation path data was extracted from power company maintenance records and simulation analysis.

[0062] Initialize the graph convolutional network parameters, including the convolutional 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), hidden layer dimensions, and learning rate, adjusting them based on validation set performance.

[0063] (3) The cross entropy loss function is used in the anomaly classification task to quantify the difference between the predicted anomaly type and severity and the true annotation. The mean square error loss function is used in the propagation path prediction task to quantify the difference between the predicted propagation path and the true path. Optimization uses the Adam optimizer, updating the GCN parameters through backpropagation and iteratively optimizing the total loss. Classification accuracy and path prediction error are evaluated on the validation set, and hyperparameters are adjusted to improve performance. The best-performing GCN parameters are saved to obtain the pretrained GCN.

[0064] (4) Input the heterogeneous graph of S3.1 into the pre-trained graph convolution network and perform graph convolution operation, specifically: 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, and the current node feature representation is updated. The weighted summation takes into account the strength of electrical connections (for example, high-weight edges reflect strong dependencies) to enhance contextual associations. Subsequently, multiple convolution operations are performed to gradually aggregate multi-order neighborhood information and generate node feature representations that include global context, capturing the complex dependencies between distribution network equipment, such as the impact of transformer overheating on conductors.

[0065] (5) Classify and decode the final node feature representation to generate abnormal classification results. The process is as follows: The node features are input into the fully connected layer and mapped to the anomaly category space. The softmax function is then applied to output the anomaly classification probability of each node. The anomaly classification results of each node include the anomaly type (such as crack, overheating, deformation) and severity level (such as minor, severe), organized by device number, and stored in the onboard storage device.

[0066] (6) Based on the final node characteristics and edge weights, analyze the anomaly propagation law and predict the diffusion path, specifically: According to the node characteristics and edge weights, the probability of anomaly propagation from a node to neighboring nodes is calculated as follows: Based on the propagation probability, a greedy algorithm or shortest path algorithm is used to generate anomaly diffusion paths. For example, if an overheating anomaly is detected at a transformer node, it is predicted to spread to the conductor nodes. The anomaly propagation path is output as a node sequence, including propagation probability and path weight.

[0067] Preferably, this step aggregates neighborhood information from heterogeneous graphs through a graph convolutional network (GCN), combines it with edge weights to update node features, and generates high-precision anomaly classification and propagation path predictions. This method leverages multimodal features and electrical connection weights to capture contextual relationships between devices (such as the impact of overheating on adjacent wires), significantly improving the comprehensiveness and accuracy of anomaly detection. Compared to traditional fixed-weight graphs, the dynamic calculation of edge weights enhances the specificity of anomaly propagation predictions, while propagation path prediction provides a new perspective for fault prevention.

[0068] S4: The bandwidth network center receives anomaly classification results and anomaly propagation path prediction results, and continuously monitors and uses historical network data and environmental factors to train bandwidth prediction models, predict bandwidth change trends, and actively adjust the allocation of computing tasks. It runs different graph convolutional neural networks at low and high bandwidths to obtain compressed data packets and enhanced inspection reports.

[0069] The specific steps include: S4.1: Receive the distribution network equipment anomaly classification results and anomaly propagation path prediction results output by S3.2, start the bandwidth network center, continuously monitor the real-time network status of the current communication link, and collect the following data: Throughput: The actual data transmission rate of a communication link, reflecting the bandwidth capacity; Latency: The time it takes for a data packet to travel from the drone to the cloud server, reflecting the network response speed; Signal strength: The received power of the wireless signal, reflecting the stability of the communication link.

[0070] Collect supplementary data from historical databases and environmental sensors, including historical network data. This includes retrieving network performance data from the power company's maintenance database over a period of time (such as throughput, latency, and signal strength time series) to reflect the long-term trend of the communication link.

[0071] Environmental Factor Information: UAV-mounted sensors collect environmental data about the current flight area, including weather conditions such as rainfall and wind speed, which affect wireless signal transmission; terrain features such as mountains and urban buildings, which block signals; and electromagnetic interference levels, such as interference from nearby high-voltage power lines or base stations, which can affect communication quality. This collected data is integrated into a unified feature vector, encompassing real-time network status, historical network data, and environmental factors. The dimensions are normalized and aligned, and stored onboard storage.

