Transformer abnormal state intelligent identification method, system and device

Abstract

The application provides a transformer abnormal state intelligent identification method, system and device, including a patrol inspection step, obtaining substation environment images shot by a UAV, and constructing a three-dimensional substation map; a transformer defect screening step, screening abnormal transformers in the three-dimensional substation map through target detection, and obtaining a UAV flight path; a flight path optimization step, calculating defect detection values and flight costs of each coordinate point, dynamically adjusting the UAV flight path based on a dynamic balance strategy, identifying corners of the transformer, generating virtual coordinate points, and planning an equidistant shooting path around the virtual marker point according to the virtual coordinate points; and a multi-modal identification step, based on the UAV flight path, synchronously collecting visual images, infrared thermal images and sound data shot by the UAV, and identifying the abnormal state of the transformer through a multi-modal data fusion strategy.

Description

Technical Field

[0001] This invention relates to transformer identification, and more particularly to intelligent identification methods, systems and devices for abnormal transformer conditions. Background Technology

[0002] Traditional substation transformer inspections rely on manual climbing, which suffers from low efficiency, high safety risks, and insufficient detection accuracy. Existing drone inspection technologies mostly employ fixed path planning, failing to dynamically adjust based on equipment inspection priorities and environmental risks. Furthermore, they lack geometric adaptation for shooting paths on complex-shaped equipment, and data from single sensors can easily lead to missed defects. This invention significantly improves inspection efficiency and accuracy through dynamic path optimization and multimodal data fusion. Summary of the Invention

[0003] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a method, system and device for intelligent identification of abnormal transformer states, so as to overcome the above-mentioned defects in the existing technology.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] Intelligent identification methods for transformer abnormal conditions include:

[0006] The inspection process involves acquiring images of the substation environment taken by drones and constructing a 3D map of the substation.

[0007] The transformer defect screening step involves screening abnormal transformers in the 3D map of the substation through target detection and obtaining the flight path of the UAV.

[0008] The flight path optimization steps include calculating the defect detection value and flight cost of each coordinate point, dynamically adjusting the UAV flight path based on a dynamic balance strategy, identifying the edges of the transformer, generating virtual coordinate points, and planning an equidistant shooting path around the virtual marker points based on the virtual coordinate points.

[0009] The multimodal recognition step involves simultaneously acquiring visual images, infrared thermal images, and sound data captured by the drone based on its flight path. Through a multimodal data fusion strategy, abnormal states of the transformer are identified.

[0010] As a preferred embodiment, a defect assessment strategy is included, which is used to acquire historical defect data of the substation and obtain a three-dimensional map of the substation. Based on the current environmental image features in the three-dimensional image of the substation and the historical defect data, the defect type and defect probability of each coordinate point are assessed, and the defect detection value is calculated based on the defect type and defect probability of each coordinate point.

[0011] Preferably, a flight cost assessment strategy is included, which includes calculating the distance between each coordinate point and the obstacle based on the three-dimensional map of the substation, obtaining the flight energy consumption of the UAV based on the UAV flight path, obtaining the distribution of the obstacle, obtaining the UAV flight posture based on the distribution of the obstacle, and obtaining the flight cost based on the distance between each coordinate point and the obstacle, the flight energy consumption, and the difficulty of the flight posture.

[0012] Preferably, the inspection step includes a 3D map component sub-step, which comprises:

[0013] The coordinate system component steps automatically generate virtual marker points at the four corners of the equipment, with the geometric center of the transformer as the origin;

[0014] The contour fitting step involves acquiring the original point cloud through LiDAR scanning, performing noise reduction processing, extracting the surface data of the equipment through power separation, and fitting a 3D model to generate the equipment contour.

[0015] The obstacle map construction steps involve identifying static and dynamic obstacles using substation environmental images. When an obstacle is static, its area is marked. When an obstacle is dynamic, its type is analyzed. If the obstacle type is a power transmission line, a swing model is constructed based on the current environmental characteristics, and the obstacle's movement area is marked according to the swing model. If the obstacle type is a moving object, a Brownian motion model is used to predict the obstacle's movement trajectory. Based on the obstacle's movement trajectory and the area of ​​the obstacle, a 3D map of the substation is constructed.

[0016] Preferably, the coordinate system component step includes a virtual point component sub-step, which includes obtaining the outer contour dimensions of the transformer, calculating the transformer coordinate offset, using AprilTag visual detection to determine whether recognition is successful, parsing the 3D position of the marker point and correcting the coordinates of the virtual marker point if recognition is unsuccessful, and starting the Qiguang radar point cloud matching and generating a safety buffer layer by extrapolating the distance along the normal vector direction with the equipment surface as the reference plane.

[0017] Preferably, the multimodal recognition step includes:

[0018] The spatiotemporal reference synchronization step involves acquiring data collected by each sensor and synchronizing the timestamps of each sensor.

[0019] The three-level fusion analysis steps generate spatiotemporally aligned raw data and obtain the abnormal features of each sensor. A multi-feature association model is constructed through a graph neural network, and the abnormal features are output through a weighting algorithm based on the saliency of the abnormal features and the reliability of the sensors.

[0020] The anomaly diagnosis output step outputs the anomaly type based on the anomaly feature type.

[0021] Preferably, the flight path optimization step includes an equidistant shooting strategy, which includes obtaining the virtual coordinate points of the transformer, obtaining the optimal shooting distance based on the actual size of the transformer and camera parameters, and generating an equidistant shooting path using two adjacent virtual coordinate points as path control points based on the optimal shooting distance and a fifth-order polynomial difference algorithm.

