Diagnosing the condition of overhead power line insulators

Abstract

The invention relates to power engineering. A method for diagnosing the condition of overhead power line insulators in real time using unmanned aerial vehicles (UAVs) includes the steps of modelling faults and collecting data. The YOLOv3 object detection and classification algorithm is used to diagnose contamination of an insulator and the YOLOv3 and YOLOv5 algorithms are used to detect mechanical damage to the insulator, and a critical condition of the insulator is modelled. Field tests are carried out to evaluate proposed models, during which the condition of insulators is diagnosed using intelligent diagnostic methods, as well as traditional diagnostic methods in the field of high-voltage engineering. An industry-accessible web interface is developed for analysing and classifying the condition of insulators on the basis of images from UAVs, and an open access platform is developed for determining the condition of an insulator. The technical result consists in enabling automated diagnostics of insulators using UAVs and also improving fault detection accuracy.

Description

[0001] DIAGNOSING THE CONDITION OF INSULATORS OF OVERHEAD POWER TRANSMISSION LINES

[0002] The invention relates to electric power engineering, in particular to diagnostics and monitoring of high-voltage power lines.

[0003] State of the art

[0004] To accurately assess the condition of transmission line insulators, expensive and time-consuming offline or online inspection methods have been widely used over the past century. To the best of the authors' knowledge, there is no similar diagnostic method in the market that can non-contactly detect the defect type (pollution / missing base / mechanical damage) and fault severity (acceptable / critical). In a study by Hu and Li (2021), channel attention in Faster RCNN and U-Net architectures was applied to insulator detection, achieving an average accuracy (AAP) of 91.9%. Wang and Yi (2021) reported an AAP of 94.5% in insulator detection by incorporating a Gaussian distribution in the bounding box prediction in YOLOv3. Han et al. (2020) proposed to improve YOLOv3-tiny by implementing a spatial pyramid (SPP) and region of interest (Rol) fusion architecture to detect and localize the missing fault. Sampedro et al.(2019) proposed an architecture combining fully convolutional Up-Net for segmentation, CNN, and Siamese CNN for insulator diagnosis purpose, which resulted in 99.25% HR. Xuan et al. (2022) proposed a one-stage CenterMask-based object detection method to detect missing insulator plate in UAV images and noise-added images, reporting accurate and high-speed performance of the proposed method. Wei et al. (2020) compared the performance of Faster R-CNN, RetinaNet, and YOLOv3 models for detecting and localizing missing string insulator cap due to self-explosion. The SSD model for aerial images was also applied to classify porcelain and composite insulators with accuracy of 93.75% and 85.29%, respectively, by Miao et al. (2019). Zan et al.(2022) claimed that the traditional tiny YOLOv4 architecture outperformed Faster R-CNN and EfficientDet in detecting the damaged insulator of overhead line, resulting in 91.75% mAP compared to 82.74% and 86.13%, respectively. Feng et al. applied YOLOv5 and its versions on UAV-captured images to identify and localize the defective part of string insulator and achieved a maximum mAP=0.995 (2021). Lui et al. (2020) proposed insulator detection based on YOLOv2 and YOLOv3 models, respectively. Zhao et al. (2020) trained Faster R-CNN model for high-speed and accurate insulator identification. Prates et al. (2019) also trained a CNN to classify the insulator type and structure material and detect faults based on images of a given insulator with a real background. Tao et al. (2020) modified the CNN algorithm using a region proposal network to detect and predict insulator faults.

[0005] Although several methods have been proposed for insulator surface contamination classification, no study has investigated state-of-the-art machine learning architectures that are capable of simultaneously performing both insulator and contamination detection and mechanical damage classification. This study is based on the application of state-of-the-art YOLO algorithms on UAV images to detect and classify insulator surface contamination and condition. Specifically, the employed architectures include the original YOLOv3, Y0L0v3-SPP, Y0L0v3-tiny, Y0L0v5n, Y0L0v5s, Y0L0v5m, Y0L0v5x, Y0L0v51, Y0L0v7, Y0L0v7-tiny, Y0L0v7x, Y0L0v8n, Y0L0v8s, Y0L0v8m, Y0L0v8x, and YOLO.v8l. These machine learning architectures have proven to be fast and robust one-step object detection models with real-time detection capabilities.