[0072] S4.2: Use the bandwidth prediction model to analyze network status and environmental factors and predict bandwidth change trends over the next period of time. The bandwidth prediction model construction and training process is as follows: (1) Constructing the structure of bandwidth prediction model: Input layer: receives unified feature vector; LSTM layer: Uses a long short-term memory (LSTM) network to process time series data and capture the temporal dependencies of network states. The number of LSTM units is set based on the complexity of the data. Fully connected layer: maps the LSTM output to the bandwidth prediction value and outputs the bandwidth change trend (continuous value) for a period of time in the future (such as 10 minutes); Activation function: Use ReLU activation function to enhance nonlinear expression.

[0073] (2) Dataset preparation: Network data from distribution network inspection scenarios are collected, including historical network performance (throughput, latency, signal strength) and environmental factors (weather, terrain, interference), which are annotated as bandwidth value sequences.

[0074] (3) The training process is as follows: using the mean squared error loss function to quantify the difference between the predicted bandwidth and the actual bandwidth, using the Adam optimizer 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.

[0075] (4) The 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 input through the input layer. The two-layer long short-term memory network layer captures the temporal dependency of the network status and environmental factors. The ReLU activation function is combined to enhance the nonlinear expression. The fully connected layer maps the long short-term memory network output to the bandwidth prediction value and outputs the bandwidth change trend prediction result for a period of time in the future. The bandwidth change trend prediction result is a time series, indicating the continuous change of the bandwidth value.

[0076] S4.3: Determine the current network conditions based on the predicted bandwidth change trend: According to the communication requirements of distribution network inspection and the data volume of multimodal feature vectors, a bandwidth threshold is preset to judge the network conditions.

[0077] 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.

[0078] 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 started.

[0079] S4.4: In the low-bandwidth scenario, based on the anomaly classification results output by S3.2 and the power company's maintenance records, a heterogeneous graph subset is generated. Specifically: The preset fault thresholds include an abnormal probability threshold and a historical failure rate threshold. The abnormal probability threshold is determined based on the probability distribution of the abnormal classification result in S3.2, and the historical failure rate threshold is determined based on the fault frequency statistics of the power company's maintenance records.

[0080] The anomaly classification results of S3.2 are read from the onboard storage device, and the anomaly classification probability and severity of each node are extracted. Combined with the heterogeneous graph of S3.1, device nodes with an anomaly classification probability higher than the anomaly probability threshold or a severe severity, or device nodes with a historical failure rate higher than the historical failure rate threshold, are selected to form a key device subset; the electrical connection relationship between the nodes of the key device subset is retained, and based on the electrical distance and load relationship in the distribution network topology data, the edge weight calculation method of S3.1 is reused to calculate the edge weight and apply min-max normalization processing; according to the device number, the multimodal feature vector generated by S2.3 is assigned to the subset node. If a node lacks a feature vector, it is filled with a preset default vector; the subset nodes, edges, and edge weights are integrated to form a heterogeneous graph subset. The lightweight graph convolutional network is run by the drone's onboard computing center to process the heterogeneous graph subset and generate preliminary results. The process is as follows: (1) Define the structure of lightweight graph convolutional network: Input layer: receives the heterogeneous graph subset of S3.1 (local device nodes and edges); Lightweight convolution layer: Use simplified graph convolution operations (reduce the number of layers and hidden dimensions, such as 1-2 layers, and reduce the dimension); Classification layer: The fully connected layer outputs the anomaly classification probability, and the dimension is adapted to the key anomaly types (such as high temperature and defects).

[0081] (2) Using a subset of heterogeneous graphs as a dataset, annotating the anomaly type and severity, and separating a portion as a validation set. During the training process of the lightweight graph convolutional network, the cross-entropy loss function is used to quantify the difference between the predicted label and the true label, where 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 is 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 is optimized by adjusting the number of layers and feature dimensions of the lightweight graph convolutional network, and the best parameters are saved as the lightweight graph convolutional network. In order to adapt to the resource constraints of the drone's onboard computing center, the lightweight graph convolutional network is compressed through weight pruning and 8-bit integer quantization methods to reduce the amount of computation and ensure the efficiency of edge computing.

[0082] (3) In low-bandwidth scenarios, a subset of key equipment is selected and fed into a lightweight graph convolutional network to generate preliminary results containing key anomaly classifications, such as "Transformer T1, high temperature, severe." The preliminary results are then semantically compressed, prioritizing high-value information, including infrared features of high-temperature areas (such as overheating values ​​in the temperature distribution feature vector), visible light features of defective areas (such as crack information in the appearance feature vector), and geometric anomaly features (such as deformation data in the laser point cloud feature vector). Entropy coding or principal component analysis is used to generate compressed data packets to reduce the data volume, and the compressed data packets are stored in an onboard storage device.