[0022] The intelligent identification system for abnormal transformer conditions includes:

[0023] An environmental perception module is used to collect environmental data of the substation and multimodal information of the transformer. The multimodal information includes image information, infrared information, temperature information, and UAV flight attitude and position information.

[0024] The path planning module is equipped with a defect assessment strategy and a flight cost assessment strategy. It calculates the defect detection value and flight value of each coordinate point, dynamically adjusts the UAV flight path based on the dynamic balance strategy, identifies the corners of the transformer, generates virtual coordinate points, and plans an equidistant shooting path around the virtual marker points based on the virtual coordinate points.

[0025] The control execution module controls the drone's flight based on its flight path and collects multimodal information.

[0026] The data processing module acquires multimodal information collected by the drone and obtains abnormal status information of the substation based on the multimodal data fusion strategy.

[0027] The intelligent identification device for abnormal transformer conditions includes a drone equipped with a visual camera, an infrared thermal imager, a temperature sensor, an inertial measurement unit, and GPS.

[0028] The beneficial effects of this invention are as follows: By quantifying the balance between defect detection value and flight costs, drones prioritize inspection of high-risk areas, reducing redundant paths and improving detection efficiency by over 30%. Marker points are designed to assist the path at the corners of square transformers, ensuring equidistant shooting from all sides, improving image analysis accuracy by 25%, and reducing the missed detection rate of minor defects such as edge cracks. Multi-source data from vision, infrared, and sound are integrated, and a cross-modal correlation model is constructed using graph neural networks to overcome the limitations of single sensors. For example, combining infrared hotspots and visual crack features can accurately distinguish between surface damage and internal overheating defects, increasing the overall detection accuracy to over 95%. The 3D map updates obstacle information in real time, and Brownian motion models are used to predict the trajectory of dynamic obstacles (such as swaying power lines). Combined with a safety buffer layer design, the drone's collision risk is reduced by 80%, making it suitable for complex power grid environments.

[0029] Specific implementation Attached Figure Description

[0030] Figure 1 This is an overall flowchart of the present invention;

[0031] Figure 2 This is a flowchart of the flight optimization strategy of the present invention;

[0032] Figure 3 This is a flowchart of the flight cost assessment strategy of the present invention;

[0033] Figure 4 This is a flowchart of the multimodal data fusion steps of the present invention;

[0034] Figure 5 This is a module connection diagram of the present invention. Detailed Implementation

[0035] 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] It should be noted that when a component is described as "fixed to" another component, it can be directly on the other component or may have a component in between. When a component is considered "connected to" another component, it can be directly connected to the other component or may have a component in between. When a component is considered "set on" another component, it can be directly set on the other component or may have a component in between. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0037] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0038] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings:

[0039] like Figure 1-5 As shown, the present invention provides an intelligent identification method for transformer abnormal states, characterized in that it includes:

[0040] The inspection process involves acquiring environmental images of the substation taken by drones and constructing a 3D map of the substation. During the inspection phase, the drone, as the core tool for data acquisition, uses its onboard high-definition camera to capture comprehensive images of the substation environment. The drone flies over the substation along a pre-set initial flight path, taking numerous multi-angle images. These images contain various information about the substation's equipment, buildings, and lines. Subsequently, 3D modeling technology is used to construct a 3D map of the substation based on the captured images. This 3D map intuitively and three-dimensionally presents the substation's spatial layout, equipment locations, and other information, providing fundamental data support for subsequent analysis and operations.

[0041] The inspection process includes a 3D map component sub-step, which includes:

[0042] The coordinate system construction step automatically generates virtual corner markers for the equipment, using the transformer's geometric center as the origin. Utilizing specialized 3D modeling algorithms, this data is processed to fit a 3D model, ultimately generating a precise equipment outline. In this process, the algorithm analyzes the spatial distribution and distance relationships of the point cloud data, connecting discrete point cloud data into a continuous surface. This achieves digital modeling of the transformer and other equipment's appearance, accurately representing the equipment's 3D structure in virtual space. The automatic generation of virtual corner markers, using the transformer's geometric center as the origin, provides a unified reference coordinate system for subsequent data acquisition and analysis. The virtual point construction sub-step first obtains the transformer's outer contour dimensions and calculates the transformer's coordinate offset. This step, based on actual measurement data, determines the transformer's approximate position and shape parameters in space.

[0043] The coordinate system component step includes a virtual point component sub-step. This sub-step involves acquiring the transformer's outer contour dimensions and calculating its coordinate offset. AprilTag vision detection is used to determine successful recognition. If successful, the 3D position of the marker point is parsed, and the virtual marker point coordinates are corrected. If recognition fails, LiDAR point cloud matching is initiated, and a safety buffer layer is generated by extrapolating the distance along the normal vector direction using the device surface as the reference plane. AprilTag is a computer vision-based marking system. It places specific label patterns on target objects. After a camera on a drone captures the label, the system can quickly identify the label's ID and attitude information. If recognition is successful, the 3D position of the marker point is parsed, and the virtual marker point coordinates are corrected to ensure a high degree of alignment between the virtual marker point and the actual device position. If recognition fails, LiDAR point cloud matching is initiated. LiDAR acquires point cloud data of the device surface by emitting laser beams and receiving reflected signals. Using the device surface as a reference plane, a safety buffer layer is generated by extrapolating the distance along the normal vector direction. This prevents the drone from colliding with the device during subsequent flight due to the inability to accurately identify virtual marker points, thus ensuring the safety of drone flight and data acquisition.

