Method and system for identifying transformer substation defects in real time based on unmanned aerial vehicle inspection

Abstract

The invention relates to the technical field of power transmission and transformation, in particular to a method and a system for recognizing transformer substation defects in real time based on unmanned aerial vehicle inspection, which are characterized in that a transformer substation inspection task is issued to an unmanned aerial vehicle carrying edge equipment through a control computer, and the inspection task comprises a set inspection route; the unmanned aerial vehicle patrols the transformer substation according to the set patrolling route and continuously shoots videos in the patrolling process; the edge device adopts a deep learning model to perform feature extraction on the shot video content to obtain a feature vector, and performs similarity calculation on the feature vector and the feature vector of the defect library photo; when the similarity reaches or exceeds a specific value, the edge device triggers a snapshot operation, and a snapshot photo is stored in a defect folder; and after the inspection task is finished, the unmanned aerial vehicle uploads the video shot in the inspection process and the defect folder to the control computer. According to the method, the defect identification time is greatly shortened, the substation inspection efficiency is improved, and the labor intensity is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of power transmission and transformation, and is a method and system for real-time identification of substation defects based on drone inspection. Background Art

[0002] In current power system operation and maintenance, the stable operation of substation equipment plays a vital role in ensuring reliable power supply across the entire power grid. Therefore, timely and accurate detection of substation equipment defects is crucial for ensuring safe power system operation. Currently, substation inspections primarily utilize a fixed-point inspection method that combines personnel, cameras, and drones.

[0003] During inspections, personnel shoulder the important responsibility of conducting comprehensive and routine inspections. During comprehensive inspections, staff members need to conduct detailed inspections of each type of equipment within the substation, including appearance, operating sounds, temperature, and other aspects, relying on their experience and expertise to determine whether any equipment is abnormal. Routine inspections, conducted according to fixed cycles and procedures, routinely check equipment to ensure that it is in normal operation. However, this manual inspection method has certain limitations. On the one hand, substation equipment is numerous and widely distributed, making it difficult for manual inspections to fully cover all equipment and every corner in real time, and potential defects may be missed. On the other hand, manual inspections are relatively inefficient, consuming a significant amount of manpower and time, and the effectiveness of the inspections depends largely on the professional level and work status of the staff.

[0004] To complement manual inspections, cameras and drones are widely used in substation inspections. Cameras are typically placed at strategic locations throughout the substation, capturing images of specific equipment at pre-defined locations. Drones, with their flexible maneuverability, can reach areas difficult for both humans and cameras to reach, similarly capturing images at pre-set locations on each piece of equipment. These images are transmitted back to the monitoring center and compared with typical images in a database of healthy equipment. This database is based on a large collection of photographs of properly functioning equipment, containing information such as the appearance characteristics of each type of equipment under normal conditions. This comparison identifies photos that do not match the images in the healthy equipment database. In theory, these mismatches could indicate equipment defects.

[0005] However, in practice, this inspection method has exposed numerous problems. First, because photos can only be taken at preset locations, existing technology cannot provide a comprehensive inspection of the equipment. Substation equipment is complex, and defects may occur at different angles and locations. However, taking photos at preset locations may not capture anomalies in hidden parts of the equipment or at specific angles, making some defects difficult to detect in a timely manner. Second, the number of mismatched photos is excessive. In actual operation, a medium-sized substation can generate over 500 mismatched photos per inspection. This massive number of photos requires manual re-screening to determine which photos correspond to defective equipment. This undoubtedly places a huge workload on staff, consumes a lot of time, and easily causes fatigue from long periods of repetitive work, which in turn affects the accuracy and efficiency of the screening.

