Marked augmented reality-based intelligent inspection method for power distribution station

Abstract

The invention discloses a marker-based augmented reality intelligent inspection method for a power distribution station, and the method comprises the steps: building a 3D map of a power distribution station house through three-dimensional scanning or BIM modeling, building a floating point map through combining with an AR identity marker attached to equipment, and achieving the one-to-one binding of virtual information and physical positions of the equipment; during routing inspection, the AR equipment provides real-time navigation through double-map fusion positioning, supports multi-mode data entry of keys, voice and the like, and completes high-speed real-time processing of'acquisition-detection-preprocessing-recognition 'of video images through a built-in concurrent modular image recognition system. According to the system, MobileNetV3 is used as a backbone network, visual features, sensor data and equipment logs are fused to realize multi-modal identification, and AR mobile terminal resource limitation is adapted through buffer queue intelligent scheduling and computing resource optimization. According to the invention, the problems of low efficiency, inaccurate data, high safety risk and high training cost of traditional inspection are solved, standardization of the inspection process, automation of data acquisition and real-time state feedback are realized, the inspection efficiency and the intelligent level of operation and maintenance are greatly improved, the comprehensive cost is reduced, and the method is suitable for daily inspection and state management of key equipment of various power distribution stations.

Description

Technical Field

[0001] This invention relates to a marker-based augmented reality (AR) intelligent inspection method for power distribution substations, and more particularly to a power distribution substation inspection technology that combines images of power distribution equipment as specific markers, augmented reality technology, and data analysis. This technology belongs to the application of computer vision technology in the intelligent operation and maintenance of power equipment. Background Technology

[0002] As a core component of the power supply chain at the end of the power system, the stable operation of the internal equipment (such as high and low voltage switchgear, transformers, distribution boxes, and circuit breakers) of the substation directly determines the power supply reliability for end users. Currently, substation inspections are still mainly conducted manually, requiring inspectors to carry paper work orders and manually check equipment status and record inspection data by sight, hearing, smell, and touch. This model has many obvious drawbacks: 1. Low inspection efficiency: Manual recording is time-consuming and labor-intensive, and the inspection progress is heavily dependent on the experience and mental state of the personnel. Inexperienced personnel need to repeatedly check paper work orders, and mental fatigue can slow down the inspection speed, resulting in the inspection time of a single substation usually lasting 1 to 2 hours, which is difficult to meet the high-efficiency operation and maintenance needs of large-scale substations. 2. Poor data accuracy: Paper records are prone to omissions (such as missing a reading on a dashboard) and misreadings (such as misreading the status of an indicator light). When the paper data is subsequently entered into the digital system, secondary errors may occur due to manual transcription, resulting in low data reliability. 3. High personnel safety risks: Inspection personnel need to stare closely at equipment nameplates and complex instruments to obtain data, which leads to excessive time spent in high-voltage equipment areas. If the equipment has potential risks such as leakage or arcing, it can easily cause safety accidents. 4. Severe information silos: Inspection data cannot be uploaded to the operation and maintenance management system in real time. It needs to be manually summarized and entered into the system, resulting in strong data lag. This makes it impossible to provide real-time data support for equipment failure prediction and preventive maintenance, which is not conducive to the intelligent management of power distribution stations. 5. High training costs: The equipment layout of substations is complex and the models are diverse. New inspection personnel need to undergo 3 to 6 months of on-site training to basically master the equipment location, inspection process and parameter standards. In the later work, they also need to conduct technical training and safety education frequently, which results in a long training cycle and high costs.

[0003] In recent years, technologies such as robotic inspection and fixed camera monitoring have emerged in the industry. While robotic inspection can replace manual labor in high-risk areas, it suffers from insufficient flexibility (e.g., inability to pass through narrow spaces) and high cost (a single inspection robot costs over 100,000 yuan). Fixed camera monitoring only enables remote video viewing and cannot provide active inspection functions such as equipment parameter reading and process guidance, still requiring manual remote assessment of equipment status, thus failing to fundamentally solve the problems of inspection efficiency and accuracy. Therefore, there is an urgent need for an inspection solution that combines the flexibility of on-site personnel with the advantages of intelligent technology, providing on-site inspection personnel with intuitive information guidance and automated data processing, effectively compensating for the shortcomings of existing technologies, and improving the quality and efficiency of substation inspections. Summary of the Invention