[0006] Most of the methods used require de-energizing the high-voltage transmission line equipment to accurately perform the assessment and monitoring methods. Contact methods are mainly used when the line is de-energized and the insulators are under autonomous control. Sometimes they are also accompanied by disconnecting the insulators from the supports, which is a risky task for repair and testing personnel in conditions of sufficiently high altitude. Manned aerial inspection involves conventional helicopter patrols near the transmission line, while unmanned inspection is carried out autonomously using UAVs or climbing robots, which also significantly reduces the risk to personnel. UAV imaging ensures high accuracy, personnel safety and low maintenance costs.At the same time, to assess the condition of the power transmission line, it is still necessary to send personnel to the site, use special equipment to lift a person (mainly a helicopter) and assess the power transmission line for an accurate solution. When inspecting sections of power transmission lines passing through marshy terrain, ravines, forest belts and water obstacles, the process becomes extremely complex. This method will simplify these tasks, it is able to identify threats at an early stage using an accurate assessment system, significantly reduce accidents, improve safety and save money. When carrying out emergency recovery work, the use of UAVs will save time on finding a faulty point and determining the causes of problems. In addition, UAV image processing based on machine learning (ML) has become more efficient compared to visual inspection or conventional image processing based on filtering, segmentation and spatial transformation.Finally, it is planned to develop an open online platform for automated classification of isolator conditions. This mechanism will enable on-site personnel to upload and analyze a dataset of isolator images and generate a full report, including the isolator condition and reliability assessment. This, in turn, has greater practical applicability, as it facilitates a more informed decision on further actions.

[0007] The objective of the invention is to develop an intelligent system based on advanced YOLO machine learning algorithms for detecting defects on power transmission line insulators.

[0008] The technical result is achieved by creating an intelligent system that allows processing images of power transmission line insulators and automatic classification of the object's condition. Classes include various mechanical damages and contamination of the insulator surface.

[0009] The description of the invention is explained by the drawings, where Fig. 1 shows the YOLOv3 network architecture, Fig. 2 shows a visualization of the bounding box predictions, Fig. 3 shows the change in the values ​​of τAP@0.5:0.95 during 20 training epochs for the studied YOLO architectures, Fig. 4 shows the change in the average value of τAP@0.5:0.95 over 100 epochs with 5-fold cross-validation and example images from a UAV, Fig. 5 shows various types of contamination and mechanical defects applied to polymer and porcelain insulators, Fig. 6 shows modeling of contamination of a glass insulator in the Ansys Maxwell program, Fig. 7 shows an example of a method for distributing contamination on the surface of a glass insulator in Ansys, Fig. 8 demonstrates the distribution of contamination on a glass suspension insulator in Ansys, Fig.9 shows the NSDD / ESDD measurement procedure, including collecting contamination from the insulator surface and measuring conductivity, drying filter paper in a desiccator and weighing the contaminant, Fig. 10 shows the operation diagram of the platform for classifying the insulator state, and Fig. 11 shows the preliminary results of applying the YOLO algorithm to localize and classify the state of various insulators.

[0010] At present, various methods for evaluating insulators have been developed, such as electrical testing, mechanical testing and routine testing. The electrical testing methods for insulators are not limited, but include low-frequency dry and wet melting and withstand voltage test, impulse breakdown and withstand voltage test, radio-frequency voltage test, visual corona test and puncture test. In parallel, the mechanical testing methods for insulator include ultimate mechanical strength test, combined mechanical and electrical strength test (applicable to suspension type insulators), withstand live load test, porosity test, heat test and spot gauge test.Conventional insulator inspections combine electrical and mechanical tests and mainly focus on (a) high voltage testing (usually applied to pin insulators), (b) proof load testing and (c) corrosion testing together with galvanizing (coating thickness) testing. In addition to the conventional standardized diagnostic tests, advanced non-contact assessment methods such as ultrasonic method, ultraviolet (UV) and infrared (IR) imaging examinations, radio interference (RI) assessment, self-normalized photothermal radiometry, manned aerial inspection using helicopter and, more recently, unmanned aerial vehicle (UAV) imaging method have also been actively developed and applied.

[0011] This paper discusses the diagnostics of the insulator condition using UAV image processing. UAV imaging provides high accuracy, personnel safety, and low maintenance costs. Technically, machine learning-based object detection algorithms can be divided into one-stage and two-stage methods. This method involves the use of the YOLO algorithm, which is a one-stage detector where object detection (regression task) and classification (classification task) are performed through a single neural network. The results show that the method successfully identifies the type of surface contamination (salt, soil, cement, snow) and mechanical damage (bird pecking, cracks, and missing caps) in various types of insulators.