[0044] The contour fitting step involves acquiring the original point cloud through LiDAR scanning, performing noise reduction, extracting surface data of the equipment through power supply segmentation, and then fitting a 3D model to generate the equipment contour. The original point cloud data acquired through LiDAR scanning typically contains a large amount of noise due to environmental interference and other factors. This noise can affect the accuracy of subsequent data processing, so noise reduction is necessary to remove unnecessary interfering data. Next, surface data of the equipment is extracted through power supply segmentation. This utilizes the differences in power supply characteristics of different parts of the equipment to filter out data relevant to the equipment surface.

[0045] The obstacle map construction process involves identifying static and dynamic obstacles using substation environmental images. For static obstacles, the system marks their location. For dynamic obstacles, the system analyzes their type. If the obstacle is a power line, a swaying model is constructed based on current environmental characteristics, and the obstacle's movement area is marked according to the swaying model. If the obstacle is a moving object, a Brownian motion model is used to predict its trajectory. Based on the obstacle's trajectory and the obstacle area, a 3D map of the substation is constructed. Utilizing substation environmental images captured by drones, the system employs image recognition technology to distinguish between static and dynamic obstacles. When a static obstacle is detected, its location is marked using image analysis algorithms, clearly defining its position within the substation. When encountering a dynamic obstacle, the system further analyzes its type. If the obstacle is a power line, which sways under the influence of wind and other environmental factors, a swaying model is constructed based on current environmental characteristics such as wind speed, direction, and temperature. This model, based on the principles of physics and mechanics, simulates the swaying amplitude and trajectory of power transmission lines under different environmental conditions, and then marks the movement area of ​​obstacles based on the swaying model. If the obstacle type is a moving object, a Brownian motion model is used to predict the obstacle's movement trajectory. Brownian motion is a mathematical model used to describe the random motion of microscopic particles; here, the motion of the moving object is analogized to Brownian motion. By analyzing the historical motion data, speed, and direction changes of the moving object, its possible future movement trajectory is predicted. Based on the obstacle movement trajectory and obstacle area information, a 3D map of the substation containing detailed information on equipment and obstacles is finally constructed. This map not only shows the spatial layout of the substation and the location of equipment, but also intuitively presents potential danger zones, providing comprehensive spatial information support for UAV flight path planning and transformer anomaly identification.

[0046] The transformer defect screening process involves using target detection to identify abnormal transformers on a 3D map of the substation and obtaining the drone's flight path. Based on the constructed 3D map, a target detection algorithm is applied to screen transformers within the map. This algorithm automatically identifies all transformers on the 3D map and determines their anomalies through feature analysis. Once an abnormal transformer is detected, the system plans a flight path for the drone to reach its location based on the transformer's position on the 3D map, combined with the substation's spatial layout and equipment distribution. This flight path considers factors such as obstacles and space constraints to ensure the drone can safely and quickly reach the target location.

[0047] The flight path optimization process involves calculating the defect detection value and flight cost at each coordinate point, dynamically adjusting the UAV's flight path based on a dynamic balancing strategy, identifying the transformer's edges and corners to generate virtual coordinate points, and planning an equidistant shooting path around the virtual marker points based on these virtual coordinate points. 1. In the flight path optimization process, the defect detection value and flight cost at each coordinate point must first be calculated. The determination of the defect detection value depends on the defect assessment strategy. By analyzing historical defect data of the substation and current environmental image features, the possible defect types and probabilities at each coordinate point are assessed, thereby calculating the importance of that coordinate point for detecting transformer defects, i.e., the defect detection value. The flight cost comprehensively considers factors such as the UAV's flight distance, altitude changes, energy consumption, flight time, and potential risks to calculate the cost required for the UAV to fly between different coordinate points. Based on the dynamic balancing strategy, the system dynamically adjusts the UAV's flight path according to the calculated defect detection value and flight cost. The core idea of ​​the dynamic balancing strategy is to minimize the UAV's flight cost while ensuring efficient defect detection. The system continuously compares the comprehensive benefits of different path schemes and selects the path that covers high-value detection points while minimizing flight costs. During the flight path optimization process, the system also identifies the transformer's edges and corners, generating virtual coordinate points based on their locations. These virtual coordinate points are set around key parts of the transformer to provide a more comprehensive view of its external features. Based on these virtual coordinate points, the system plans an equidistant shooting path around the virtual markers. The drone flies along this path, enabling it to photograph the transformer from multiple angles and distances, acquiring richer and more comprehensive image data.

[0048] This includes a defect assessment strategy, which acquires historical defect data and a 3D map of the substation. Based on the current environmental image features in the 3D substation image and historical defect data, the strategy assesses the defect type and probability at each coordinate point, and calculates the defect detection value based on these parameters. All transformer defect records since the substation was put into operation are collected, covering detailed information such as the time of occurrence, specific location coordinates, defect type (e.g., winding fault, core overheating, insulation aging), handling method, and handling results. For example, if transformers in a certain area frequently experienced core overheating during the high-temperature summer period in the past three years, these records will become important data for subsequent analysis. The strategy also combines the constructed 3D substation map with current environmental image features, including the transformer's appearance (e.g., presence of oil leaks, broken insulators), surrounding equipment layout, ambient temperature and humidity, and lighting conditions. For example, if the 3D map reveals a transformer with limited space for heat dissipation and a high ambient temperature, this information will be included in the assessment system. Finally, machine learning algorithms, such as decision trees and random forests, are used to perform correlation analysis between historical defect data and current environmental image features. Taking core overheating defects as an example, the algorithm analyzes the combination of factors such as ambient temperature, load conditions, and equipment age that make this defect more likely to occur. By learning from a large amount of historical data, a relationship model is established between defect types and various influencing factors. Based on this model, the probability of different defect types occurring at each coordinate point is calculated. For example, after analysis, under the current environment, the probability of a transformer at a certain coordinate point experiencing winding faults is 0.3, the probability of core overheating is 0.2, and the probability of insulation aging is 0.1. These probability values ​​provide an important foundation for subsequently determining the value of defect detection. Based on the defect type and probability at each coordinate point, combined with the severity and difficulty of detection, the defect detection value is calculated. Defect types that may lead to serious consequences, such as winding faults, are assigned higher weights; defects that are easier to detect are assigned lower weights. For example, if the probability of a winding fault at a certain coordinate point is 0.3, and this defect could lead to a large-scale power outage, its severity is high, so it is assigned a weight of 0.8; the probability of insulation aging is 0.1, and its severity is relatively low, so it is assigned a weight of 0.3. The defect detection value of the coordinate point is calculated using a formula, such as: Defect detection value = winding fault probability × winding fault weight + insulation aging probability × insulation aging weight + ..., thereby determining the importance of the coordinate point for detecting transformer defects.

