Transformer rack acceptance inspection data set construction method and system

Abstract

The invention relates to a transformer rack acceptance inspection data set construction method and system, and the method comprises the steps: cooperatively collecting multi-angle and multi-exposure images of a transformer rack through an industrial camera and an unmanned aerial vehicle, and synchronously recording the geographic position, attitude information and environmental parameters; a pre-training target detection model is adopted to generate an initial label, an electric power engineer manually corrects a component target frame and an installation key point, and a high-precision label file is formed; performing normalization, random rotation, noise addition and color transformation enhancement on the data, and dividing the data into a training set, a verification set and a test set in proportion; and through classified storage of a distributed file system, efficient access and model training are supported. Through multi-dimensional data fusion collection, semi-automatic labeling and standardized processing, the problems that an existing data set is small in scale and low in labeling quality are solved, the detection precision and generalization ability of the deep learning model for the installation defects of the transformer rack are remarkably improved, and reliable data support is provided for intelligent acceptance of power equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment detection, and in particular to a method and system for constructing a transformer stand acceptance detection data set. Background Art

[0002] In modern power systems, transformers are core equipment for power transmission. The installation quality and operating status of their test benches directly impact power system stability. As power networks expand and the number of test benches increases, manual inspections are inefficient and subjective, making them difficult to meet demand. While automated inspection methods based on computer vision and deep learning are gaining popularity, existing datasets for target-based key point inspections for transformer test bench installation acceptance suffer from small size, limited scenarios, and poor annotation. Constructing high-quality datasets is crucial.

[0003] Traditional manual inspection relies on visual inspection and simple tools to check the size and position of transformer racks. It has low efficiency and accuracy in complex environments and is greatly affected by subjective factors. Computer vision-based inspection uses algorithms such as image edge detection and template matching to determine the installation status of the rack. Although it improves efficiency, the detection effect is poor and the versatility is weak in complex backgrounds. Deep learning-based inspection uses CNN, YOLO and other algorithms to achieve rack target and key point detection, but due to the lack of dedicated large-scale high-quality data sets, the model generalization ability and detection accuracy are insufficient.

[0004] To this end, this application designs a method and system for constructing a transformer stand acceptance inspection dataset. By collecting a rich variety of transformer stand installation scene image data, covering different models of stands, multiple installation environments and working conditions, and combining it with high-precision annotation technology, the scale, diversity and accuracy of the data are ensured. This dataset is used to train a deep learning model to improve the model's detection accuracy, efficiency and generalization ability for transformer stand targets and key points, achieving rapid, accurate and automated acceptance inspection of transformer stand installation quality, overcoming the limitations of manual inspection, and solving problems existing in existing computer vision and deep learning detection technologies. This provides reliable technical support and data assurance for the installation and acceptance of power system transformer stands, and improves the safety and stability of power system operation. Summary of the Invention

[0005] In order to overcome the deficiencies in the prior art, the present invention provides a method and system for constructing a transformer stand acceptance test data set.

[0006] A method for constructing a transformer stand acceptance test data set, characterized by comprising the following steps: S1, collects transformer stand images through multi-angle, multi-exposure shooting and drone assistance; S2, synchronously records geographic location, posture information and environmental parameters; S3, uses a combination of semi-automatic annotation and manual correction to generate high-precision annotation data; S4 normalizes, enhances and partitions the dataset and stores it in a distributed file system.

[0007] Furthermore, in order to better realize the present invention, S1 is specifically, using a high-resolution industrial camera to perform multi-angle image acquisition of the transformer stand at intervals of 30° within the horizontal range of 360° and vertical directions including upward, horizontal and downward angles, and each group of angles acquires at least 10 images with different exposures. At the same time, a drone equipped with an industrial camera is used to perform stable photography of the transformer stand installed at high altitude, and the drone's flight altitude is 5-15 meters and the flight speed is 1-3 meters / second.