[0012] Due to the highly competitive global electricity market and demand, utilities will require reliable, fast, accurate and responsive predictive or diagnostic systems to evaluate their assets, analyze results and make the right decisions. In addition, the transition from human-assisted test results interpretation to an automated system is currently considered necessary in the industry. This proposal will integrate a new and significant contribution to the monitoring and support of high-voltage insulators to solve one of the long-term problems associated with decision making regarding insulator performance. This invention is relevant for both utility operators and insulator owners.Since Kazakhstan is a fairly large country with significant harsh climatic conditions, and the energy network of Kazakhstan is growing very quickly, the results will be of great importance to Kazakhstan's generation, transmission and distribution networks in their strategy for maintaining assets, planned replacements, and minimizing the possibility of a wrong decision. This invention is also directly useful for entire industries that use electrical insulators for both high and low voltage levels.

[0013] This invention will identify various critical indicators for insulators, propose a new algorithm to determine the criticality of insulators, evaluate the state of the active part of the insulator, and implement effective asset management to improve the reliability of insulators. At this stage, there is no such reliable automated decision-making system in the world that can evaluate the operability and integrity of insulators and help the operator make a confident decision to de-energize the line. A new insulator evaluation model with an error evaluation report will be developed to detect the undesirable state of insulators to recommend proactive actions to utilities. The application of deep learning algorithms together with YOLO will facilitate accurate fault detection of insulators, and thus be the first step towards UAV-based automated diagnostics of insulators in the industry.This, in turn, will facilitate the timely detection of damage to insulators before extreme accidents and irreversible faults occur on the power transmission line.

[0014] 1) A large dataset containing images of porcelain, glass and composite post and suspension insulators with various surface contaminations and mechanical impurities was obtained in laboratory conditions. In particular, images of insulators with surface deposits, namely snow, soil, wet soil, cement and salt, as well as clean surfaces; and defects such as bird pecking damage, cracks and missing plates were obtained under different illumination, distances and camera angles.

[0015] 2) YOLOv3, -v5, -v7 and -v8 architectures and their variants are trained and compared to detect external insulators and classify their surface contamination and defects from UAV images.

[0016] 3) The model selection procedure is performed by comparing possible architectures in terms of performance and computational requirements. Two networks, namely InsuNet-Y8 and InsuNet-Y5, are proposed based on the best performance and the optimal trade-off between complexity and performance. The results show that the lighter architecture (i.e., InsuNet-Y5) can effectively explore isolators at the expense of a slight decrease in accuracy - this may be an attractive choice for embedded devices such as UAVs.

[0017] 4) The implementation of InsuNet-Y8 and InsuNet-Y5 in an industrial environment is discussed and demonstrated.

[0018] As an example, the architecture of the YOLOv3 network is shown in Fig. 1. The network is constructed by stacking 53 convolutional layers on a modified Darknet-53 network, resulting in 106 convolutional layers. Three detector heads are arranged in layers 79, 91 and 103 to detect large, medium and small objects, respectively. Predictions are generated in layers 82, 94 and 106 as feature maps whose sizes are 1 / 32, 1 / 16 and 1 / 8 of the input image size, respectively. We propose to use the YOLOv3 object detection process, which is carried out in detection blocks at three different scales. Using the output layer of the detection blocks, we create a feature map containing the bounding box prediction information, see Fig. 2. During the detection process, YOLOv3 predicts the width, height and center coordinates of the bounding box that completely surrounds the object and gives the predicted label a confidence score associated with the bounding box.The bounding box prediction is illustrated in Fig. 2, where the red rectangle represents the predicted bounding box around the object, the yellow dot indicates the coordinates of the center of the bounding box, and the blue rectangle highlights the cell responsible for the prediction.

[0019] Although the estimation of the weights of each individual model from the training set is called the learning process, the model selection in this study is a two-step procedure: Phase (I) selects the best epoch (the epoch that resulted in the highest wAR value at an IOU threshold in the range from 0.5 to 0.95, denoted as wAR@0.5:0.95) and its corresponding trained model, which we call the best individual model, and Phase (II) selects the “best” model for insulator surface diagnostics among the best individual models (i.e., the “best” of each used YOLO architecture trained in the best epochs).

