Artificial Intelligence-Based Method and System for Unmanned Aerial Vehicle Inspection and Control of Power Distribution Networks

Abstract

This invention discloses an artificial intelligence-based method and system for the inspection and control of power distribution networks using unmanned aerial vehicles (UAVs). Relating to the field of UAV power inspection technology, the method includes: adjusting the UAV's pose in real time based on target perception feedback during the inspection process to actively align with the target and acquire images of power components; uploading valid images from the selected images to a cloud-based management platform; identifying power component entities in the selected images and constructing a spatial semantic graph based on the spatial distribution relationships between components; performing topological logic reasoning and feature consistency analysis on the spatial semantic graph to identify abnormal component states and determine the power logic attributes of faulty components; generating maintenance instructions based on the identified abnormal states and power logic attributes and pushing them to the terminal. This invention addresses the problems of high redundancy in original images, insufficient physical positioning accuracy leading to ambiguity in component logical identity identification, and difficulty in identifying long-tail atypical hidden dangers during power distribution network inspection.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) power inspection technology, and more specifically, to an artificial intelligence-based method and system for the inspection and control of power distribution networks using UAVs. Background Technology

[0002] As the final link in the power system reaching users, the power distribution network is characterized by its wide geographical distribution, dense tower arrangement, and complex operating environment, making it a critical infrastructure for ensuring the reliability of electricity supply to society. With the deepening of smart grid construction, the use of drones for automated inspections has become a mainstream method to improve operation and maintenance efficiency and reduce human risks. Especially in areas with complex terrain, drones can acquire images of the operating status of key components such as towers, insulators, and crossarms from multiple angles, providing first-hand visual data for identifying potential power grid hazards.

[0003] Current drone inspections primarily rely on pre-set GIS (Geographic Information System) waypoints for programmed flight and GPS for spatial positioning. For data processing, a common approach is to use drones to take blind photos and transmit all raw images back to a cloud-based backend. The backend system then uses a trained supervised learning model to perform target detection on the acquired images, thereby classifying and determining predefined typical defects such as insulator damage and foreign objects entangled in the wire mesh.

[0004] However, existing technologies still face the following bottlenecks in practical applications: First, the distribution network environment suffers from strong signal interference, causing the GPS positioning of drones to frequently drift at the meter level. In multi-circuit parallel scenarios, it is difficult to accurately distinguish the logical affiliation of adjacent components based solely on physical coordinates. For example, it is difficult to accurately determine which phase (A, B, or C) a fault point belongs to, leading to ambiguity in identification. Second, the identification algorithm based on supervised learning is highly dependent on labeled samples. For atypical "long-tail" hazards in the distribution network that have low probability of occurrence and varied forms, such as micro-cracks at specific angles and unconventional discharge traces, there is a high rate of missed detection. Finally, the full image backhaul mode puts enormous pressure on the limited wireless communication bandwidth of the distribution network, and the backend system is filled with a large amount of invalid redundant data due to compositional deviations or blurry images, which seriously restricts the closed-loop timeliness of inspection and control. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an artificial intelligence-based method for the inspection and control of power distribution networks using unmanned aerial vehicles (UAVs). This method addresses the problems of high data redundancy, ambiguity in the logical identity identification of power components, and difficulty in identifying atypical faults during power distribution network inspections through active alignment and acquisition based on perception feedback, real-time filtering of effective images, topological reasoning of spatial semantic graphs, and feature consistency analysis.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An AI-based method for the inspection and control of power distribution networks using drones includes the following steps: adjusting the drone's pose in real time based on target perception feedback to align with the target, and acquiring images of power components containing multimodal information; filtering valid images from the images and uploading them to a cloud-based management platform; identifying power component entities in the valid images and constructing a spatial semantic graph based on the spatial distribution relationship between components; performing topological logic reasoning and feature consistency analysis on the spatial semantic graph to identify abnormal states of components and determine the power logic attributes of faulty components; generating maintenance instructions based on the abnormal states and power logic attributes, and pushing them to the terminal.

[0008] In a preferred embodiment, the real-time adjustment of the UAV pose based on target perception feedback includes: extracting feature points of power components in the inspection image in real time and obtaining their image coordinates, wherein the feature points are key points characterizing the geometric structure or semantic position of the power components; calculating the feature deviation between the current position of the feature points and the preset mapping position; generating control commands based on the feature deviation to adjust the attitude of the UAV or gimbal in real time so that the feature points move toward the preset mapping position, thereby achieving active target tracking.