[0008] Furthermore, in order to better implement the present invention, the S2 specifically involves synchronously collecting the geographic location information, posture information and environmental parameters of the transformer stand, wherein the geographic location information is obtained through a combined positioning module of GPS and IMU with a positioning accuracy of centimeters. The environmental parameters include temperature, humidity and light intensity, and are recorded in real time through temperature and humidity sensors and light sensors.

[0009] Furthermore, in order to better implement the present invention, the S3 is specifically: S31: Input the collected image data into the pre-trained Faster R-CNN model for preliminary target detection, generating an initial annotation file containing the target boxes and key point coordinates of the transformer stand components, including brackets, crossarms, and insulators. Key points include bolt connection points and horizontal reference points. At S32, a team of power engineers manually revised the initial annotation files based on the power industry installation specifications. These revisions included adjusting the target frame position, precisely calibrating key point coordinates, and supplementing missing annotations. The annotation results were saved in JSON format and associated with the acquisition time, longitude and latitude coordinates, and environmental parameter metadata.

[0010] Furthermore, in order to better implement the present invention, the S4 is specifically, S41, preprocessing the annotated dataset, including normalizing the image to a resolution of 1024 × 1024 pixels, removing ambiguous or conflictingly annotated data, and performing data augmentation by random rotation of ±15°, horizontal or vertical translation of no more than 10% of the image size, scaling by 0.8–1.2 times, adding Gaussian or salt-and-pepper noise, and adjusting brightness and contrast; S42, divide the enhanced dataset into training set, validation set and test set in a ratio of 8:1:1, and store them through a distributed file system. The distributed file system adopts the Ceph architecture. The data directory is stored by test bench model and acquisition scene, and contains independent image folders, annotation folders and metadata description files.

[0011] Furthermore, in order to better realize the present invention, the resolution of the high-resolution industrial camera is not less than 5 million pixels, is equipped with a wide-angle lens, and achieves uniform lighting through a fill light device in a low-light environment, and the fill light intensity is 300-1000 lux.

[0012] Furthermore, in order to better realize the present invention, during the drone acquisition process, the ground control station presets the flight path as a closed polygon surrounding the platform, and the flight altitude is dynamically adjusted according to the platform installation height to ensure that the image covers the overall structure of the platform and the surrounding environment within 1 meter.

[0013] Furthermore, in order to better implement the present invention, the random rotation operation in the data enhancement is randomly selected within the range of ±15° with a step size of 5°, the probability of noise addition is 20%-50%, the Gaussian noise variance is 0.01-0.05, and the salt and pepper noise density is 0.1%-0.5%.

[0014] Based on the above transformer bench acceptance test dataset construction method, the transformer bench acceptance test dataset construction system includes: Data acquisition module: Integrates a high-resolution industrial camera, GPS+IMU positioning module, temperature and humidity sensors, light sensors, and a drone control unit, and is connected to the host computer via USB or Ethernet; Annotation processing module: deploys pre-trained Faster R-CNN models and annotation tools, supporting semi-automatic annotation generation and manual correction interfaces; Data enhancement module: configure random rotation, translation, scaling, noise addition and color transformation algorithms, and support parameter customization; Storage management module: implements data classification storage based on the Ceph distributed file system, and provides multi-user concurrent access and data version control functions.

[0015] Furthermore, in order to better implement the present invention, the annotation tool is a customized version of Labelme, which adds power component category templates, key point annotation shortcuts and metadata automatic filling functions.