[0006] Furthermore, when staff have questions about the content of photos, existing inspection methods lack effective verification methods. Currently, judgments can only be made based on photos taken at fixed, pre-set locations, without the ability to obtain equipment information from a wider range of angles and time periods for reference. If an anomaly is unclear or difficult to identify in the photos, staff are forced to conduct on-site inspections at the substation. This not only increases the complexity and cost of the work, but also can delay on-site inspections, potentially preventing potential equipment defects from being addressed in a timely manner, posing a risk to the safe and stable operation of the power system. Summary of the Invention

[0007] The present invention provides a method and system for real-time identification of substation defects based on drone inspections, which overcomes the shortcomings of the above-mentioned existing technologies. It can effectively solve the problems that the existing technologies cannot conduct comprehensive inspections of equipment, the workload of manual screening of mismatched photos is large, and on-site inspections are required when there are questions about the content of photos.

[0008] One of the technical solutions of the present invention is achieved through the following measures: a method for real-time identification of substation defects based on drone inspection, comprising the following steps: S1. Task issuance: The computer is used to issue a substation inspection task to the drone equipped with the edge device, and the inspection task includes a set inspection route. S2. Inspection and video collection: The drone inspects the substation according to the set inspection route and continuously captures videos during the inspection process; S3. Real-time defect comparison and capture: The edge device uses a deep learning model to extract features from the captured video content, obtaining a feature vector. This feature vector is then compared with the feature vectors of defect library photos to determine similarity. When the similarity reaches or exceeds a specific value, the edge device triggers a capture operation and saves the captured photo to the defect folder. S4. Data upload: After the inspection mission is completed, the drone will upload the video and defect folder taken during the inspection to the control computer.

[0009] The following is a further optimization and / or improvement of one of the above-mentioned technical solutions: In the above step S3, the specific value may be 85%-95%.

[0010] In the above step S3, the deep learning model may be a ResNet model or an ArcFace model.

[0011] The above-mentioned edge devices are mounted on drones and can rely on GPUs for accelerated processing.

[0012] In the above step S3, the similarity calculation may adopt a cosine similarity calculation method.

[0013] The feature vectors of the defect library photos can be pre-stored in a vector database for real-time retrieval.

[0014] In the above step S3, if the edge device determines that the similarity between a certain frame in the video and the defect library photo reaches or exceeds a specific value, in addition to triggering the snapshot operation, it will also mark the time point of the frame in the video for subsequent quick positioning and viewing.

[0015] In the above step S3, before the edge device extracts features from the video content, it can first use a video frame extraction tool to extract images from the video stream according to the frame rate.

[0016] The control computer may be installed with at least one open source framework among OpenCV, Dlib, and TensorFlowLite for further analysis and processing of the uploaded videos and captured defect photos.

[0017] The second technical solution of the present invention is achieved through the following measures: a system for real-time identification of substation defects based on drone inspections, including a drone, a control computer, a vector database and an edge device; wherein, the drone is equipped with an edge device, which is used to inspect the substation along a set route according to the task issued by the control computer, shoot videos and upload data processed by the edge device; wherein, the control computer is used to issue inspection tasks and receive uploaded data; wherein, the vector database is used to store feature vectors of defect library photos; wherein, the edge device has a built-in deep learning model and a similarity calculation algorithm, which is used to extract video feature vectors and compare them with feature vectors in the library, and trigger capture and saving when the similarity meets the standard.