[0009] In a preferred embodiment, filtering valid images from the images includes: using image processing algorithms to perform exposure detection and sharpness evaluation on the images, and removing images that do not meet preset standards; determining the validity of the images based on the proportion of electrical component entities in the images, and removing invalid and redundant images.

[0010] In a preferred embodiment, the construction of the spatial semantic graph includes: identifying power component entities as graph nodes; establishing graph edges connecting each image node based on the spatial relationships of the power component entities in the image; extracting the visual features of the power component entities, and generating node attributes that include component category features and morphological features, so as to form a spatial semantic graph that integrates spatial structure and component features.

[0011] In a preferred embodiment, topological logic reasoning of the spatial semantic graph includes: inputting the spatial semantic graph into a preset graph neural network model to generate enhanced node features that integrate topological context information; mapping the enhanced node features to a preset power logic label space to calculate the probability distribution of each power component entity under different logic categories; and determining the power logic attributes of the faulty component based on the probability distribution, wherein the power logic attributes include at least one of the power phase, circuit number, or connection level to which the faulty component belongs.

[0012] In a preferred embodiment, the feature consistency analysis includes: reconstructing local image features of an electrical component entity using a self-supervised feature reconstruction network, and calculating the structural consistency deviation entropy between the original image features and the reconstructed image features; if the deviation entropy exceeds a preset threshold, it is determined that the component has atypical morphological anomalies.

[0013] In a preferred embodiment, before generating the maintenance instructions, the method further includes: spatial matching and data fusion of damage features in the visible light image and temperature field distribution features in the infrared thermal imaging image; determining the severity level of the fault based on the fused features; and determining the execution priority of the maintenance instructions based on the severity level.

[0014] In a preferred embodiment, before acquiring images of power components containing multimodal information, the method further includes: acquiring GIS topology data, tower coordinates, and equipment ledgers of the distribution network to be inspected, and generating the optimal trajectory instruction set for the task using a multi-objective path planning algorithm.

[0015] In a preferred embodiment, the spatial relationship of the electrical component entities in the image includes at least one of the relative positional relationship, adjacency relationship, and physical connection relationship between the components.

[0016] This invention provides an AI-based unmanned aerial vehicle (UAV) inspection and control system for power distribution networks, comprising: an image acquisition module for real-time adjustment of the UAV's pose to align with the target based on target perception feedback, and for acquiring images of power components containing multimodal information; an edge filtering module for filtering valid images from the images and uploading them to a cloud-based management platform; a spatial semantic modeling module for identifying power component entities in the valid images and constructing a spatial semantic graph based on the spatial distribution relationships between components; an intelligent reasoning and diagnosis module for performing topological logic reasoning and feature consistency analysis on the spatial semantic graph, identifying abnormal states of components, and determining the power logic attributes of faulty components; and an operation and maintenance management module for generating operation and maintenance instructions based on the abnormal states and power logic attributes, and pushing them to the terminal.

[0017] The technical effects and advantages of the artificial intelligence-based unmanned aerial vehicle (UAV) inspection and control method for power distribution networks of this invention are as follows:

[0018] This invention achieves real-time and accurate alignment of inspection targets and standardized multimodal data acquisition through target perception feedback, solving the mapping deviation problem caused by environmental interference at its source. It significantly reduces data backhaul redundancy and communication bandwidth pressure by utilizing an image validity screening mechanism. At its core, it overcomes the identity identification ambiguity caused by insufficient accuracy in traditional physical positioning by constructing a spatial semantic graph and performing topological logic reasoning, achieving accurate mapping of the power logic attributes of faulty components. Combined with feature consistency analysis, it greatly enhances the detection capability for long-tail atypical defects. Finally, through a closed-loop command push system, it improves the immediacy of distribution network operation and maintenance response and the scientific nature of auxiliary decision-making. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the process of the AI-based unmanned aerial vehicle (UAV) inspection and control method for power distribution networks provided in an embodiment of the present invention.

[0020] Figure 2 A schematic diagram of an inspection trajectory based on multi-objective optimization provided in an embodiment of the present invention;

[0021] Figure 3 A visualization diagram of anomaly diagnosis based on self-supervised reconstruction provided in an embodiment of the present invention;

[0022] Figure 4 This is a block diagram of an AI-based power distribution network drone inspection and control system provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1, Figure 1 This invention presents an artificial intelligence-based method for the inspection and control of power distribution networks using unmanned aerial vehicles (UAVs), comprising the following steps:

[0025] S1, during the drone inspection process, adjusts the drone's attitude in real time based on target perception feedback to actively align with the target and collect images of power components containing multimodal information.