[0020] To fine-tune the proposed “baseline models” InsuNet-Y8 and InsuNet-Y5, we first select the best epoch using 5-fold cross-validation. More precisely, the training set (i.e., 267 images acquired at the local substation) was randomly divided into 5 groups (folds). Each group was subjected to validation sequentially, and the remaining four groups were used to train the surrogate classifier for 100 epochs. After each training epoch, the models were evaluated on the validation groups using the wAP@0.5:0.95 metric. The average wAP@0.5:0.95 value for each epoch is calculated across all the retained folds. Finally, the epoch that resulted in the highest average wAP@0.5:0.95 value across all the retained folds was selected as the optimal (best) number of epochs. In Fig. 4 shows the change in the mean value of wAP@0.5:0.95 across different training epochs during 5-fold cross-validation.For InsuNet-Y5, the highest average wAP@0.5:0.95 value of 0.5698 was recorded at epoch 100, while InsuNet-Y8 achieved the highest average wAP@0.5:0.95 value of 0.7398 at epoch 88. As a result, the InsuNet-Y8 and InsuNet-Y5 models were fully tuned using the full training set (i.e., 267 industrial images) for 88 and 100 epochs, respectively.

[0021] This platform will be easily used by the user / operator by uploading images captured by the UAV in various formats. The dataset will be automatically analyzed and the platform will generate the status classification results. This is advantageous since this methodology does not require any specific technical skills from the personnel, such as programming or image processing. The reported classification results and operator confidence score will provide strong support to the operator in the decision-making process. The developed engine will provide a full report of the classification results.

[0022] Since this proposal plans to develop codes using Python programming language, it is possible to use cloud environment for this task and then develop application (app) for any portable device, see Fig. 10. To implement this, different IoT platforms can be used in the cloud computing section. This will be done through the web interface of the cloud service provider or through the application programming interface (API). It is planned to use and implement Amazon IoT platform or thinger.io because of its compatibility with the existing Ethernet Shield 2. After uploading the data to the cloud, we plan to store it in a cloud database such as Relational Database Service (RDS) or a data lake such as Amazon Simple Storage Service (S3). After that, we plan to connect the mobile application to the cloud using the API of the cloud service provider to retrieve the data.It is necessary to use the API keys and authentication mechanisms provided by the cloud service provider to secure the connection, and then upload the data to the mobile application. The corresponding libraries should be configured in the cloud, and the processor data (image data captured by the handheld device) should be integrated with the cloud server. Then, the proposed platform allows defining the variables and data to be transmitted by the program. However, a special technical code should be developed and defined for this task. Finally, a program should be developed to send the information through the cloud system to the Application. Photos can be taken using any handheld device, including photos captured by UAVs, and then processed in the cloud environment. Fig. 11 shows the preliminary results of applying the YOLO algorithm to localize and classify the state of different insulators. Fig.11 demonstrates the confident performance of the trained model in various challenging situations. For example, the model demonstrates robustness to additional background elements and multiple insulators in a single image. Moreover, the model successfully detects the insulator and predicts the damage class with high confidence despite low brightness and long-range imaging.

Claims

Invention formula An intelligent system for diagnosing the condition of overhead power line insulators in real time using unmanned aerial vehicles, characterized in that the system consists of systematized steps in which they model the contamination of the insulator surface and mechanical faults, perform practical selection and collection of data, model the contamination of the insulator surface, collect data on the contamination of the insulator surface, collect data on mechanical damage to the insulator, prepare a data set with the development of a system for detecting insulators and assessing the surface condition using intelligent methods and algorithms, use the most modern algorithms for detecting and classifying YOLO objects, train and select the YOLOv3 model for diagnosing insulator contamination, train and select the YOLOv3 and YOLOv5 models for detecting mechanical damage to the insulator, evaluate the trained models on a test set,obtained from the laboratory setup, conduct field tests to evaluate the proposed models, which diagnose the state of insulators using intelligent methods and traditional diagnostic methods in high-voltage technology, study the modeling of the critical state of the insulator, diagnose the pollution of the insulator and the severity of mechanical damage using experimental measurements of DDF, ESDD and NSDD, develop a web interface for industrial access to analyze and classify the state of insulators from UAV images, develop a free access platform for determining the state of the insulator.

Images