[0004] Purpose of the invention: To overcome the shortcomings of existing technologies, this invention provides a marker-based augmented reality (AR) intelligent inspection method for substations. By recognizing AR markers, overlaying virtual information, and interacting with data in real time, it seamlessly integrates virtual equipment information with real-world scenarios, enabling intelligent guidance of the inspection process, automatic data collection and analysis, and real-time alarms for anomalies. Ultimately, it achieves standardization, efficiency, and safety in the inspection process, improving inspection efficiency, ensuring personnel safety, and reducing operation and maintenance costs. It is applicable to the daily inspection and status management of key power equipment such as high and low voltage switchgear, transformers, and distribution boxes in substations.

[0005] Technical solution: To achieve the above objectives, the technical solution adopted by this invention is as follows: An intelligent inspection method for power distribution substations based on tag-based augmented reality (AR) involves: 3D modeling of the internal environment of the power distribution substation; attaching AR identity tags to designated devices to establish a one-to-one correspondence between AR identity tags and designated devices; constructing a 3D map of the internal environment of the power distribution substation and a floating-point map covering all designated devices, where each designated device is marked as a floating point in spatial location on the floating-point map; before inspection, selecting a list of devices to be inspected from the designated devices according to the inspection task requirements and specifying the inspection order of the devices to be inspected; during inspection, first identifying the current environment using AR devices and determining the current location by referring to the 3D map; simultaneously, identifying the first uninspected device in the current inspection task using the floating-point map, providing a navigation path from the current location to the identified uninspected device, recording inspection data using AR devices and marking the inspection as completed, and completing the inspection of all devices to be inspected one by one.

[0006] The core of this project is to achieve intelligent substation inspection through "AR tag binding + dual map navigation + multi-module concurrent recognition + real-time data interaction". Compared with the problems of low efficiency, poor data accuracy, high safety risks, information lag, and high training costs of existing technologies, as well as the shortcomings of insufficient flexibility of robot inspection and single function of fixed monitoring, this project solves the core pain points of the traditional inspection mode by using the technical means of 3D modeling and floating point map fusion positioning, AR tag precise association with equipment, concurrent image recognition system high-speed processing, and operation and maintenance server real-time data analysis. It achieves standardized inspection process, automated data collection, and real-time status feedback.

[0007] Specifically, 3D scanning or BIM modeling is used to create a 3D model of the internal environment of the power distribution station, constructing a 3D map. 3D scanning can quickly obtain the spatial dimensions and equipment layout of the real environment, while BIM modeling can accurately reproduce the equipment structure and installation relationships. Both can generate high-fidelity, interactive virtual maps, adaptable to scenarios with dense equipment and complex layouts in power distribution stations. This provides accurate spatial benchmarks for inspection navigation and environmental identification, avoiding inspection omissions or inefficient routes caused by environmental perception biases.

[0008] Specifically, the floating-point map binds the spatial location, historical data, real-time data, threshold data, standard inspection procedures, and AR identity tags of designated devices. During inspection, the spatial locations of the devices to be inspected earlier and later are simultaneously bound. All floating-point information is stored on the operations and maintenance server. This design achieves "one-stop integration" of device information. Inspection personnel can simultaneously obtain full-dimensional data of devices through AR devices, eliminating the need to repeatedly consult paper documents or switch systems. This solves the problems of scattered information and cumbersome querying in traditional inspections. At the same time, the binding of the spatial locations of preceding and following devices provides a foundation for continuous navigation, ensuring a coherent and efficient inspection process and preventing disorientation or missed inspection nodes in complex substations.

[0009] Specifically, both the 3D map and the floating-point map are stored on the operations and maintenance server, and the AR device connects to the server wirelessly. On the operations and maintenance server, by comparing real-time data and threshold data collected during inspections, the current status of the equipment is determined. The server analyzes and provides feedback on possible causes of abnormal states, operational suggestions, and related equipment. If a related equipment is not on the list of equipment to be inspected, it is added to the list, and its inspection priority is increased. If a related equipment is on the list but has not been inspected, its inspection priority is increased. This design achieves an "intelligent closed loop" for inspections. Real-time data comparison can quickly locate equipment anomalies, avoiding the subjectivity and lag of manual judgment. Anomaly cause analysis and operational suggestions directly guide on-site handling, reducing reliance on the experience of inspection personnel. Adjusting the priority of related equipment can promptly identify potential cascading failures, solving the shortcomings of traditional inspections that focus on "single-point inspection and ignoring related risks," thus improving the comprehensiveness and preventative nature of inspections.