[0026] It should be noted that before conducting drone inspections, it is necessary to obtain the GIS topology data, tower coordinates, and equipment ledgers of the power distribution network to be inspected. A multi-objective path planning algorithm is then used to generate the optimal flight path command set for the task. Specifically, the spatial vector distribution map of the target distribution area is retrieved through the Power Production Management System (PMS), and the latitude, longitude, and height information of each tower is extracted as three-dimensional spatial discrete nodes. Combined with the drone's remaining battery power, single takeoff and landing coverage radius, and no-fly zone restrictions, a multi-objective function is established with the shortest flight distance and highest inspection efficiency as its objectives. The cost calculation formula for the multi-objective path planning is as follows:

[0027] (1)

[0028] in, For the total inspection cost, For the first The physical distance of the flight path. In the first The time spent taking photos at each inspection point and These are the distance weight coefficient and the time weight coefficient, respectively, and satisfy the following conditions: Solving the problem using the particle swarm optimization algorithm makes The minimized node access sequence generates the optimal mission trajectory instruction set, which includes takeoff and landing points, waypoints, and camera pose information. For example... Figure 2 As shown in the figure, the inspection trajectory generated based on the above algorithm is illustrated. Discrete points represent the three-dimensional coordinates of the towers, and the connecting lines represent the optimal access sequence path optimized by the particle swarm optimization algorithm. This figure visually demonstrates the improvement in task efficiency brought about by trajectory planning. After generating the above instruction set and uploading it to the UAV flight control system, the UAV begins to execute autonomous flight tasks. When the UAV navigates to the vicinity of the target waypoint, the system automatically switches to visual guidance mode to enter the active alignment phase.

[0029] In this embodiment, the real-time adjustment of the UAV pose based on target perception feedback includes:

[0030] Feature points of power components in inspection images are extracted in real time, and the image coordinates of the feature points are obtained. The feature points are key points that characterize the geometric structure or semantic location of the power components.

[0031] Calculate the feature deviation between the current position of the feature point and the preset composition position;

[0032] Based on the aforementioned characteristic deviation, control commands are generated, and the flight attitude angle or gimbal angle of the UAV is adjusted in real time through a feedback control algorithm for active centering tracking of the inspection target.

[0033] The above steps are as follows: After the UAV arrives at the preset waypoint, the onboard visual sensor captures a low-resolution preview image. A lightweight convolutional neural network based on heatmap regression and using the HRNet-W18 architecture is then used to extract the coordinates of semantic key points of the power components in real time. These semantic key points are the center point of the insulator string or the vertex of the crossarm edge. The real-time coordinates of these semantic key points in the current image coordinate system are defined as follows: The image center reference coordinates of the preset composition are defined as follows: .

[0034] Furthermore, in order to achieve accurate angle compensation, the feature deviation in the pixel dimension is first decomposed into horizontal pixel deviation. Vertical pixel deviation The calculation formula is as follows:

[0035] (2)

[0036] Then, using the camera's focal length With pixel size Convert pixel deviation into an increment of angular deviation in physical space. The formula for calculating the angle deviation is as follows:

[0037] (3)

[0038] The obtained angle deviation is input into the proportional-integral-derivative (PID) control algorithm, which outputs control commands for the UAV yaw angle adjustment and gimbal pitch angle adjustment. The specific steps of the feedback control adjustment are as follows: the UAV yaw angle is adjusted through the horizontal control loop to eliminate... Simultaneously, the gimbal pitch angle is adjusted via the pitch control loop to eliminate... The control algorithm formula is as follows:

[0039] (4)

[0040] in, For the angular velocity control quantity of the actuator, These are the proportional, integral, and derivative gains, respectively. This control loop ensures that the inspected target is consistently centered on the screen.

[0041] The system determines the stability of the image composition by calculating the variance of feature point coordinates in five consecutive frames. When the variance is lower than a preset stability threshold, alignment is considered complete and synchronous acquisition is triggered. The acquired multimodal information includes visible light image information and infrared thermal imaging information. Specifically, the dual-light payload mounted on the UAV is synchronously triggered to acquire high-resolution visible light images and raw infrared images, and a pre-calibrated homography matrix is ​​used... The infrared images are resampled and subjected to affine transformation to achieve spatial alignment. The spatial alignment calculation formula is as follows:

[0042] (5)

[0043] in, These are the pixel coordinates of the infrared image. The coordinates are aligned with the mapped visible light, thereby enabling simultaneous observation of structural damage and thermal defects of the target component in the same spatial dimension.