[0018] This invention uses drones equipped with edge devices to achieve automated inspections and real-time data processing. The drones automatically capture substation equipment according to a set route. The edge devices' built-in deep learning models and similarity calculation algorithms analyze the video content in real time, directly capturing defect images without the need for manual screening of large numbers of photos. Compared to traditional methods that manually screen for unmatched photos, this significantly reduces defect identification time. For example, for a medium-sized substation, manual screening, which could previously take hours or even days, can be completed simultaneously during a drone inspection, significantly improving the overall efficiency of substation inspections and defect identification. Leveraging the powerful feature extraction capabilities of deep learning models, they can learn the complex and subtle features of substation equipment, significantly improving identification accuracy compared to traditional methods that rely solely on visual judgment or simple photo comparison. Furthermore, by setting a scientifically appropriate specific value for similarity calculation, the occurrence of false positives and missed detections is reduced. The edge devices process video data in real time, capturing every frame and capturing all-around equipment defects. This addresses the problem of traditional pre-set point photography, which prevents comprehensive equipment inspections. They can detect defects in hidden areas or at unusual angles, ensuring accurate identification of substation equipment defects. This reduces the manual sifting through large numbers of mismatched photos, freeing staff from lengthy, tedious photo comparisons and avoiding the fatigue and inefficiencies associated with repetitive work. Task setup and subsequent data review and analysis are performed solely on the control computer, reducing both manpower and labor intensity. Furthermore, it reduces the number of on-site inspections required due to concerns about photo content, further saving both manpower and time. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Attachment Figure 1 The figure is a flow chart of a method for real-time identification of substation defects based on drone inspections according to an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The present invention is not limited to the following embodiments, and specific implementation methods can be determined based on the technical solutions of the present invention and actual conditions.

[0021] The present invention will be further described below in conjunction with the embodiments: Example 1: As shown in the attached Figure 1 As shown, the method for real-time identification of substation defects based on drone inspection includes the following steps: S1. Task issuance: The computer is used to issue a substation inspection task to the drone equipped with the edge device, and the inspection task includes a set inspection route. S2. Inspection and video collection: The drone inspects the substation according to the set inspection route and continuously captures videos during the inspection process; S3. Real-time defect comparison and capture: The edge device uses a deep learning model to extract features from the captured video content, obtaining a feature vector. This feature vector is then compared with the feature vectors of defect library photos to determine similarity. When the similarity reaches or exceeds a specific value, the edge device triggers a capture operation and saves the captured photo to the defect folder. S4. Data upload: After the inspection mission is completed, the drone will upload the video and defect folder taken during the inspection to the control computer.

[0022] In this embodiment of the present invention, an edge device is a computing device located close to the data source or user. In this case, it is mounted on a drone and used for real-time data processing. A deep learning model is a model that learns based on data representation and can automatically learn features from large amounts of data. In this case, it is used to extract video content features. A feature vector is a digital representation of image or video content features and can be used for similarity comparison.

[0023] In this embodiment of the present invention, a control computer, acting as the command initiator, issues drones a mission encompassing patrol routes. The drone captures video during its patrols, and the edge device uses a deep learning model to analyze the video content, extracting feature vectors and comparing them against a defect database. If a defect meets certain criteria, a snapshot is captured and the data is transmitted back to the control computer. This enables automatic, real-time identification of substation defects, significantly improving work efficiency and reducing manual workload compared to traditional methods of manually screening photos. Furthermore, the entire patrol process can be recorded for easy subsequent review.

[0024] In this embodiment of the present invention, a control computer serves as the control center for the entire system. On this computer, the operator sets a detailed patrol route based on the substation's actual conditions and patrol requirements. This route information is packaged into patrol mission instructions and transmitted via wireless communication to a drone equipped with an edge device. Upon receiving the instructions, the drone confirms the mission details and prepares to execute them. The drone then conducts an autonomous patrol flight within the substation area, following the designated patrol route. During the flight, the drone's onboard camera continuously captures the substation equipment at a constant frame rate, recording its operating status as a video. The captured video data is transmitted in real time to the onboard edge device, providing raw data for subsequent defect analysis. After receiving the video data, the edge device first uses a deep learning model to extract features from each frame. Trained with a large number of defective and healthy samples, the deep learning model accurately captures device features. The extracted features are converted into feature vectors, which the edge device then compares with the feature vectors of the photos stored in the defect database. When the calculated similarity reaches or exceeds a specific value, the edge device determines that the equipment in the current video footage is defective, triggering a snapshot operation to capture the footage containing the defect and save it to a dedicated defect folder. Once the drone completes its inspection, it uploads the complete video captured during the inspection, along with the defect folder containing the defective footage, to the control computer via wired or wireless means. Once the control computer receives the data, the operator can review, analyze, and process it to promptly identify any problems with substation equipment and take appropriate measures.