[0010] Specifically, the AR device visually identifies AR identity markers, which can be device photos, QR codes, or pre-designed visual reference images. This design boasts high adaptability and reliability. Device photos and visual reference images can be adapted to older equipment without QR code labels, while QR codes support fast and accurate recognition, meeting the labeling needs of power distribution equipment of different ages and types. The visual recognition method eliminates the need for contact with the equipment, avoiding the safety risks of close-range operation of high-voltage equipment. Furthermore, it is unaffected by electromagnetic interference within the power distribution station, ensuring the stability of the label recognition and solving the problems of cumbersome and error-prone equipment identification verification in traditional inspections.

[0011] Specifically, the AR device records inspection data through methods such as key input, handwriting input, image input, or voice input. For content requiring recognition, the AR device performs the recognition and confirmation, or simultaneously saves the original material. This design provides recording methods adaptable to multiple scenarios. Voice and image input achieve "what you see is what you get," significantly shortening data recording time and solving the problem of low efficiency in manual handwriting recording. Multiple input methods adapt to different inspection scenarios (e.g., handwriting can be selected when voice input is not possible in noisy environments), improving operational flexibility. Saving the original material provides evidence for data traceability, avoiding subsequent disputes, while device-side recognition and confirmation reduce errors from manual transcription and improve data accuracy.

[0012] Specifically, after all the equipment to be inspected in the inspection task has been inspected, the operation and maintenance server performs background analysis on all inspection data to generate an inspection report. This design realizes the automated collection and analysis of inspection data, avoiding the time-consuming and labor-intensive process of manually compiling reports and significantly shortening the inspection closed-loop cycle. The report can integrate information such as equipment status, anomaly details, and handling suggestions, providing data support for operation and maintenance management. It solves the problems of delayed report generation and low data utilization in traditional inspections, and helps to make refined and intelligent decisions in the operation and maintenance of power distribution stations.

[0013] Specifically, the AR device has a built-in image recognition system that uses a deep learning network and adopts a concurrent modular system architecture. It includes a video image acquisition module, an input-level object detection module, an intermediate-level preprocessing module, and an output-level object recognition module and video image recognition module. The video image acquisition module directly sends the acquired video images to the object detection module. The object detection module performs motion estimation on the video images, sending dynamically identified video images to an optical flow analysis thread to generate dynamic bounding boxes, and sending statically identified video images to a keyframe extraction thread to generate static images. The identified dynamic bounding boxes or static images are stored in a waiting-to-process buffer queue. The preprocessing module retrieves an image from the head of the waiting-to-process buffer queue and performs image enhancement operations, storing the enhanced image in a waiting-to-recognize buffer queue. The object recognition module retrieves an image from the head of the recognition buffer queue and performs target recognition operations, directly sending the target recognition result to the video image recognition module. The video image recognition module outputs the target recognition result. The proposed concurrent modular system architecture incorporates a concurrent multi-threaded working mode on top of a modular design. It abandons the traditional communication method relying on reserved interfaces between modules, instead employing a message buffer queue combined with a "sample class" communication method. This transforms the traditional serial structure into a parallel system structure, accelerating the efficiency of the video image recognition system and achieving a truly high-speed, real-time video image recognition system. This design resolves the contradiction between the limited resources of AR mobile devices and the real-time recognition requirements. The concurrent multi-threaded architecture allows the acquisition, detection, preprocessing, and recognition modules to work in parallel, significantly improving processing speed. The message buffer queue combined with the "sample class" communication method reduces coupling between modules, enhancing system stability. Dynamic / static frame classification processing avoids redundant calculations, ensuring rapid focus on core detection targets even in complex substation scenarios (such as personnel movement or slight equipment vibration), achieving second-level recognition and feedback of equipment status.