[0044] This step, through visual feedback control and multimodal spatial alignment technology, compensates for the mapping deviation caused by unstable positioning signals in the power distribution network environment, solving the problems of target loss and mismatch of multi-source data caused by shooting composition offset in traditional inspections. The technical effect is to improve the standardization and accuracy of image acquisition, providing a high-quality, spatially consistent multimodal data foundation for subsequent logical reasoning based on spatial semantic graphs.

[0045] S2, After filtering the valid images from the images, upload them to the cloud management platform.

[0046] It should be noted that the selection of valid images from the images includes: using image processing algorithms to perform exposure detection and sharpness evaluation on the acquired images, and removing images that do not meet preset standards. In this embodiment, blurry, overexposed, or underexposed images are removed. Specifically, the acquired original images are processed in real-time grayscale in the UAV's onboard computing module, the high-frequency components of the image are extracted using the Laplacian operator, and the sharpness of the image edges is measured by calculating the variance of the high-frequency components. The image sharpness evaluation calculation formula is as follows:

[0047] (6)

[0048] in, The grayscale distribution of the input image. For the Laplace convolution operator, Var(·) represents the variance calculation for the calculated sharpness score. The system presets a sharpness threshold. ,like If the image is determined to be blurred due to motion or defocus, it will be removed. For images that pass the sharpness detection, the system further uses the brightness histogram distribution to evaluate the exposure quality, specifically by calculating the average brightness of all pixels in the entire image. The formula for calculating the exposure brightness is as follows:

[0049] (7)

[0050] in, For pixels brightness value, and These are the width and height of the image, respectively. This represents the average brightness. When Greater than the highlight threshold or less than the low brightness threshold When an image is overexposed or underexposed, it is removed from the upload queue to ensure the visual recognizability of the uploaded data.

[0051] After initial image quality screening, the validity of the images is further determined based on the proportion of electrical components in the images, eliminating invalid and redundant images. In this embodiment, images that do not contain the preset electrical component entities are mainly eliminated. Specifically, the airborne lightweight deep learning operator YOLOv5-lite is used to extract semantic regions from the image content, identifying candidate boxes for preset electrical components such as insulators, crossarms, and towers, and calculating the total pixel area covered by all component candidate boxes in the image plane. The formula for calculating the validity proportion of electrical components is as follows:

[0052] (8)

[0053] in, To identify the first The bounding box of an electrical component entity. The total number of identified components. This represents the effective proportion of electrical components in the image. If... Less than the preset significance threshold Specifically, when the inspection target deviates from the center of the field of view or is composed of pure background due to accidental triggering, the system determines that the image is an invalid and redundant image and performs local truncation, and only transmits the high-quality and simplified dataset that has passed the double screening back to the cloud management platform via the wireless communication link.

[0054] As shown in Table 1, the table lists the various judgment parameters and their threshold settings when the airborne terminal performs the filtering task. Through this parameter matrix, the airborne module can accurately truncate invalid and redundant data.

[0055] Table 1

[0056]

[0057] This step utilizes an onboard edge processing mechanism to achieve real-time load reduction of massive amounts of raw inspection data, addressing the technical challenges of insufficient wireless communication bandwidth and redundant backend storage in large-scale power distribution network inspections. The technical effect is that it filters out a large amount of invalid image information at the source, significantly improving the timeliness of data transmission and providing high-quality, high-purity training and diagnostic materials for subsequent high-complexity topology logic reasoning in the cloud.

[0058] S3 identifies electrical component entities in the filtered images and constructs a spatial semantic graph based on the spatial distribution associations between components.

[0059] It should be noted that traditional solutions typically rely solely on single-object deep learning detection operators to obtain the rectangular bounding boxes of components and attempt to perform hard matching between these bounding boxes and equipment ledgers using geographic coordinate information. The problem with this approach is that in multi-circuit parallel distribution networks or complex geographical environments, GPS positioning suffers from meter-level random drift errors, and adjacent components visually overlap significantly, making it difficult for the system to accurately distinguish the logical affiliation of components (difficult to distinguish between side-by-side A-phase and B-phase insulators), easily leading to ambiguity in identification. The method in this step introduces spatial semantic association, transforming isolated detection results into a structured graph model with topological constraints, effectively solving the positioning problem caused by insufficient physical coordinate accuracy. The technical effect is that it not only achieves component category identification but also extracts the logical dependencies between components, providing a structured decision-making basis for subsequent accurate power phase derivation.

[0060] In this embodiment, the construction of the spatial semantic graph includes:

[0061] 1) The identified power component entities are used as graph nodes. Specifically, the Faster R-CNN feature extraction network deployed in the cloud is used to perform a global search on the image to obtain the category label and bounding box of each power component entity. and the centroid coordinates in the pixel coordinate system Define each individual entity as a graph. One of the nodes .