[0025] Traditional substation inspections rely on manual photo screening, which is a massive workload. This drone-based real-time substation defect identification method, however, significantly reduces manual intervention and improves inspection efficiency by utilizing automated drone patrols and real-time data processing by edge devices. Drones can quickly cover various areas of the substation and efficiently complete filming tasks according to pre-set routes. Edge devices analyze video data in real time to promptly identify defects, eliminating the need for manual review of numerous photos, saving significant time and labor costs. The application of deep learning models and similarity calculation algorithms enables more accurate identification of substation equipment defects. Deep learning models can learn the complex characteristics of equipment, making them more accurate and stable than human visual judgment. By setting a reasonable similarity threshold, false positives and missed detections are reduced, improving the reliability of defect identification, helping to promptly identify potential equipment issues and ensuring safe and stable substation operation. During the inspection process, drone videos capture a complete record of the equipment's operating status. Defect photos captured by edge devices can also be combined with video for review and traceability. When personnel have questions about a defect, they can directly retrieve the corresponding video clips to observe and analyze the equipment from different angles and time points, providing richer information for accurately determining the defect nature and formulating repair plans.

[0026] The present invention issues an inspection mission to a drone equipped with an edge device through a control computer. The drone begins to inspect along the route set by the mission. The inspection process takes the form of video shooting to record the equipment in all directions. During the inspection process, the video shooting content is continuously compared with the photos in the defect library through the algorithm arranged in the edge device carried by the drone. When a device is encountered whose status is about 90% similar to the photo in the defect library, the drone automatically takes a snapshot and saves the photo of the defective device to a defect folder. After the inspection is completed, the drone uploads the inspection video and the defect folder to the control computer through a data connection for the staff to view. The present invention records the entire inspection process by recording a video. The edge device algorithm takes a snapshot when it matches the photo with a similarity of about 90% with the photo in the defect library. The number of defective photos can be greatly reduced, reducing the amount of photos that need to be manually reviewed. When there are questions about the content of the photos, the inspection video can also be retrieved to analyze the defective equipment.

[0027] Preferably, in step S3, the specific value is 85%-95%. When the edge device performs similarity calculations, 85%-95% is used as the threshold for determining whether a video image is defective. When the calculated similarity falls within this range, a snapshot operation is triggered. By defining the similarity threshold range, defect judgment has a relatively scientific and reasonable quantitative standard, improving the accuracy and consistency of defect identification and avoiding misjudgments or missed detections due to improper threshold settings.

[0028] In an embodiment of the present invention, in step S3, the deep learning model is a ResNet model or an ArcFace model. The ResNet model is a deep residual network model that solves problems such as gradient vanishing in deep neural network training by introducing a residual structure, and can effectively extract deep features of images. The ArcFace model is mainly used in the field of face recognition. It improves the accuracy of face recognition by constraining the angle of features. In this method, it can be used to extract substation equipment features and enhance the discrimination. By utilizing the powerful feature extraction capabilities of the ResNet model or the ArcFace model, the video images taken by the drone are analyzed, and the image information is converted into feature vectors to provide more representative data for subsequent similarity calculations. The accuracy and efficiency of feature extraction are improved, so that the edge device can more accurately identify the defect features of the substation equipment, further improving the reliability of defect identification.

[0029] In an embodiment of the present invention, the edge device is mounted on a drone and relies on a GPU for accelerated processing. A GPU (graphics processing unit) is a processor specifically designed for processing graphics and image data. It has powerful parallel computing capabilities and can accelerate the calculation process of deep learning models. When processing video data, running deep learning models, and performing similarity calculations, the edge device leverages the parallel computing capabilities of the GPU to quickly process large amounts of data, shorten calculation time, and achieve real-time analysis. Ensuring that the edge device can process video data in real time during drone inspections and meet the processing speed requirement of at least 10 FPS ensures the real-time nature of defect identification and avoids missed defects due to slow processing speeds.