[0014] Specifically, the preprocessing module adopts an enhanced pipeline architecture, sequentially performing geometric transformation, color transformation, noise and blur processing, and advanced transformation operations on the input image. Simultaneously, it corrects radial, tangential, and perspective distortions through geometric correction, and optimizes resources and improves performance through computational resource scheduling. Finally, the standardized enhanced image is pushed to the recognition buffer queue. This design addresses scenarios with complex lighting conditions in power distribution stations (such as shadows and strong light reflections) and varying shooting angles (such as close-up and oblique shots). It improves image quality through image enhancement and corrects shooting distortions through geometric correction, ensuring the accuracy of subsequent recognition. Computational resource scheduling dynamically allocates computing power based on the AR device's load, avoiding device lag caused by processing complex images, balancing recognition accuracy and battery life, and solving the problem of "difficulty in balancing quality and efficiency" in mobile image processing.

[0015] Specifically, the object recognition module uses MobileNetV3 as its backbone network. A feature extraction accelerator adaptively selects the resolution based on the target size and performs multi-level feature fusion, fusing visual features, sensor data, and device logs to achieve multimodal recognition. Real-time processing optimizes target features and status determination. In this design, the lightweight MobileNetV3 network adapts to the limited computing resources of AR mobile devices, reducing energy consumption while ensuring recognition accuracy. The feature extraction accelerator dynamically adjusts the resolution based on the target size, enabling accurate recognition of small targets (such as indicator lights) and large targets (such as transformers), adapting to scenarios with significant differences in the size of equipment in power distribution stations. Multimodal data fusion overcomes the limitations of single-vision recognition, combining sensor data (such as voltage and current) and device logs (such as maintenance records) to achieve more comprehensive status determination, reducing misjudgments caused by insufficient visual information and solving the problems of "single information and weak anti-interference ability" in traditional image recognition.

[0016] Beneficial Effects: The label-based augmented reality intelligent inspection method for substations provided by this invention has the following advantages compared to existing technologies: 1. Significantly Improved Efficiency: 3D map navigation and standardized process guidance, combined with multiple methods for rapid data entry, reduce the inspection time of a single substation by more than 50%; the concurrent image recognition system achieves second-level response, avoiding process bottlenecks; 2. Accurate and Reliable Data: MobileNetV3 network and multimodal fusion technology ensure that the equipment recognition accuracy is ≥95% and the error rate is ≤0.5%, eliminating human misreading and omissions; data is uploaded in real time, without lag or secondary errors; 3. Reduced Safety Risks: Non-contact recognition reduces the time spent in high-voltage areas, and real-time alarms and standardized visual guidance avoid risks such as electric shock and operational violations; 4. Intelligent Upgrade of Operation and Maintenance: Anomaly linkage and tracing enable preventive maintenance, reducing the number of failures by more than 30%; data-driven decision-making and remote collaboration improve the speed of operation and maintenance response; 5. Effective Cost Control: The training cycle for new personnel is shortened by more than 60%, the hardware cost of AR equipment is 70% lower than that of inspection robots, resource scheduling optimization extends equipment battery life, and the overall cost is significantly reduced. Attached Figure Description

[0017] Figure 1 An architecture diagram of the image recognition system built into an AR device; Figure 2 A schematic diagram illustrating the resource optimization process for computing resource scheduling; Figure 3 This is a schematic diagram of the intelligent scheduling process for the buffer queue; Figure 4 This is a schematic diagram of the marking process of a new frame before it enters the buffer queue. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0019] An intelligent inspection method for power distribution substations based on marker-based augmented reality (AR) involves creating a 3D model of the substation's internal environment using 3D scanning or BIM modeling, and attaching AR identification tags (equipment photos, QR codes, or pre-designed visual reference maps, etc.) to designated devices, establishing a one-to-one correspondence between AR identification tags and designated devices. A 3D map of the substation's internal environment and a floating-point map covering all designated devices are constructed, with each designated device marked as a floating point in spatial location on the floating-point map. Before inspection, a list of devices to be inspected is selected from the designated devices according to the inspection task requirements, and the inspection order is specified. During inspection, the current environment is first identified using AR devices, and the current location is determined by referring to the 3D map. Simultaneously, the system uses a floating-point map to identify the first uninspected device in the current inspection task, provides a navigation path from the current location to the identified uninspected device, records inspection data using an AR device, and marks the inspection as complete, completing the inspection of all devices one by one. After all devices in the inspection task have been inspected, the system uses the maintenance server to perform background analysis on all inspection data and generate an inspection report. Inspection data is recorded using an AR device, with recording methods including key input, handwriting input, image input, or voice input. For content that needs to be recognized, recognition and confirmation are performed on the AR device, or the original material is saved simultaneously.