[0062] 2) Based on the spatial relationships of each power component entity in the image, graph edges connecting each image node are established. In this embodiment, spatial relationships are represented by relative positional relationships, adjacency relationships, or physical connection relationships in a preset coordinate system. Specifically, the Euclidean distance between the centroids of each node and the intersection-over-union ratio (IoU) of the bounding boxes are calculated, and the spatial coupling degree between components is evaluated by combining prior knowledge of power topology (such as the insulator string must be hung below the crossarm). If the distance between two component nodes is less than the preset spatial influence radius (specifically implemented by setting a coupling threshold), the spatial coupling degree between components is determined. If there is a physical connection between the two nodes, then a connecting edge is established between them. The formula for calculating the spatial correlation strength between the components is as follows:

[0063] (9)

[0064] in, For nodes and The adjacency weights between elements are used to represent the elements of the graph's adjacency matrix. The coordinates of the centroid of the component are... For component bounding boxes, and The preset distance weight and overlap weight coefficients are used. This is a spatial scale parameter. Using this formula, the system can automatically connect entities with physical constraints, such as "tower-crossarm," "crossarm-insulator," and "insulator-conductor," forming a topological network that reflects the actual physical connection logic.

[0065] 3) Extract visual features from each power component entity to generate node attributes that include component category and morphological features. Specifically, a 1024-dimensional pooled feature vector is extracted from the bounding box region of each component using a pre-trained ResNet-50 backbone network and defined as a deep visual feature. And generate morphological feature vectors by combining the geometric distribution of components in the image. The formula for calculating the morphological feature vector is as follows:

[0066] (10)

[0067] in, For the first The aspect ratio of the bounding box of each component. This represents the area of ​​a component relative to the entire image. The tilt angle of the component's principal axis relative to the image's horizontal line. These features are then concatenated to obtain the initial attribute vectors for each node. The formula for merging node attributes is:

[0068] (11)

[0069] in, This represents the vector concatenation operator. In summary, the specific steps for constructing the spatial semantic graph are as follows: First, use a high-precision cloud-based detection operator to obtain the pixel-level coordinates and categories of all components; then, based on the spatial association strength formula and combined with a threshold... The algorithm determines the edge relationships between nodes and constructs the adjacency matrix of the graph. Finally, it encapsulates the enhanced attribute vectors of each node through a feature extraction network to complete the construction of the spatial semantic graph. This graph model not only fully preserves the visual information of power components, but also solidifies the spatial topological constraints between components through the graph edge structure, achieving a technological leap from pixel-level recognition to semantic-level understanding.

[0070] S4. Perform topological logic reasoning and feature consistency analysis on the spatial semantic graph to identify abnormal states of components and determine the power logic attributes of faulty components.

[0071] It should be noted that traditional solutions typically rely on fixed geographic coordinate matching or simple local image classification algorithms. The problem with these solutions is that GPS signal drift in power distribution network environments often leads to incorrect identification of towers and connected components. Furthermore, traditional supervised learning models struggle to identify rare and atypical defects outside the training samples, resulting in significant "long-tail" fault detection. The method presented in this paper extracts topological constraint information between components using graph neural networks and combines this with a self-supervised learning mechanism to detect subtle morphological distortions, effectively solving the challenges of ambiguous identification and early warning of unknown defects in complex environments. The resulting technical benefits include the system's ability to autonomously correct physical positioning deviations through the relative arrangement logic between components and its proactive ability to detect new types of hidden dangers.

[0072] In this embodiment, performing topological logic reasoning on the spatial semantic graph includes:

[0073] (1) Input the spatial semantic graph into a preset graph neural network model to generate enhanced node features that integrate topological context information. Specifically, use multi-layer graph convolution operators to process the graph node attributes constructed in S3. Message passing and feature aggregation are performed to enable each node to capture the spatial arrangement patterns of components such as crossarms and insulators within its neighborhood. The aggregation calculation formula for the enhanced node features is as follows:

[0074] (12)

[0075] in, For the first Enhanced node features resulting from the fusion of topological context among individual nodes For nodes The set of neighboring nodes, Let be the degree of the node. The state transition matrix is ​​a learnable matrix. This is the activation function.