[0016] The beneficial effects of the present invention are: Comprehensive and diverse data: This dataset utilizes a multi-dimensional data collection system to cover data from different transformer models, diverse installation environments, and operating conditions. This system not only captures image information but also simultaneously records geographic location and environmental parameters. This combined multi-angle, multi-exposure photography and drone-assisted aerial data collection effectively avoids data monotony, providing rich and realistic scene data for model training and significantly enhancing the dataset's coverage and diversity. High-Precision Annotation and Standardization: This system utilizes a combination of semi-automatic annotation and manual revision, relying on a pre-trained model to generate initial annotations. Professional power engineers then refine these annotations according to industry standards, ensuring that the annotations accurately capture the target frames of test bench components and the coordinates of key installation points. The annotation results are formatted uniformly and include detailed metadata, significantly improving both quality and standardization compared to existing datasets, which often suffer from vague and haphazard annotation. Strong application adaptability and scalability: Distributed file system storage and standardized directory organization facilitate data access and updates. Whether applied to different algorithm research or to meet the ever-changing detection needs of power engineering sites, it has excellent adaptability and good expansion potential. DETAILED DESCRIPTION

[0017] This embodiment provides a method and system for constructing a transformer stand acceptance test data set.

[0018] The existing transformer bench acceptance test has the following problems: Manual inspection: Relying on the inspector's visual inspection and simple tools, it is inefficient and difficult to carry out work in complex environments such as high altitude and low light conditions. The inspection results are greatly influenced by the operator's experience and subjective judgment, and the ability to identify subtle defects and potential problems is poor, which cannot meet the inspection needs of the large-scale development of power networks. Computer vision-based inspection: This method uses algorithms such as image edge detection and template matching to perform bench installation inspection. However, it is susceptible to interference in complex backgrounds and has difficulty accurately identifying targets and key points. A large number of templates must be pre-made, which lacks universality for transformer benches of different models and installation scenarios, making it difficult to adapt to diverse inspection tasks.

[0019] Deep learning-based detection: Although automated detection is achieved using algorithms such as CNN and YOLO, there is a lack of large-scale, high-quality dedicated datasets for transformer test bench installation and acceptance scenarios. As a result, the trained models have poor generalization capabilities in practical applications, and the detection accuracy and stability of test benches in different environments and installation conditions are poor, which cannot meet the high-precision requirements of engineering practice.

[0020] The data acquisition system of this embodiment consists of an image acquisition device, a positioning device, and an environmental parameter acquisition device. The image acquisition device uses a high-resolution industrial camera (e.g., with a resolution of at least 5 megapixels) equipped with a wide-angle lens to ensure a complete image of the transformer stand and its surroundings. A fill light device is also used to address low-light inspection scenarios. The positioning device uses a high-precision GPS module (with centimeter-level positioning accuracy) combined with an inertial measurement unit (IMU) to obtain real-time information about the device's geographic location and posture during image acquisition. The environmental parameter acquisition device includes temperature and humidity sensors and light intensity sensors to record and collect environmental parameters. Each device is connected to a host computer via USB or Ethernet. The host computer runs customized data acquisition software to control the devices and simultaneously acquire data. Data collection is conducted on transformer rigs of various models and installation methods at various power project sites. The camera is controlled to capture images of the rig from multiple angles (horizontally, images are captured every 30° across a 360° range; vertically, images are captured from upward, horizontal, and downward perspectives). Each capture consists of at least 10 images at varying exposures to cover varying lighting conditions. For high-altitude transformer rigs, image acquisition equipment is carried out using drones. The drone's flight altitude and speed are set by the ground control station to ensure stable and accurate image acquisition. Simultaneously, data from positioning equipment and environmental parameter acquisition equipment is recorded during the acquisition process to provide auxiliary information for subsequent data annotation and analysis. Data annotation utilizes a combination of semi-automatic annotation and manual correction. Pre-trained general-purpose object detection models (such as the ResNet-based Faster R-CNN model) are first used to perform preliminary detection on the captured images, identifying the transformer rack's main structure and possible keypoint locations. Initial annotation boxes and keypoint coordinates are generated. A team of professional power engineers and annotators then manually corrects and supplements the initial annotation results based on power industry standards and transformer rack installation specifications. This annotation process includes object box annotations for transformer rack components (such as brackets, crossarms, and insulators), as well as precise coordinate annotations for key installation points (such as component connection points, fixing bolt locations, and horizontal reference points). Professional annotation tools (such as Labelme) are used during the annotation process. The annotation results are stored in JSON format, including information such as the image file name, annotation category, object box coordinates, keypoint coordinates, and acquisition location and environmental parameters. The annotated data was preprocessed, including image normalization (resizing images to a uniform size of 1024×1024 pixels) and data cleaning (removing blurry, damaged, and incorrectly labeled images). To increase the diversity of the dataset, data augmentation techniques were employed, including random rotation (within ±15°), translation (horizontally and vertically no more than 10% of the image size), scaling (0.8-1.2x), noise addition (Gaussian and salt-and-pepper), and color transformation (adjusting brightness, contrast, and saturation). The processed data was divided into training, validation, and test sets in an 8:1:1 ratio, used for deep learning model training, hyperparameter tuning, and performance evaluation, respectively. The processed dataset is organized into a unified directory structure, consisting of image folders, annotation file folders, and metadata files. The image folders store image data categorized by acquisition scene, rig model, and other factors; the annotation file folders store JSON files with annotations for the corresponding images; and the metadata files record basic dataset information (such as acquisition time, location, equipment model, and data size). The dataset is stored on a disk array on a high-performance server and managed using a distributed file system (such as Ceph). This ensures data storage security, scalability, and fast accessibility, facilitating subsequent deep learning model training and related research based on the dataset.