[0083] 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 from S3.2 are uploaded to the cloud database via the drone communication center for refined reasoning. The cloud server runs the same full-scale graph convolutional network as in S3.2, employing a 3-4-layer convolutional architecture and high dimensionality. It performs graph convolution, classification decoding, and propagation path prediction, outputting anomaly classification results and anomaly propagation path predictions, which are stored in the cloud database.

[0084] Based on cloud or edge computing results, combined with distribution network topology and anomaly classification results, a dynamic risk assessment graph is generated. Nodes are labeled with anomaly type and severity (for example, red indicates severe overheating), and edges are labeled with propagation probability (for example, 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 graph is mapped to three-dimensional space and a 3D visualization model is generated using visualization methods. Nodes represent devices, edges represent electrical connections, color and size indicate anomaly severity, and arrows indicate propagation paths. Finally, the anomaly classification results, anomaly propagation path predictions, risk assessment graph, and 3D visualization model are integrated to generate an enhanced inspection report. This report is output in a structured format and uploaded to a cloud database or stored onboard an onboard storage device.

[0085] Preferably, this step optimizes the real-time performance and accuracy of distribution network inspections through bandwidth prediction models and adaptive computing strategies. Compared to methods that rely solely on cloud-based processing, this method utilizes lightweight graph convolutional networks in low-bandwidth scenarios, reducing communication dependencies and adapting to complex environments (such as areas with strong electromagnetic interference). Semantic compression prioritizes high-value features (such as high temperatures and defects), outperforming traditional data compression and reducing bandwidth requirements. The cloud-based collaborative model utilizes the full graph convolutional network to provide refined analysis, dynamic risk assessment maps, and 3D visualization models, enhancing the intuitiveness and accuracy of maintenance decisions.

[0086] This embodiment also provides a computer device, which is suitable for the case of a real-time recognition method for drone distribution network inspection images 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 computer-executable instructions to implement the real-time recognition method for drone distribution network inspection images based on artificial intelligence proposed in the above embodiment.

[0087] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device 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 an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0088] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for real-time recognition of UAV distribution network inspection images based on artificial intelligence 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 read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0089] In summary, the present invention: by constructing a heterogeneous graph, the device type, number and electrical connection relationship in the distribution network topology data are modeled as nodes and edges, and assigned with multimodal feature vectors, and a pre-trained graph convolutional neural network is used to aggregate neighborhood information, output anomaly classification results and abnormal propagation path prediction results, thereby realizing the capture of complex dependencies between devices and accurate prediction of abnormal propagation paths, significantly improving the comprehensiveness of inspections and fault prevention capabilities, and overcoming the limitations of traditional image analysis that only focuses on a single device. At the same time, by monitoring the real-time network status and adaptively allocating computing tasks, efficient data processing that dynamically adapts to bandwidth fluctuations is achieved, significantly reducing delays in weak network environments and improving edge computing efficiency, and enhancing the adaptability and real-time performance of inspections in complex environments.

[0090] 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 the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A real-time recognition method for UAV distribution network inspection images based on artificial intelligence, characterized by: include, The drone generates an inspection route through a path planning algorithm and collects multimodal data to generate a multimodal data packet; Perform spatiotemporal alignment on multimodal data packets, extract features from each modality, and fuse them using a cross-modal attention mechanism to obtain a multimodal feature vector. 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 neighboring device features. Node representations are updated, and the anomaly classification results and anomaly propagation path prediction results for distribution network devices are output. The bandwidth network center receives anomaly classification results and anomaly propagation path prediction results, and continuously monitors and uses historical network data and environmental factors to train bandwidth prediction models, predict bandwidth change trends, actively adjust the allocation of computing tasks, and run different graph convolutional neural networks at low and high bandwidths to obtain compressed data packets and enhanced inspection reports.

2. The method for real-time recognition of UAV distribution network inspection images based on artificial intelligence according to claim 1, characterized in that: The UAV generates an inspection path and collects multimodal data through a path planning algorithm. Generating a multimodal data packet means obtaining a distribution network topology map, determining the equipment location as a waypoint based on the distribution network topology map, using the A* path planning algorithm to generate an inspection path, importing the inspection path into the UAV, collecting appearance images, temperature distribution images, three-dimensional point cloud data, timestamps, and location information to obtain a multimodal data packet.