[0076] (2) Map the enhanced node features to a preset power logic label space, and calculate the probability distribution of each power component entity under different logic categories. Specifically, map the enhanced node features... The fully connected layer is subjected to dimensionality transformation, and a normalized exponential function is used to map it to a classification space containing phase labels (A, B, C phases), loop labels, and hierarchy labels. The formula for calculating the classification probability distribution is as follows:

[0077] (13)

[0078] in, For nodes Belongs to the The probability of power-like logical attributes. The output layer's mapping weight matrix. This represents the total number of logical label categories.

[0079] (3) Based on the probability distribution, determine the power logic attributes of the faulty component, wherein the power logic attributes include at least one of the power phase, circuit number, or connection level to which the faulty component belongs. The steps are as follows:

[0080] 1) Extract the neighborhood topological features of each node in the graph;

[0081] 2) Determine the logical identity of each node through probability calculation;

[0082] 3) Complete the attribute labeling of components from the physical space to the power business logic space according to the maximum probability criterion.

[0083] The feature consistency analysis of the spatial semantic graph includes:

[0084] (1) A self-supervised feature reconstruction network is used to reconstruct the local image features of the power component entity, and the structural consistency deviation entropy between the original image features and the reconstructed image features is calculated. Specifically, a self-supervised encoder based on mask modeling is used to reduce the dimensionality of the identified component image region, and a decoder is used to attempt to restore the pixel details hidden by the mask, thereby generating reconstructed image features. Then, by comparing the original features The distribution deviation is assessed, and the formula for calculating the structural consistency deviation entropy is as follows:

[0085] (14)

[0086] in, The entropy is the structural consistency deviation. The pixel space range of the component image. For pixel position index. For example... Figure 3 As shown in the figure, the image reconstruction comparison during the anomaly diagnosis process is presented. The left side is the original inspection image (a). The Insulator: 0.98 above the red box in the figure is the probability that the logical identity is an insulator calculated by the algorithm. The middle is the healthy state reconstruction image generated by the self-supervised network (b). The right side is the deviation heat map obtained by subtracting the two (c). The dark red area in the figure is the detected atypical morphological anomaly. This visualization result clearly shows the system's sensitivity to unknown faults.

[0087] (2) If the deviation entropy exceeds the preset threshold If the component exhibits an atypical morphological anomaly, it is determined that the component has an atypical morphological anomaly. Specifically, when the component has previously unseen fine cracks, foreign object entanglement, or unconventional deformation, the reconstructed network cannot accurately restore its abnormal details, resulting in a significant increase in deviation entropy, thereby achieving automatic early warning of unknown long-tail defects.

[0088] This step endows the inspection data with clear business semantics through topological reasoning and enhances the sensitivity to complex abnormal states through consistency analysis, significantly improving the diagnostic robustness of the management and control platform under non-ideal operating conditions.

[0089] S5 generates operation and maintenance instructions based on the identified abnormal states and power logic attributes, and pushes them to the terminal.

[0090] It should be noted that before generating the maintenance instructions, the process also includes: spatial matching and data fusion of damage features in the visible light image and temperature field distribution features in the infrared thermal imaging image. Specifically, this involves using the homography matrix obtained in step S1. The visible light damage score is extracted by pixel-level overlay mapping between the physical defect regions identified in the visible light image and the thermal anomaly regions in the infrared image. Infrared temperature rise anomaly score The visible light damage score The quantification is based on the degree of component morphological distortion identified in step S4, while the infrared temperature rise anomaly score is... The multimodal fusion feature score is obtained by calculating the difference between the highest temperature at the fault point and the ambient reference background temperature and then normalizing it. The calculation formula for the multimodal fusion feature score is as follows:

[0091] (15)

[0092] in, The integrated fault intensity index is the result of fusion. For visible light damage scoring, For infrared temperature rise anomaly scoring, and These are the preset weighting coefficients. This represents the peak temperature of the fault area in the infrared image. The average temperature of the local environmental background. This is the maximum permissible operating temperature rise threshold for this type of equipment.

[0093] The formula for calculating the infrared temperature rise anomaly score is as follows:

[0094] (16)

[0095] in, The maximum saturation entropy value for structural damage is preset; when the deviation entropy exceeds the threshold, the score increases linearly in the range of 0 to 1 as the entropy value increases, thereby realizing a quantitative description of the severity of atypical morphological abnormalities.