[0021] The following is an example of a JSON file: { "filename": "Outdoor Overhead Stand_S11_20250511_Horizontal 30 Degrees.jpg", "labels": [ { "label":"bracket", "bbox": [100, 200, 300, 400] / / target box coordinates, [x_min, y_min, x_max, y_max] }, { "label": "Insulator", "bbox": [400, 300, 500, 450] }, { "label": "Fixing Bolt", "keypoints": [ / / Key point coordinates [250, 350], [350, 350] ] } ], "metadata": { "Collection Time": "2025-05-11 10:30:00", "Geographic Location": "Longitude:116.4074, Latitude:39.9042", "Environmental parameters": { "Temperature": 25, "Humidity": 40, "Light Intensity": 500 } } } The key points of this embodiment are: Multi-dimensional data collection system: A collection system consisting of high-resolution industrial cameras, high-precision positioning equipment, and environmental parameter collection equipment is constructed to enable the simultaneous collection of multi-dimensional data, from image information, geographic location, and environmental parameters. Through multi-angle and multi-exposure photography, as well as high-altitude drone acquisition, comprehensive data acquisition covers various types of transformer rigs in various installation scenarios, addressing the existing data collection issues of limited single-scenario scenarios and insufficient information dimensionality. Precise labeling and quality control: A labeling method that combines semi-automatic labeling with manual correction is adopted. Initial labeling is generated based on a pre-trained general model, and then manually corrected by a professional team based on power industry standards to ensure that the labeling content covers the target frame of the test bench components and the precise coordinates of key installation points. The labeling results contain rich metadata information, effectively improving the accuracy and standardization of the labeling. Data augmentation: Using a variety of data augmentation techniques, such as random rotation, translation, and noise addition, to significantly expand the size and diversity of the dataset; Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention and are not limiting. Other modifications or equivalent substitutions made to the technical solution of the present invention by ordinary technicians in this field should be included in the scope of the claims of the present invention as long as they do not depart from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for constructing a transformer stand acceptance test data set, characterized in that: The following steps are involved: S1, collects transformer stand images through multi-angle, multi-exposure shooting and drone assistance; S2, synchronously records geographic location, posture information and environmental parameters; S3, uses a combination of semi-automatic annotation and manual correction to generate high-precision annotation data; S4 normalizes, enhances and partitions the dataset and stores it in a distributed file system.

2. The method for constructing a transformer stand acceptance test data set according to claim 1, characterized in that: Specifically, S1 is to use a high-resolution industrial camera to capture multi-angle images of the transformer stand at intervals of 30° within the horizontal range of 360° and in the vertical direction including upward, horizontal and downward angles, and to capture at least 10 images with different exposures for each group of angles. At the same time, a drone equipped with an industrial camera is used to perform stable photography of the transformer stand installed at high altitude. The drone's flight altitude is 5-15 meters and the flight speed is 1-3 meters per second.