[0030] In an embodiment of the present invention, in step S3, the similarity calculation adopts the cosine similarity calculation method. Cosine similarity measures the similarity of two vectors by calculating the cosine value of the angle between them. The value range is between [-1, 1]. The closer the value is to 1, the more similar the two vectors are. The video feature vector extracted by the edge device and the feature vector of the defect library photo are compared according to the cosine similarity calculation method to obtain a numerical value representing the similarity between the two, so as to determine whether the device in the video has defects. The cosine similarity calculation method is simple and efficient, and can quickly and accurately measure the similarity between two vectors, providing a reliable quantitative basis for defect judgment and helping to improve the accuracy of defect identification.

[0031] In an embodiment of the present invention, the feature vectors of the defect library photos are pre-stored in a vector database for real-time retrieval. The vector database is a database specifically used to store and manage vector data. It supports efficient vector retrieval operations and can quickly find vector data similar to the target vector. Before defect identification, the feature vectors of the defect library photos are first extracted and stored in the vector database. During the inspection process, after the edge device extracts the video feature vector, the vector database quickly retrieves the feature vectors of the defect library photos similar to it based on the input feature vector for the edge device to perform similarity comparison. The speed of retrieving the feature vectors of the defect library photos is improved to meet the real-time requirements, thereby improving the operating efficiency of the entire defect identification system, so that the edge device can compare the feature vectors in a timely manner to determine whether the equipment has defects.

[0032] In an embodiment of the present invention, in step S3, if the edge device determines that the similarity between a certain frame in the video and a photo in the defect library reaches or exceeds a specific value, in addition to triggering the snapshot operation, the time point of the frame in the video will be marked for subsequent quick location and viewing. When the edge device detects a frame that meets the defect judgment criteria, while executing the snapshot operation, it records the timestamp information of the frame in the video and associates the time point with the snapshot photo for storage. This allows staff to quickly locate the specific time location of the defect when viewing the inspection video and defect photos, analyze and process the defects more efficiently, save review time, and improve work efficiency.

[0033] In an embodiment of the present invention, in step S3, before the edge device extracts features from the video content, it first uses a video frame extraction tool to extract images from the video stream at a frame rate. The video frame extraction tool is a tool used to extract a single image from a continuous video stream at a set frame rate, such as FFmpeg, OpenCV, etc., which is used in the present invention to obtain image data required for the deep learning model. The video frame extraction tool processes the video stream shot by the drone at a certain frame rate, converts the continuous video into a series of single images, and provides basic data for the subsequent feature extraction of the deep learning model. The standardization and consistency of the data processed by the deep learning model are guaranteed, so that the model can accurately extract features from the video, improve the quality of feature extraction, and thus improve the accuracy of defect identification.

[0034] In an embodiment of the present invention, the control computer is installed with at least one open source framework: OpenCV, Dlib, or TensorFlow Lite, for further analysis and processing of uploaded videos and captured defect photos. OpenCV is an open source library for computer vision tasks, providing a variety of image processing and computer vision algorithms for operations such as image filtering and feature detection. Dlib is a modern C++ machine learning toolkit that includes a variety of machine learning algorithms and computer vision functions, such as face detection and feature point location. TensorFlow Lite is a lightweight version of TensorFlow, specifically designed for running machine learning models on mobile and embedded devices, and can be used for model inference and simple model training on the control computer. The control computer uses these installed open source frameworks to perform secondary processing on the videos and defect photos uploaded by the drone, such as image enhancement using OpenCV, more accurate feature recognition using Dlib, and more complex model analysis using TensorFlow Lite. This further enriches the analysis methods for videos and photos, improves the accuracy and processing power of defect analysis, helps discover defect information that may be missed by initial identification by edge devices, and enhances the defect recognition capabilities of the entire system.