[0049] This includes a flight cost assessment strategy. This strategy involves calculating the distances between each coordinate point and obstacles based on a 3D map of the substation, obtaining the drone's flight energy consumption based on its flight path, acquiring the obstacle distribution, determining the drone's flight attitude based on the obstacle distribution, and calculating the flight cost based on the distances between each coordinate point and obstacles, flight energy consumption, and flight attitude difficulty. In the substation's 3D map, each coordinate point is used as a reference, and spatial geometric algorithms are used to calculate the shortest distance between it and surrounding obstacles (including buildings, other equipment, power lines, etc.). For example, for a given coordinate point, the 3D map data calculates its distance to the nearest building to be 5 meters and its distance to the adjacent power line to be 3 meters. These distance data reflect the collision risk faced by the drone when flying near that coordinate point. The closer to an obstacle, the higher the flight cost. A distance-cost function is defined; for example, when the distance is less than a safety threshold (assumed to be 2 meters), the flight cost increases exponentially; as the distance increases, the rate of increase in flight cost gradually slows down. This transforms the distance factor into a quantifiable flight cost indicator. 1. Establish a flight energy consumption model based on the drone's model, performance parameters (such as motor power, battery capacity, etc.), and factors such as flight speed and altitude. For example, for a certain model of drone, the power consumption per kilometer of flight in level flight is proportional to the square of the flight speed; during climb, the power consumption per meter of ascent is related to the drone's payload and ascent rate. Calculate flight energy consumption in real time based on the drone's actual flight path and the energy consumption model. Assume the drone flies from coordinate point A to coordinate point B, a distance of 2 kilometers, with one climb of 10 meters. The energy consumption for this segment is calculated using the energy consumption model to be a certain value (e.g., 500mAh), and this energy consumption value is included in the flight cost assessment. Based on the actual flight path of the UAV and combined with an energy consumption model, the flight energy consumption is calculated in real time. Assuming the UAV flies from coordinate point A to coordinate point B, a distance of 2 kilometers, with one climb (10 meters altitude increase), the energy consumption for this flight segment is calculated using the energy consumption model to be a certain value (e.g., 500mAh), and this energy consumption value is included in the flight cost assessment. 1. Analyze the distribution of obstacles in the 3D map of the substation, including the shape, size, location, and movement trajectory of dynamic obstacles. For example, in a certain area, there are multiple intersecting power lines, some of which sway under wind, and other equipment obstructs the flight path. Based on the obstacle distribution, plan the UAV's flight attitude.When encountering narrow passages, drones may need to adjust to a side-flying posture; when filming equipment at high altitudes, pitch and yaw angles need to be adjusted. Different flight postures correspond to different operational difficulties; the greater the operational difficulty, the higher the flight cost. For example, a side-flying posture requires more precise attitude control from the drone, and its flight cost will increase by a certain percentage (e.g., 20%) compared to a level flight posture. The flight cost is calculated by comprehensively considering factors such as the distance between each coordinate point and obstacles, flight energy consumption, and flight posture difficulty. A weighted summation method is used to assign a corresponding weight to each factor. For example, the distance factor has a weight of 0.4, the energy consumption factor has a weight of 0.3, and the flight posture difficulty factor has a weight of 0.3. The flight cost is calculated using the formula: Flight Cost = Distance Cost × Distance Weight + Energy Consumption Cost × Energy Consumption Weight + Flight Posture Difficulty Cost × Flight Posture Difficulty Weight. This provides a quantitative reference for optimizing the drone's flight path, enabling the drone to minimize flight costs and risks while ensuring detection effectiveness.

[0050] The multimodal recognition step, based on the drone's flight path, simultaneously acquires visual images, infrared thermal images, and sound data captured by the drone. Through a multimodal data fusion strategy, it identifies abnormal states of the transformer. During the drone's optimized flight path, multiple types of data are simultaneously collected, including visual images, infrared thermal images, and sound data. Visual images reveal the transformer's appearance, surface damage, stains, etc.; infrared thermal imaging detects temperature distribution across different parts of the transformer, identifying potential overheating faults; and sound data reflects the transformer's operational acoustic characteristics, determining the presence of abnormal vibrations or noise. The multimodal data fusion strategy integrates and analyzes these different data types. This strategy utilizes machine learning algorithms and data processing techniques to extract key features from each modality and fuse these features to form a more comprehensive and accurate feature description. Based on the fused features, the system can more accurately identify abnormal states of the transformer, improving the accuracy and reliability of the identification.