3. The method for constructing a transformer stand acceptance test data set according to claim 1, characterized in that: Specifically, S2 is to synchronously collect the geographic location information, posture information and environmental parameters of the transformer stand, wherein the geographic location information is obtained through a combined GPS and IMU positioning module with a positioning accuracy of centimeters. The environmental parameters include temperature, humidity and light intensity, and are recorded in real time through temperature and humidity sensors and light sensors.

4. The method for constructing a transformer stand acceptance test data set according to claim 1, characterized in that: The S3 is specifically: S31: Input the collected image data into the pre-trained Faster R-CNN model for preliminary target detection, generating an initial annotation file containing the target boxes and key point coordinates of the transformer stand components, including brackets, crossarms, and insulators. Key points include bolt connection points and horizontal reference points. At S32, a team of power engineers manually revised the initial annotation files based on the power industry installation specifications. These revisions included adjusting the target frame position, precisely calibrating key point coordinates, and supplementing missing annotations. The annotation results were saved in JSON format and associated with the acquisition time, longitude and latitude coordinates, and environmental parameter metadata.

5. The method for constructing a transformer stand acceptance test data set according to claim 1, characterized in that: The S4 is specifically: S41, preprocessing the annotated dataset, including normalizing the image to a resolution of 1024 × 1024 pixels, removing ambiguous or conflictingly annotated data, and performing data augmentation by random rotation of ±15°, horizontal or vertical translation of no more than 10% of the image size, scaling by 0.8–1.2 times, adding Gaussian or salt-and-pepper noise, and adjusting brightness and contrast; S42, divide the enhanced dataset into training set, validation set and test set in a ratio of 8:1:1, and store them through a distributed file system. The distributed file system adopts the Ceph architecture. The data directory is stored by test bench model and acquisition scene, and contains independent image folders, annotation folders and metadata description files.

6. The method for constructing a transformer stand acceptance test data set according to claim 2, characterized in that: The resolution of the high-resolution industrial camera is not less than 5 million pixels, and it is equipped with a wide-angle lens. In low-light environments, uniform lighting is achieved through a fill light device, and the fill light intensity is 300-1000 lux.

7. The method for constructing a transformer stand acceptance test data set according to claim 2, characterized in that: During the drone acquisition process, the ground control station presets the flight path as a closed polygon surrounding the platform, and the flight altitude is dynamically adjusted according to the platform installation height to ensure that the image covers the overall structure of the platform and the surrounding environment within 1 meter.

8. The method for constructing a transformer stand acceptance test data set according to claim 5, characterized in that: In the data augmentation, the random rotation operation is randomly selected within the range of ±15° with a step size of 5°, the probability of noise addition is 20%-50%, the Gaussian noise variance is 0.01-0.05, and the salt and pepper noise density is 0.1%-0.5%.

9. A transformer stand acceptance test data set construction system, based on the transformer stand acceptance test data set construction method according to any one of claims 1 to 8, characterized in that: include: Data acquisition module: Integrates a high-resolution industrial camera, GPS+IMU positioning module, temperature and humidity sensors, light sensors, and a drone control unit, and is connected to the host computer via USB or Ethernet; Annotation processing module: deploys pre-trained Faster R-CNN models and annotation tools, supporting semi-automatic annotation generation and manual correction interfaces; Data enhancement module: configure random rotation, translation, scaling, noise addition and color transformation algorithms, and support parameter customization; Storage management module: implements data classification storage based on the Ceph distributed file system, and provides multi-user concurrent access and data version control functions.

10. The transformer stand acceptance test data set construction system according to claim 9, characterized in that: The labeling tool is a customized version of Labelme, which adds power component category templates, key point labeling shortcuts and metadata automatic filling functions.