[0096] After completing the quantitative integration of the above cross-modal characteristics, a comprehensive evaluation index that can fully characterize the operating status of the component is obtained. Based on this index, risk classification and operation and maintenance instructions are generated, specifically including the following steps:

[0097] (1) Determine the severity level of the fault based on the fused characteristics, and determine the execution priority of the operation and maintenance instructions based on the severity level. Specifically, through a preset tiered fault judgment decision matrix, the comprehensive fault intensity index is... The structural consistency deviation entropy obtained in step S4 A multi-dimensional joint analysis is conducted. To ensure the objectivity of the rating, the system presets a comprehensive intensity early warning threshold. Comprehensive intensity critical threshold and structural distortion entropy threshold The formula for assessing the severity level of the fault is as follows:

[0098] (17)

[0099] in, The core logic behind the output severity level result lies in: when the combined value of temperature rise and structural damage... When the highest critical threshold is reached, or when the temperature rise reaches the warning value and is accompanied by obvious structural morphological abnormalities (high deviation entropy), the system automatically determines it to be in the highest level of "emergency" state.

[0100] (2) Assign the corresponding execution priority value according to the severity level. and the time limit for defect elimination Specifically, by establishing a hierarchical response mapping function, the abstract levels are transformed into digital indicators with time constraints. The calculation formulas for execution priority and response time limit are as follows:

[0101] (18)

[0102] in, This is a linear mapping function from rank to weight. This is an inverse proportional function from priority to time limit. Specifically: when When set to "emergency", The highest level is 1, corresponding to Within 24 hours; when When it is "significant", set The value is 2, corresponding to Within 72 hours; when When set to "normal", It is 3, corresponding to Within 15 working days. Through this quantification mechanism, the system can provide operations and maintenance personnel with clear guidance on the urgency of tasks.

[0103] (3) Generate operation and maintenance instructions based on the identified abnormal states and power logic attributes, and push them to the terminal. The specific steps include:

[0104] 1) Perform data encapsulation of operation and maintenance instructions. The cloud management and control platform will structurally integrate the power logic attributes determined in step S4 (including the line name, tower number, and phase information of the faulty component), the severity level, priority coefficient, and fault elimination time limit determined in this step, as well as the timestamp of the fault occurrence, the geographical coordinates of the fault point, and the fault comparison map (including the visible light annotation map and the infrared alignment map) to generate a standardized fault elimination task dataset.

[0105] 2) Intelligent matching of operation and maintenance entities. The system calls the business rule engine to compare the logical geographical coordinates of the fault point with the grid management data in the power ledger, automatically retrieves the operation and maintenance team responsible for the transformer area or the line of the voltage level and its current workload status, and determines the specific team ID or individual account identifier that receives the instruction.

[0106] 3) Real-time push via encrypted link. The cloud management platform uses a secure certificate-encrypted wireless communication link (such as a 5G sliced ​​network or a power wireless private network) and the Message Queuing Telemetry Transport (MQTT) protocol or Representational State Transfer (REST) ​​interface to push the encapsulated fault elimination task dataset in the form of structured messages to the mobile handheld terminals of designated maintenance personnel. After receiving the message, the terminal triggers a vibration or voice alarm and displays a task card on the interface. Maintenance personnel can directly access the logical location and visual diagnostic evidence of the fault point through the terminal, realizing a digital workflow from discovery to handling.

[0107] This step, through deep fusion of multimodal data and a quantitative priority evaluation mechanism, solves the problems of high false alarm rates from single sensors and strong subjectivity in manual grading, achieving a closed-loop process from "visual discovery" to "logical location" and then to "precise dispatching." The technical effect is a significant improvement in the standardization of operation and maintenance decisions. Through tiered time constraints and automated matching and push notifications, it ensures that high-risk hazards are prioritized for handling, greatly reducing the possibility of hidden defects in the distribution network escalating into sudden accidents.

[0108] Example 2, Figure 4 An AI-based unmanned aerial vehicle (UAV) inspection and control system for power distribution networks is presented, including:

[0109] The image acquisition module is used to adjust the drone's pose in real time based on target perception feedback during drone inspection to actively align with the target and acquire images of power components containing multimodal information.

[0110] The edge filtering module is used to filter out valid images from the image and then upload them to the cloud management platform;

[0111] The spatial semantic modeling module is used to identify electrical component entities in the filtered images and construct a spatial semantic graph based on the spatial distribution associations between components.

[0112] The intelligent reasoning and diagnosis module is used to perform topological logic reasoning and feature consistency analysis on the spatial semantic graph, identify abnormal states of components, and determine the power logic attributes of faulty components.

[0113] The operation and maintenance management module is used to generate operation and maintenance instructions based on the identified abnormal states and power logic attributes, and push them to the terminal.