[0035] It should be noted that in the present invention, the video resolution needs to reach 1080P, a processing speed of at least 10FPS is required, it relies on GPU (graphics card) acceleration (such as edge device NVIDIA Jetson), and requires a high-definition image library. In the present invention, video frame extraction uses tools such as FFmpeg and OpenCV to extract images from the video stream at frame rate; deep learning models (such as ResNet, ArcFace) are used to extract feature vectors; when performing vectorized comparison, the feature vectors are calculated with the photo vectors in the image library for similarity (such as cosine similarity). When performing real-time retrieval, vector databases (such as Milvus and Faiss) are used to accelerate the retrieval of massive data; when triggering a snapshot, when the similarity exceeds the threshold, the current frame is saved or an alarm is sent. In the present invention, when selecting key algorithms and tools, YOLO and EfficientDet are used for object recognition; OpenCV, Dlib, and TensorFlowLite (suitable for edge device deployment) are used as open source frameworks. In the present invention, existing equipment or software on the market can be used, such as the cloud service API using Google Cloud Video Intelligence, which supports object scene recognition and can be integrated with BigQuery; Dahua's SmartPSS platform, which supports image search and millisecond-level response; Uniview's "Kunlun" video structured server, which supports concurrent analysis of 10,000 videos; and SenseTime's SenseFoundry platform, which supports cross-camera tracking and has a comparison accuracy rate of over 99%.

[0036] Example 2: This embodiment provides a system for real-time substation defect identification based on drone inspections, including a drone, a control computer, a vector database, and an edge device. The drone, equipped with an edge device, is configured to patrol the substation along a set route according to tasks assigned by the control computer, capture video, and upload data processed by the edge device. The control computer is configured to issue inspection tasks and receive uploaded data. The vector database is configured to store feature vectors of defect photos. The edge device has a built-in deep learning model and similarity calculation algorithm to extract video feature vectors and compare them with feature vectors in the database. When the similarity meets a certain threshold, a snapshot is triggered and saved. The system, with the control computer as the control center, issues inspection tasks to drones equipped with edge devices. The drone patrols and captures video according to the task requirements. The edge device processes the video data in real time, and the vector database provides data support. All components work together to identify defects, and the data is then transmitted back to the control computer. This creates a complete and efficient substation defect identification system with clear division of labor and collaborative operation. This fully automates the entire process from task issuance, data collection and processing, to result feedback, significantly improving the efficiency and accuracy of substation defect identification and reducing labor costs.

[0037] As the front-end data collection device, the drone, equipped with edge devices, flies over the substation according to mission instructions issued by the control computer. During flight, the drone's camera captures video of the substation equipment and transmits the video data to the edge device in real time. The edge device's built-in deep learning model and similarity calculation algorithm immediately processes the video data, enabling real-time detection of equipment defects. Once a defect image that meets the similarity requirement is detected, the edge device triggers a capture and saves the image. The processed data, including the captured image and related analysis information, is uploaded to the drone. The control computer plays a core control and management role in the system. During the mission issuance phase, the operator plans the drone's patrol route on the control computer, generates the patrol mission, and sends it to the drone. During the patrol, the control computer monitors the drone's flight status and mission progress in real time. After the drone completes its patrol mission, the control computer receives the uploaded video and defect data. The operator can then review, analyze, and organize this data in detail to further assess the operating status of the substation equipment. The vector database pre-stores a large number of feature vectors of defect database photos. When the edge device needs to perform a similarity calculation, it quickly retrieves the defect library photo feature vector that matches the current video feature vector from the vector database. The vector database, with its efficient retrieval algorithm, can quickly return relevant data, providing strong data support for real-time processing on the edge device and ensuring rapid and accurate defect comparison and identification.