[0051] Multimodal recognition steps include:

[0052] The spatiotemporal reference synchronization step involves acquiring data from each sensor and timestamping each sensor. In the complex operating environment of a substation, devices such as visual cameras, infrared thermal imagers, and sound sensors mounted on a drone simultaneously collect data. However, due to differences in the operating frequencies and data transmission speeds of each sensor, the collected data may deviate in time and space. Therefore, the spatiotemporal reference synchronization step is crucial, serving as the foundation for subsequent data analysis. First, the system acquires data from each sensor in real time and extracts the timestamp information contained within the data. The timestamp records the specific moment of data acquisition and is the key basis for achieving time synchronization. For time synchronization, a high-precision clock synchronization algorithm, such as Network Time Protocol (NTP) or Precise Time Protocol (PTP), is used to unify the timestamps of each sensor onto a standard time reference, ensuring consistency in the time dimension of data collected by different sensors. For spatial synchronization, based on the substation's 3D map and the drone's real-time positioning information (such as obtained through GPS or inertial navigation systems), the spatial location corresponding to the data acquisition time of each sensor is determined. By establishing a unified spatial coordinate system, data collected by different sensors can be mapped to this coordinate system, enabling accurate spatial correspondence and generating spatiotemporally aligned raw data. For example, the location of a transformer in a visual image can be matched with the location during infrared thermal imaging and sound data acquisition, ensuring that subsequent analysis involves multimodal data from the same time and location.

[0053] The three-level fusion analysis step generates spatiotemporally aligned raw data and acquires anomalous features from each sensor. A multi-feature association model is constructed using a graph neural network, and based on the saliency of the anomalous features and sensor reliability, a weighted algorithm outputs the anomalous features. After completing spatiotemporal benchmark synchronization, spatiotemporally aligned raw data is obtained, which then proceeds to the three-level fusion analysis step. This step deeply mines anomalous features from the data through three layers of processing.

[0054] The spatiotemporally aligned raw data undergoes preprocessing, including image denoising and audio signal filtering, to improve data quality. Then, anomalous features are extracted from visual images, infrared thermal imaging, and audio data. In visual images, computer vision algorithms, such as convolutional neural networks (CNNs), are used to identify anomalous features on the transformer surface, such as cracks, oil leaks, and loose components. Infrared thermal imaging data is analyzed using thermal imaging algorithms to detect temperature anomalies in various parts of the transformer, such as hot spots and uneven temperature distribution. Audio data is processed using signal processing algorithms to extract features such as abnormal vibration frequencies and abnormal noise.

[0055] The extracted anomaly features from each sensor are used as nodes, and a multi-feature association model is constructed using a graph neural network (GNN). Graph neural networks can effectively handle data with complex relationships. In this model, different types of anomaly features are treated as nodes in the graph, and the relationships between features are treated as edges. For example, cracks on the surface of a transformer may be associated with internal vibration anomalies; this association is modeled and learned using a graph neural network. By training the graph neural network, the model can automatically uncover the potential correlations between various anomaly features, thereby gaining a more comprehensive understanding of the transformer's operating status.

[0056] In multi-feature association models, different anomalous features have varying degrees of significance, meaning they differ in their importance in determining transformer anomalies. Furthermore, the reliability of different sensors also varies. To accurately output features valuable for anomaly diagnosis, a weighting algorithm is introduced. This algorithm assigns appropriate weights to each anomalous feature based on its significance and the reliability of the sensors. For example, data collected by sound sensors, which are easily affected by environmental interference, are assigned relatively low weights; while visual image features that directly reflect the transformer's surface condition are assigned appropriate weights based on their degree of anomaly. Through weighted calculation, representative anomalous features are ultimately output, highlighting key anomaly information and providing accurate data support for subsequent anomaly diagnosis.

[0057] The anomaly diagnosis output step outputs the anomaly type based on the anomaly feature type. Based on the anomaly features output from the three-level fusion analysis step, the system proceeds to the anomaly diagnosis output step. The system has pre-established a correspondence database between anomaly features and anomaly types, constructed through learning and summarizing a large amount of historical data. For example, when a crack is detected on the transformer surface and the corresponding area shows an abnormal temperature rise, the database indicates an insulation fault; if an abnormal vibration frequency matches the frequency characteristics of a loose component, it is diagnosed as a mechanical component fault.

[0058] The system matches the output anomaly characteristics with standard characteristics in the relational database, determines the anomaly type of the transformer based on the matching results, and outputs the diagnostic results. The output anomaly type information can be presented in intuitive charts, text reports, and other forms, allowing maintenance personnel to quickly understand the transformer's abnormal situation, take timely and appropriate maintenance measures, and ensure the safe and stable operation of the substation.

[0059] Through the three sub-steps of spatiotemporal benchmark synchronization, three-level fusion analysis, and anomaly diagnosis output, the multimodal identification step achieves intelligent and accurate identification of transformer abnormal states, providing strong technical support for transformer operation and maintenance management.