[0114] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0115] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0116] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0117] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0118] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0119] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for the inspection and control of power distribution networks using unmanned aerial vehicles (UAVs) based on artificial intelligence, characterized in that, Includes the following steps: Based on target perception feedback, the drone's pose is adjusted in real time to align with the target, and images of power components containing multimodal information are acquired. Valid images are selected from the images and uploaded to the cloud management platform; Identify electrical component entities in the valid images and construct a spatial semantic map based on the spatial distribution relationship between the components; Topological logic reasoning and feature consistency analysis are performed on the spatial semantic graph to identify abnormal states of components and determine the power logic attributes of faulty components. Based on the abnormal state and power logic attributes, operation and maintenance instructions are generated and pushed to the terminal; The topological logic reasoning of the spatial semantic graph includes: The spatial semantic graph is input into a preset graph neural network model to generate enhanced node features that incorporate topological context information; The enhanced node features are mapped to a preset power logic label space, and the probability distribution of each power component entity under different logic categories is calculated. Based on the probability distribution, the power logic attributes of the faulty component are determined, and the power logic attributes include at least one of the power phase, circuit number, or connection level to which the faulty component belongs. The feature consistency analysis includes: A self-supervised feature reconstruction network is used to reconstruct local image features of power component entities, and the structural consistency deviation entropy between the original image features and the reconstructed image features is calculated. If the deviation entropy exceeds a preset threshold, the component is determined to have an atypical morphological abnormality.

2. The method for unmanned aerial vehicle (UAV) inspection and control of power distribution networks based on artificial intelligence as described in claim 1, characterized in that, The real-time adjustment of the UAV pose based on target perception feedback includes: Feature points of power components in inspection images are extracted in real time, and their image coordinates are obtained. The feature points are key points that characterize the geometric structure or semantic location of the power components. Calculate the feature deviation between the current position of the feature point and the preset composition position; Based on the feature deviation, control commands are generated to adjust the attitude of the UAV or gimbal in real time, so that the feature points move toward the preset mapping position, thereby achieving active tracking of the target.

3. The method for unmanned aerial vehicle (UAV) inspection and control of power distribution networks based on artificial intelligence as described in claim 2, characterized in that, Filtering the images into valid ones includes: Image processing algorithms are used to perform exposure detection and sharpness evaluation on images, and images that do not meet the preset standards are removed. The validity of an image is determined based on the proportion of electrical components in the image, and invalid or redundant images are removed.

4. The method for unmanned aerial vehicle (UAV) inspection and control of power distribution networks based on artificial intelligence as described in claim 3, characterized in that, The construction of the spatial semantic graph includes: Identify electrical component entities as graph nodes; Based on the spatial relationships of electrical component entities in the image, establish graph edges connecting each image node; Visual features of the power component entities are extracted to generate node attributes that include component category features and morphological features, thereby forming a spatial semantic map that integrates spatial structure and component features.

5. The method for unmanned aerial vehicle (UAV) inspection and control of power distribution networks based on artificial intelligence according to claim 4, characterized in that, Before generating the operation and maintenance instructions, the following is also included: Spatial matching and data fusion are performed on damage features in visible light images and temperature field distribution features in infrared thermal imaging images; The severity level of the fault is determined based on the merged characteristics, and the execution priority of the operation and maintenance instructions is determined based on the severity level.

6. The method for unmanned aerial vehicle (UAV) inspection and control of power distribution networks based on artificial intelligence according to claim 5, characterized in that, Before acquiring images of power components containing multimodal information, the process also includes: obtaining GIS topology data, tower coordinates, and equipment ledgers of the distribution network to be inspected, and generating the optimal trajectory instruction set for the task using a multi-objective path planning algorithm.

7. The method for monitoring and controlling unmanned aerial vehicles (UAVs) in power distribution networks based on artificial intelligence according to claim 6, characterized in that, The spatial relationships of the electrical component entities in the image include at least one of the following: relative positional relationships, adjacency relationships, and physical connection relationships between components.

8. A system using the artificial intelligence-based unmanned aerial vehicle (UAV) inspection and control method for power distribution networks as described in any one of claims 1-7, characterized in that, include: The image acquisition module is used to adjust the UAV's pose in real time based on target perception feedback to align with the target and acquire images of power components containing multimodal information. The edge filtering module is used to filter out valid images from the image and upload them to the cloud management platform; The spatial semantic modeling module is used to identify electrical component entities in the valid images and construct a spatial semantic map based on the spatial distribution relationship between the components; The intelligent reasoning and diagnosis module is used to perform topological logic reasoning and feature consistency analysis on the spatial semantic graph, identify abnormal states of components, and determine the power logic attributes of faulty components. The operation and maintenance management module is used to generate operation and maintenance instructions based on the abnormal status and power logic attributes, and push them to the terminal.

Images