[0038] This system seamlessly integrates drones, control computers, a vector database, and edge devices. Each component performs its own function, collaborating to complete substation defect identification tasks. From task assignment to data collection and processing to final result analysis, a complete and efficient workflow is formed, significantly improving the system's overall performance and reliability. Operators can remotely control drone inspections from a control computer, eliminating the need to visit the substation. They can also access real-time drone footage and defect data processed by edge devices, enabling real-time monitoring of substation equipment operating status. This remote monitoring and management approach not only improves work efficiency but also reduces safety risks faced by workers on-site. The system's components are relatively independent yet collaborative, offering excellent scalability. For example, higher-performance drones can be replaced, edge device hardware and algorithms can be upgraded, or the vector database's storage and retrieval strategies can be optimized based on actual needs. This flexibility enables the system to be continuously optimized and upgraded as technology advances and application requirements change, continuously enhancing its substation defect identification capabilities and capabilities.

[0039] The above technical features constitute the embodiments of the present invention, which have strong adaptability and implementation effect. Non-essential technical features can be added or removed according to actual needs to meet the requirements of different situations.

Claims

1. A method for real-time identification of substation defects based on drone inspection, characterized in that The following steps are involved: S1. Task issuance: The computer is used to issue a substation inspection task to the drone equipped with the edge device, and the inspection task includes a set inspection route. S2. Inspection and video collection: The drone inspects the substation according to the set inspection route and continuously captures videos during the inspection process; S3. Real-time defect comparison and capture: The edge device uses a deep learning model to extract features from the captured video content, obtaining a feature vector. This feature vector is then compared with the feature vectors of defect library photos to determine similarity. When the similarity reaches or exceeds a specific value, the edge device triggers a capture operation and saves the captured photo to the defect folder. S4. Data upload: After the inspection mission is completed, the drone uploads the video and defect folder taken during the inspection to the control computer.

2. The method for real-time identification of substation defects based on drone inspection according to claim 1 is characterized in that In step S3, the specific value is 85%-95%.

3. The method for real-time identification of substation defects based on drone inspection according to claim 1 is characterized in that In step S3, the deep learning model is a ResNet model or an ArcFace model.

4. The method for real-time identification of substation defects based on drone inspection according to claim 1, 2 or 3 is characterized in that The edge device is mounted on a drone and relies on a GPU for accelerated processing.

5. The method for real-time identification of substation defects based on drone inspection according to claim 1, 2 or 3 is characterized in that In step S3, the similarity calculation adopts a cosine similarity calculation method.

6. The method for real-time identification of substation defects based on drone inspection according to claim 1, 2 or 3 is characterized in that The feature vectors of the defect library photos are pre-stored in a vector database for real-time retrieval.

7. The method for real-time identification of substation defects based on drone inspection according to claim 1, 2 or 3 is characterized in that In step S3, if the edge device determines that the similarity between a certain frame in the video and a photo in the defect library reaches or exceeds a specific value, in addition to triggering the snapshot operation, it will also mark the time point of the frame in the video for subsequent quick positioning and viewing.

8. The method for real-time identification of substation defects based on drone inspection according to claim 1, 2 or 3 is characterized in that In step S3, before the edge device extracts features from the video content, it first uses a video frame extraction tool to extract images from the video stream at a frame rate.

9. The method for real-time identification of substation defects based on drone inspection according to claim 1, 2 or 3 is characterized in that The control computer is installed with at least one open source framework among OpenCV, Dlib, and TensorFlowLite, which is used to further analyze and process the uploaded video and captured defect photos.

10. A system for real-time identification of substation defects based on drone inspections, characterized in that It includes a drone, a control computer, a vector database and an edge device; the drone is equipped with an edge device, which is used to patrol the substation along the set route according to the task issued by the control computer, shoot videos and upload data processed by the edge device; the control computer is used to issue patrol tasks and receive uploaded data; the vector database is used to store feature vectors of defect library photos; the edge device has a built-in deep learning model and similarity calculation algorithm, which is used to extract video feature vectors and compare them with feature vectors in the library, and trigger capture and saving when the similarity meets the standard.

Images