[0060] The flight path optimization process includes an equidistant shooting strategy. This strategy involves acquiring virtual coordinates of the transformer, determining the optimal shooting distance based on the transformer's actual dimensions and camera parameters, and using adjacent virtual coordinates as path control points to generate an equidistant shooting path through a fifth-order polynomial difference algorithm. Computer vision algorithms, such as edge detection algorithms (Canny operator, etc.), are used to identify the transformer's contour edges, thereby determining the transformer's corner positions. Alternatively, based on point cloud data acquired by LiDAR, feature extraction algorithms are used to find key geometric feature points on the transformer surface. Based on these corners and key feature points, virtual coordinates of the transformer are automatically generated. The derivation is based on imaging principles and the geometric relationships of similar triangles. Specifically, the calculation process is based on imaging principles and the geometric relationships of similar triangles. Assume the actual dimensions of a key part of the transformer are... The ideal size for this part to appear in the image is The camera's focal length is The sensor size is According to the imaging formula (in The optimal shooting distance is calculated by combining the required image clarity based on sensor size and pixel resolution. For example, if a key interface of a transformer is desired to occupy a certain proportion of pixels in the image, the optimal shooting distance between the drone and the transformer can be determined using the above formula and related parameters. This ensures that the captured image can fully display the transformer details while maintaining image clarity and resolution sufficient for subsequent analysis. After obtaining the virtual coordinate points and the optimal shooting distance, an equidistant shooting path is generated using a fifth-order polynomial interpolation algorithm, with adjacent virtual coordinate points serving as path control points. The fifth-order polynomial interpolation algorithm generates smooth, continuous curves, meeting the drone's requirements for path smoothness during flight and preventing drastic attitude changes and speed jumps during flight, thus ensuring flight safety and shooting stability.

[0061] Let two adjacent virtual coordinate points be and The flight time range of the drone is The general form of a fifth-degree polynomial is: By constraining boundary conditions such as the location of path control points, the speed and acceleration of the UAV at the start and end points, the coefficients of the polynomial are solved. For example, requiring drones to start from the origin. The velocity at that point is acceleration is At the finish line The velocity at that point is acceleration is Substituting these conditions into the polynomial and its derivative expression, solving the simultaneous equations for the coefficients, we can obtain the result of the UAV's operation. The flight path function is defined for the given time period. Following this method, each group of adjacent virtual coordinate points is processed sequentially to generate a complete equidistant shooting path surrounding the virtual marker point of the transformer. When the drone flies along this path, it can capture images of the transformer at equal intervals from the optimal shooting distance.

[0062] The intelligent identification system for abnormal transformer conditions includes:

[0063] The environmental perception module is used to collect environmental data from the substation and multimodal information from the transformer. This multimodal information includes image data, infrared data, temperature data, and the UAV's flight attitude and position information. As the system's data source, the environmental perception module undertakes comprehensive, multi-dimensional data acquisition tasks. This unit is primarily mounted on a UAV platform and equipped with various high-precision sensors for collecting substation environmental data and multimodal information from the transformer.

[0064] Employing a high-resolution dual-spectrum camera, it can simultaneously acquire visible light and infrared images. Visible light images clearly reveal the transformer's external details, such as casing damage and oil leaks; infrared images are used to detect temperature distribution on the equipment's surface, identifying potential overheating faults. Equipped with a professional temperature sensor, it can monitor the temperature data of key parts of the transformer in real time, and combined with infrared thermal imaging data, more accurately determine the equipment's temperature status. An integrated high-precision GPS module and inertial measurement unit (IMU) acquire the UAV's flight attitude (pitch angle, yaw angle, roll angle) and position information (latitude, longitude, and altitude) in real time. This data is not only used for the UAV's own navigation and control but also provides a spatial positioning reference for subsequent data processing.

[0065] Each sensor operates synchronously according to a preset sampling frequency to ensure that the collected data is consistent in time. For example, the image acquisition frequency is set to 30 frames / second, the temperature sensor sampling frequency is 10 times / second, and the flight information acquisition frequency is 50 times / second.

[0066] The collected data will be precisely timestamped and transmitted in real time to the drone's data buffer via a high-speed data transmission link, awaiting further processing.

[0067] The path planning module is equipped with defect assessment and flight cost assessment strategies. These strategies calculate the defect detection value and flight value of each coordinate point, respectively. Based on a dynamic balancing strategy, the module dynamically adjusts the UAV's flight path, identifies transformer edges, generates virtual coordinate points, and plans equidistant shooting paths around virtual marker points. This path planning module is the system's "intelligent brain," responsible for planning the optimal flight path for the UAV to achieve efficient defect detection. This unit integrates the defect assessment and flight cost assessment strategies and achieves their organic combination through a dynamic balancing strategy.

[0068] First, historical defect data is extracted from the substation's historical defect database, including information such as the types of defects that occurred in previous transformers, their occurrence time, location, and corresponding environmental conditions. Simultaneously, by combining the substation's 3D map with current environmental image features, machine learning algorithms (such as random forests and support vector machines) are used to construct a defect prediction model.

[0069] This model analyzes each coordinate point on a 3D map of a substation, assessing the potential defect types and their probabilities at each point. For example, historical data reveals that transformers in a certain area are prone to winding overheating during the high-temperature summer months. Therefore, under the current summer conditions, the probability of winding overheating defects occurring at coordinate points in that area would be assessed as relatively high.

[0070] Based on the defect type and probability at each coordinate point, and considering factors such as defect severity and detection difficulty, the defect detection value is calculated. High-risk defects that may lead to serious consequences are assigned a higher detection value weight.

[0071] Based on a 3D map of the substation, the distance between each coordinate point and obstacles (including buildings, other equipment, power lines, etc.) is calculated. Using spatial geometry algorithms, the shortest distance between the UAV and obstacles is accurately measured when flying at different coordinate points. The closer the distance, the higher the flight risk and the greater the corresponding distance cost.

[0072] Based on the drone's flight path planning, and taking into account factors such as the drone's model parameters (e.g., motor power, battery capacity), flight speed, and altitude changes, a flight energy consumption model is established to predict the drone's energy consumption on different paths. For example, the climb process and high-speed flight will consume more power, and the corresponding energy cost will also increase.

[0073] Analyzing obstacle distribution determines the required flight attitude (such as level flight, side flight, and pitch flight) for the UAV when flying in different areas. Complex flight attitudes place higher demands on UAV control, increasing the difficulty of flight and thus raising the cost of flight attitude difficulty.