3. The method for real-time recognition of UAV distribution network inspection images based on artificial intelligence according to claim 2, characterized in that: The A* path planning algorithm is used to generate the inspection path, which is to use the UAV take-off and landing point as the starting point, define the cost function through the Euclidean distance between waypoints and the flight priority, and generate the inspection path by minimizing the total cost search.

4. The method for real-time recognition of UAV distribution network inspection images based on artificial intelligence as claimed in claim 3 is characterized by: Temporal and spatial calibration of multimodal data packets refers to temporal alignment of appearance images, temperature distribution images, and three-dimensional point cloud data using a linear interpolation method based on timestamps, and then projecting the spatial pose transformation matrix onto the imaging coordinate system of the appearance image.

5. The method for real-time recognition of UAV distribution network inspection images based on artificial intelligence according to claim 4, characterized in that: The extraction of each modal feature is specifically as follows: The pre-trained ResNet network is used to extract appearance features from the spatiotemporally aligned appearance image and generate an appearance feature map. 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; The pre-trained PointNet network is used to extract geometric structure features from 3D point cloud data and generate a geometric structure feature map.

6. The method for real-time recognition of UAV distribution network inspection images based on artificial intelligence according to claim 5, characterized in that: The cross-modal attention mechanism is used to fuse features to obtain a multimodal feature vector, specifically, The appearance feature map, temperature distribution feature map, and geometric structure feature map are input to the cloud server for cross-modal attention fusion. The feature dimensions are aligned through global average pooling to generate appearance feature vectors, temperature distribution feature vectors, and laser point cloud feature vectors. The dot product attention mechanism is used to calculate the correlation score between the appearance feature vector, temperature distribution feature vector and laser point cloud feature vector, and the softmax function is applied to generate the attention weight. Based on the attention weight, each feature vector is weightedly fused to generate a multimodal feature vector.

7. The method for real-time recognition of UAV distribution network inspection images based on artificial intelligence according to claim 6, characterized in that: A heterogeneous graph is constructed through the distribution network topology data, nodes and edges are set, and multimodal feature vectors are assigned to nodes. Specifically, Obtaining distribution network topology data from a database maintained by a power company, wherein the distribution network topology data includes device type, device number, and electrical connection relationship; Set each device as a node and the electrical connection relationship as an edge to build a basic graph structure; Based on the basic graph structure, the multimodal feature vectors are assigned to the corresponding nodes through the device number, and normalization is performed after the edge weights are calculated based on the electrical distance to generate a heterogeneous graph.

8. The method for real-time recognition of UAV distribution network inspection images based on artificial intelligence according to claim 7, characterized in that: The graph convolutional neural network is used to aggregate neighbor device features, update node representations, and output abnormal classification results and abnormal propagation path prediction results of distribution network equipment. Specifically, The heterogeneous graph is input into the pre-trained graph convolutional neural network. Through multi-layer graph convolution operations, neighborhood information is aggregated according to edge weights, node feature representation is updated, and the anomaly classification results are output through the fully connected layer and softmax function. The anomaly propagation probability is calculated based on the node and edge weights, and the shortest path algorithm is used to generate the anomaly propagation path prediction results.

9. The method for real-time recognition of UAV distribution network inspection images based on artificial intelligence according to claim 8, characterized in that: The bandwidth network center receives the anomaly classification results and anomaly propagation path prediction results, and continuously monitors and uses historical network data and environmental factors to train the bandwidth prediction model to predict bandwidth change trends, specifically, Receive 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 based on historical network data and environmental factors; The unified feature vector is analyzed through the pre-trained bandwidth prediction model to obtain the bandwidth change trend.

10. The method for real-time recognition of UAV distribution network inspection images based on artificial intelligence according to claim 9, characterized in that: The active adjustment of the distribution of computing tasks runs different graph convolutional neural networks at low bandwidth and high bandwidth to obtain compressed data packets and enhanced inspection reports, specifically, Determine low-bandwidth and high-bandwidth scenarios based on bandwidth change trends and preset bandwidth thresholds; In low-bandwidth scenarios, a heterogeneous graph subset is generated based on the heterogeneous graph and anomaly classification results. A lightweight graph convolutional network is run to process the subset, generating preliminary anomaly classification results. Principal component analysis is then used to generate compressed data packets. In high-bandwidth scenarios, heterogeneous graph requirements are uploaded to the cloud database, and the pre-trained graph convolutional network is run to output anomaly classification results and anomaly propagation path prediction results. Dynamic risk assessment maps and three-dimensional visualization models are also generated and integrated into enhanced inspection reports.

Images