[0020] In the floating-point map, the spatial location, historical data, real-time data, threshold data, standard inspection procedures, and AR identity tags of specified devices are bound together. During inspection, the spatial locations of the devices to be inspected earlier and later are bound simultaneously. All information from the floating-point map is stored on the operations and maintenance server. Both the 3D map and the floating-point map are stored on the operations and maintenance server, and the AR devices connect to the operations and maintenance server wirelessly. On the operations and maintenance server, the current status of the devices is determined by comparing the real-time data and threshold data obtained during inspection. The server analyzes and provides feedback on possible causes of abnormal states, operational suggestions, and related devices. If a related device is not in the list of devices to be inspected, it is added to the list, and its inspection priority is increased. If a related device is in the list of devices to be inspected but has not been inspected, its inspection priority is increased.

[0021] The AR device has a built-in image recognition system that uses a deep learning network and adopts a concurrent modular system architecture, such as... Figure 1As shown, the system includes a video image acquisition module, an input-level object detection module, an intermediate-level preprocessing module, and an output-level object recognition module and a video image recognition module. The video image acquisition module directly sends the acquired video images to the object detection module. The object detection module performs motion estimation on the video images, sends video images identified as dynamic to an optical flow analysis thread to generate dynamic bounding boxes, and sends video images identified as static to a keyframe extraction thread to generate static images. The identified dynamic bounding boxes or static images are stored in a waiting-to-process buffer queue. The preprocessing module retrieves an image from the head of the waiting-to-process buffer queue and performs image enhancement operations, storing the enhanced image in a waiting-to-recognize buffer queue. The object recognition module retrieves an image from the head of the recognition buffer queue and performs target recognition operations, directly sending the target recognition result to the video image recognition module. The video image recognition module outputs the target recognition result.

[0022] In the optical flow analysis thread, the motion information of objects between adjacent frames is calculated by utilizing the temporal changes of pixels in the image sequence and the correlation between adjacent frames, based on the correspondence between the previous and current frames. In the keyframe extraction thread, a keyframe is a frame that describes the main content of the video. After obtaining the keyframes, content-based still image retrieval techniques can be used to search for them, thus transforming the video retrieval problem into an image retrieval problem. Therefore, keyframes should be representative; for video information, they should represent thematic features, while for image information, the specific characteristics depend on the extracted features.

[0023] Convolutional neural networks (CNNs) have been widely used in visual tasks such as image classification and object detection, achieving great success. However, CNNs typically require significant computation and memory, limiting their application in resource-constrained environments such as mobile devices and embedded systems, thus necessitating network compression. MobileNetV3 is a highly efficient and lightweight CNN optimized for mobile devices. The object detection module in this case also uses MobileNetV3 as its backbone network. After motion estimation of the input image by the precision decision unit, the INT8 mode is used for high-confidence dynamic images to identify dynamic bounding boxes through fast inference paths, while the FP16 mode is used for keyframes to accurately identify static images.

[0024] In target recognition networks, the inputs and outputs of different modules are typically connected via interfaces. This case introduces caching technology (reducing data access and computation time), widely used in application system development, to restructure the business processing flow. Taking into account the slightly slower system processing speed and the phased nature of system input data in AR device scenarios, this approach avoids performance bottlenecks during system processing, thereby improving performance. In this case, both the pending processing buffer queue and the pending recognition buffer queue utilize caching technology, and a multi-level caching design of "raw frame cache → preprocessing queue → detection pool → inference engine → result cache → postprocessing queue" is adopted to further optimize data flow efficiency.

[0025] The preprocessing module adopts an enhanced pipeline architecture, applying a series of image enhancement operations sequentially to the input image. These image enhancement operations include geometric transformations (horizontal flip, vertical flip, rotation, random cropping, resizing), color transformations (randomly adjusting brightness and contrast, adjusting hue, saturation and value, randomly shifting RGB channels), noise and blurring (adding Gaussian noise, Gaussian blur), and advanced transformations (randomly occluding image regions, elastic transformation, mesh distortion), etc.

[0026] To ensure image quality and accuracy, the preprocessing module simultaneously performs ensemble correction of geometric distortions in the image using mathematical models and algorithms to improve the accuracy of the vision system. Geometric distortions typically include radial, tangential, and perspective distortions, which affect the geometric features of the image, causing positional deviations or shape distortions of objects. The selection of algorithms and software tools is crucial during geometric correction. Commonly used geometric correction algorithms include least squares methods, nonlinear optimization algorithms, and adaptive filtering algorithms. These algorithms minimize distortions in the image by adjusting correction parameters, thus achieving accurate geometric correction. Before correction, a pinhole camera model and distortion model must be established, and correction parameters are obtained through methods such as checkerboard calibration to ensure the correction effect.