[0074] The flight cost is calculated by weighted summation, taking into account factors such as the distance between each coordinate point and obstacles, flight energy consumption, and the difficulty of flight posture. The weights of different factors can be dynamically adjusted according to actual mission requirements.

[0075] Based on a dynamic balancing strategy, the optimal balance between defect detection value and flight cost is sought. Intelligent optimization algorithms (such as genetic algorithms and particle swarm optimization) are employed to dynamically adjust the UAV's flight path. This ensures maximum coverage of high-defect-detection-value areas while reducing flight costs and improving inspection efficiency.

[0076] Computer vision technology is used to identify the edges and corners of a transformer. Edge detection algorithms (such as the Canny operator) are used to extract the outline edges of the transformer, thereby determining the position of the edges and corners and generating virtual coordinate points.

[0077] Based on virtual coordinate points, an equidistant shooting path is planned around the virtual marker points. By setting appropriate shooting radii and angle intervals, the drone can comprehensively photograph the transformer from multiple angles and different distances, acquiring rich detection data.

[0078] The control execution module controls the drone's flight based on its flight path and collects multimodal information.

[0079] The control execution module is the "hands and feet" of the system. It is responsible for precisely controlling the flight of the UAV according to the flight path generated by the path planning module and completing the task of collecting multimodal information.

[0080] The system receives flight path data from the path planning module and converts it into control commands for the UAV. Through the UAV's flight control system, it adjusts the UAV's flight attitude and speed in real time to ensure the UAV flies along the predetermined path.

[0081] By combining data from sensors onboard the drone (such as GPS and IMU), real-time flight path correction is achieved. When the drone deviates from the preset path, the flight control system automatically calculates the deviation and generates corresponding control signals to adjust the drone's flight direction, bringing it back to the correct path.

[0082] The system controls the sensor's operation based on the flight path and preset shooting points. When the drone flies to the designated coordinates, it triggers operations such as image acquisition and temperature monitoring to ensure that the collected data matches the path planning.

[0083] The acquired multimodal information undergoes preliminary processing and packaging, including data format conversion and compression, to reduce data transmission volume and improve data transmission efficiency. The processed data is then transmitted to the ground data processing module via a wireless communication link.

[0084] The data processing module acquires multimodal information collected by the UAV and, based on a multimodal data fusion strategy, obtains abnormal state information of the substation. This module serves as the system's "analysis center," responsible for in-depth processing and analysis of the multimodal information collected by the UAV, ultimately deriving abnormal state information of the transformer.

[0085] After receiving multimodal data transmitted from the UAV, the timestamp and spatial location information are first extracted from the data. A high-precision clock synchronization algorithm (such as Network Time Protocol NTP or Precision Time Protocol PTP) is used to uniformly calibrate the timestamps of data collected by different sensors to ensure data consistency in the time dimension.

[0086] Based on the 3D map of the substation and the spatial positioning information of UAVs, a unified spatial coordinate system is established. Data collected by different sensors are mapped to this coordinate system to achieve precise spatial alignment of the data, generating spatiotemporally aligned raw data, laying the foundation for subsequent analysis.

[0087] Three-level fusion analysis

[0088] Raw data preprocessing and anomaly feature extraction: The spatiotemporally aligned raw data undergoes preprocessing, such as image denoising (using algorithms like median filtering and Gaussian filtering) and audio signal filtering (removing background noise interference), to improve data quality. Then, anomaly features are extracted from visual images, infrared thermal imaging, and audio data respectively. Convolutional neural networks (CNNs) are used to analyze visual images to identify anomalies such as cracks and deformations on the transformer surface; infrared thermal imaging analysis algorithms are used to detect abnormal temperature areas on the transformer; and signal processing algorithms (such as Fast Fourier Transform, FFT) are used to perform spectral analysis on the audio data to extract abnormal vibration frequencies and noise features.

[0089] Constructing a multi-feature association model: The extracted abnormal features from each sensor are used as nodes to construct a multi-feature association model using a graph neural network (GNN). This model can uncover potential correlations between different types of abnormal features. For example, overheating inside a transformer may cause an increase in the outer casing temperature, accompanied by abnormal vibrations. By training the graph neural network, the correlation patterns between these features are learned, leading to a more comprehensive understanding of the transformer's operating status.

[0090] Anomaly Feature Output Based on Weighted Algorithm: Considering the saliency of different anomaly features and the varying reliability of different sensors, a weighted algorithm is introduced. Based on the importance of the anomaly feature in determining the transformer's abnormal state, as well as factors such as sensor stability and accuracy, a corresponding weight is assigned to each anomaly feature. For example, sound sensor data, which is highly susceptible to environmental interference, is assigned a relatively low weight; visual image features that directly reflect the transformer's surface condition are assigned appropriate weights based on their degree of anomaly. Through weighted calculation, key anomaly information is highlighted, and representative anomaly features are output.

[0091] The system pre-establishes a database of correspondences between anomaly features and anomaly types, built upon a large amount of historical fault data and expert experience. The anomaly features output after three levels of fusion analysis are then matched with the standard features in this database.

[0092] Based on the matching results, the type of transformer anomaly is determined, and a detailed anomaly diagnostic report is generated. The diagnostic report is presented in intuitive charts and text, including information such as the anomaly type, location, and severity assessment. This allows maintenance personnel to quickly understand the transformer's abnormal situation and take timely maintenance measures to ensure the safe and stable operation of the substation.

[0093] The intelligent identification device for transformer abnormal conditions includes a drone equipped with a visual camera, an infrared thermal imager, a temperature sensor, an inertial measurement unit, and GPS.

[0094] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent identification of abnormal transformer conditions, characterized in that, include: The inspection process involves acquiring images of the substation environment taken by drones and constructing a 3D map of the substation. The transformer defect screening step involves screening abnormal transformers in the 3D map of the substation through target detection and obtaining the flight path of the UAV. The flight path optimization steps include calculating the defect detection value and flight cost of each coordinate point, dynamically adjusting the UAV flight path based on a dynamic balance strategy, identifying the edges of the transformer, generating virtual coordinate points, and planning an equidistant shooting path around the virtual marker points based on the virtual coordinate points. The system includes a defect assessment strategy, which is used to acquire historical defect data of the substation and obtain a three-dimensional map of the substation. Based on the current environmental image features in the three-dimensional image of the substation and the historical defect data, the system assesses the defect type and defect probability at each coordinate point. Based on the defect type and defect probability at each coordinate point, the system calculates and obtains the defect detection value. The system includes a flight cost assessment strategy, which includes calculating the distance between each coordinate point and the obstacle based on a 3D map of the substation, obtaining the flight energy consumption of the UAV based on the UAV's flight path, obtaining the distribution of the obstacle, obtaining the UAV's flight posture based on the distribution of the obstacle, and obtaining the flight cost based on the distance between each coordinate point and the obstacle, the flight energy consumption, and the difficulty of the flight posture. The multimodal recognition step involves simultaneously acquiring visual images, infrared thermal images, and sound data captured by the drone based on its flight path. Through a multimodal data fusion strategy, abnormal states of the transformer are identified.

2. The intelligent identification method for abnormal transformer states according to claim 1, characterized in that, The inspection process includes a 3D map construction sub-step, which comprises: The coordinate system construction steps involve automatically generating virtual marker points at the four corners of the equipment, with the geometric center of the transformer as the origin. The contour fitting step involves acquiring the original point cloud through LiDAR scanning, performing noise reduction processing, extracting the surface data of the equipment through power separation, and fitting a 3D model to generate the equipment contour. The obstacle map construction steps are as follows: Static and dynamic obstacles are identified using substation environmental images. When an obstacle is static, its area is marked. When an obstacle is dynamic, its type is analyzed. If the obstacle type is a power transmission line, a swing model is constructed based on current environmental characteristics, and the obstacle's movement area is marked according to the swing model. If the obstacle type is a moving object, a Brownian motion model is used to predict the obstacle's trajectory. Based on the obstacle's trajectory and the area it is located in, a 3D map of the substation is constructed.

3. The intelligent identification method for abnormal transformer states according to claim 2, characterized in that, The coordinate system construction step includes a virtual point construction sub-step, which includes obtaining the outer contour dimensions of the transformer and calculating the transformer coordinate offset. The AprilTag visual detection is used to determine whether the recognition is successful. If the recognition is successful, the 3D position of the marker point is parsed and the coordinates of the virtual marker point are corrected. If the recognition is unsuccessful, the Qiguang radar point cloud matching is started, and a safety buffer layer is generated by extrapolating the distance along the normal vector direction with the device surface as the reference plane.

4. The intelligent identification method for abnormal transformer states according to claim 1, characterized in that, The multimodal recognition steps include: The spatiotemporal reference synchronization step involves acquiring data collected by each sensor and synchronizing the timestamps of each sensor. The three-level fusion analysis steps generate spatiotemporally aligned raw data and obtain the abnormal features of each sensor. A multi-feature association model is constructed through a graph neural network, and the abnormal features are output through a weighting algorithm based on the saliency of the abnormal features and the reliability of the sensors. The anomaly diagnosis output step outputs the anomaly type based on the anomaly feature type.

5. The intelligent identification method for abnormal transformer states according to claim 1, characterized in that, The flight path optimization step includes an equidistant shooting strategy, which includes obtaining the virtual coordinates of the transformer, obtaining the optimal shooting distance based on the actual size of the transformer and camera parameters, and generating an equidistant shooting path using two adjacent virtual coordinates as path control points based on the optimal shooting distance and a fifth-order polynomial difference algorithm.

6. A transformer abnormality state intelligent identification system, characterized in that, include: An environmental perception module is used to collect environmental data of the substation and multimodal information of the transformer. The multimodal information includes image information, infrared information, temperature information, and UAV flight attitude and position information. The path planning module is equipped with a defect assessment strategy and a flight cost assessment strategy. It calculates the defect detection value and flight value of each coordinate point, dynamically adjusts the UAV flight path based on the dynamic balance strategy, identifies the corners of the transformer, generates virtual coordinate points, and plans an equidistant shooting path around the virtual marker points based on the virtual coordinate points. The system includes a defect assessment strategy, which is used to acquire historical defect data of the substation and obtain a three-dimensional map of the substation. Based on the current environmental image features in the three-dimensional image of the substation and the historical defect data, the system assesses the defect type and defect probability at each coordinate point. Based on the defect type and defect probability at each coordinate point, the system calculates and obtains the defect detection value. The system includes a flight cost assessment strategy, which includes calculating the distance between each coordinate point and the obstacle based on a 3D map of the substation, obtaining the flight energy consumption of the UAV based on the UAV's flight path, obtaining the distribution of the obstacle, obtaining the UAV's flight posture based on the distribution of the obstacle, and obtaining the flight cost based on the distance between each coordinate point and the obstacle, the flight energy consumption, and the difficulty of the flight posture. The control execution module controls the drone's flight based on its flight path and collects multimodal information. The data processing module acquires multimodal information collected by the drone and obtains abnormal status information of the substation based on the multimodal data fusion strategy.

Images