[0027] Considering the limited computing resources in AR device scenarios, this case addresses resource optimization and performance improvement through computing resource scheduling. The main objectives of computing resource scheduling include: ① Resource optimization: maximizing resource utilization through reasonable configuration and scheduling; ② Performance improvement: ensuring good system performance under high load; ③ Cost control: reducing resource usage costs and improving economic efficiency; ④ Scalability: supporting system expansion to meet ever-increasing computing demands. Figure 2As shown, the system optimizes the computing resource scheduling process by monitoring CPU load, CPU utilization, and temperature status in real time through a resource monitor. When a computing task occurs, the dynamic allocator selects a strategy based on the monitoring data from the resource monitor: if the computing task is a low-load task, CPU multi-threading is given priority; if the computing task is a high-load task, CPU acceleration is given priority; if the computing task is an urgent task, NPU-dedicated resources are given priority.

[0028] The object recognition module uses MobileNetV3 as its backbone network. Through a feature extraction accelerator, it adaptively selects the resolution based on the target size (4x downsampling for small targets, 2x downsampling for regular targets, and original size for large targets) and performs multi-level feature fusion. This fuses visual features, sensor data (voltage Ua / Ub / Uc, current Ia / Ib / Ic, etc.), and equipment logs (commissioning time, maintenance records, etc.) to achieve multimodal recognition. Real-time processing optimizes target features and status determination. The feature extraction process sequentially completes data preprocessing (cleaning, denoising, normalization), feature selection, feature extraction, and feature dimensionality reduction steps to ensure the effectiveness of the feature vectors.

[0029] The image recognition system used in this case adopts a concurrent modular system architecture. It is based on three concurrently executing work modules: object detection, preprocessing, and object recognition. They communicate with each other through a shared memory buffer queue. The object detection module reads the image resources acquired by the video image acquisition module, extracts object samples, and pushes them to the buffer queue to guide the overall system operation. The preprocessing module is awakened by the buffer queue to be processed and pushes the standardized samples to the buffer queue to be recognized. The object recognition module, as the output level of the system, is awakened by the buffer queue to be recognized and exports the results in the window after completing the recognition work.

[0030] Considering the primary purpose of intelligent inspection within the substation—namely, safety and timely handling of safety issues—intelligent scheduling of the two buffer queues can be implemented to address urgent problems promptly. For example... Figure 3 As shown, this is the intelligent scheduling process of the buffer queue. The resource monitor monitors the buffer queue depth, CPU / GPU load and memory pressure in real time. When a new frame arrives in the buffer queue, the system dynamically adjusts the system based on the monitoring data from the resource monitor and reference Table 1.

[0031] Table 1 Dynamic Adjustment Parameters Adjustment Dimensions Adjustment range Impact Factor Queue capacity ±30% Memory utilization Batch size 1~16 CPU utilization Frame resolution 360p~4k Network bandwidth Processing frame rate 15~60fps CPU temperature like Figure 4As shown, before a new frame enters the buffer queue, it needs to be prioritized by the queue manager (e.g., abnormal frames are marked with high priority) so that urgent images can be processed first. In addition, the queue manager can also store new frames in a circular buffer or a dynamically expanded area according to the memory allocation strategy to avoid memory leaks or overflows.

[0032] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the present invention in any way, and all technical solutions obtained by equivalent substitution or equivalent transformation fall within the protection scope of the present invention.

Claims

1. A tag-based augmented reality intelligent inspection method for substations, characterized in that: 3D modeling of the internal environment of the power distribution room is performed, and AR identity tags are attached to designated devices to establish a one-to-one correspondence between AR identity tags and designated devices; a 3D map of the internal environment of the power distribution room and a floating-point map covering all designated devices are constructed, and each designated device is marked as a floating point in spatial location in the floating-point map; Before the inspection, select the list of equipment to be inspected from the designated equipment according to the inspection task requirements, and specify the inspection order of the equipment to be inspected. During the inspection, the current environment is first identified using AR devices, and the current location is determined by referring to the 3D map. At the same time, the first uninspected device in the current inspection task is identified using a floating map, and a navigation path from the current location to the identified uninspected device is given. Inspection data is recorded using AR devices and the inspection is marked as completed. The inspection of all devices to be inspected is completed one by one.

2. The label-based augmented reality intelligent inspection method for substations according to claim 1, characterized in that: A 3D model of the internal environment of the power distribution station is created using 3D scanning or BIM modeling to construct a 3D map.

3. The label-based augmented reality intelligent inspection method for substations according to claim 1, characterized in that: In the floating-point map, the spatial location, historical data, real-time data, threshold data, standard inspection process, and AR identity tag of the specified device are bound together; During the inspection, the spatial location of the device to be inspected earlier and the spatial location of the device to be inspected later are bound at the same time; all floating-point information is stored in the operation and maintenance server.

4. The label-based augmented reality intelligent inspection method for substations according to claim 1, characterized in that: Both 3D maps and floating-point maps are stored on the operations and maintenance server. AR devices connect to the operations and maintenance server wirelessly. On the operations and maintenance server, the current status of the device is determined by comparing the real-time data and threshold data detected during inspection. The server analyzes and provides feedback on possible causes of abnormal status, operation suggestions, and related devices. If the related device is not in the list of devices to be inspected, it is added to the list and its inspection priority is increased. If the associated device is on the list of devices to be inspected but has not been inspected, its inspection priority will be increased.

5. The label-based augmented reality intelligent inspection method for substations according to claim 1, characterized in that: The AR device identifies AR identity markers visually, and the AR identity markers are device photos, QR codes, or pre-designed visual reference images.

6. The label-based augmented reality intelligent inspection method for substations according to claim 1, characterized in that: The inspection data is recorded using AR devices, and the recording methods include key input, handwriting input, image input, or voice input. For content that needs to be recognized, the AR device is used for recognition and confirmation, or the original material is saved at the same time.

7. The label-based augmented reality intelligent inspection method for substations according to claim 1, characterized in that: Once all the equipment to be inspected in the inspection task has been inspected, the operation and maintenance server performs background analysis on all inspection data and generates an inspection report.

8. The label-based augmented reality intelligent inspection method for substations according to claim 1, characterized in that: The AR device has a built-in image recognition system that uses a deep learning network and a concurrent modular system architecture. It includes a video image acquisition module, an input-level object detection module, an intermediate-level preprocessing module, and an output-level object recognition module and video image recognition module. The video image acquisition module directly sends the acquired video images to the object detection module. The object detection module performs motion estimation on the video images, sending dynamically identified video images to an optical flow analysis thread to generate dynamic bounding boxes, and sending statically identified video images to a keyframe extraction thread to generate static images. The identified dynamic bounding boxes or static images are stored in a waiting-to-process buffer queue. The preprocessing module retrieves data from the waiting-to-process buffer queue... The image is retrieved from the head of the queue and an image enhancement operation is performed. The enhanced image is then stored in the recognition buffer queue. The object recognition module retrieves the image from the head of the recognition buffer queue and performs a target recognition operation. The target recognition result is then directly sent to the video image recognition module. The video image recognition module outputs the target recognition result. The concurrent modular system architecture adds a concurrent multi-threaded working mode to the modular design. It abandons the traditional communication method through reserved interfaces between modules and adopts a message buffer queue combined with a "sample class" communication method. This changes the traditional serial structure to a parallel system structure, thereby accelerating the working efficiency of the video image recognition system and completing a truly high-speed real-time video image recognition system.

9. The label-based augmented reality intelligent inspection method for substations according to claim 8, characterized in that: The preprocessing module adopts an enhanced pipeline architecture, which sequentially performs geometric transformation, color transformation, noise and blur processing, and advanced transformation operations on the input image. At the same time, it corrects radial, tangential, and perspective distortions through geometric correction, and achieves resource optimization and performance improvement by combining computing resource scheduling. Finally, the standardized enhanced image is pushed to the buffer queue to be recognized.

10. The label-based augmented reality intelligent inspection method for substations according to claim 1, characterized in that: The object recognition module uses MobileNetV3 as its backbone network. It adaptively selects the resolution based on the target size through a feature extraction accelerator and performs multi-level feature fusion. It integrates visual features, sensor data, and device logs to achieve multimodal recognition. The target features and state determination are optimized through real-time processing.

Images