Separation and identification method for multi-source hybrid ultra-high frequency partial discharge diagram

Abstract

The present invention discloses a separation and identification method for multi-source hybrid ultra-high frequency partial discharge diagram, which comprises: obtaining a PRPD diagram for ultra-high

Description

TECHNICAL FIELD The present application belongs to the field of partial discharge signal identification, and particularly relates to a separation and identification method for a multi-source hybrid ultra-high frequency partial discharge diagram based on digital image target detection and pairing technology. BACKGROUND Gas Insulated Switchgear (GIS) equipment has been widely used not only in high-voltage and ultra-high voltage fields, but also in extra-high voltage fields since its commercialization in the 1960s. However, in some complex working environments or in the process of manufacturing and installation, it is inevitable to produce some internal defects and safety hazards, such as poor conductor contact, conductive particles, metal tips, air gaps in insulating parts, etc., leading to various types of partial discharges, which will in turn lead to GIS insulation failures and power system accidents. Therefore, regularly conducting ultra-high frequency partial discharge detection on GIS equipment in operation, analyzing and diagnosing internal partial discharge of GIS based on the detected partial discharge PRPD diagram, timely discovering and eliminating internal defects and hidden dangers of GIS, has become an important part of GIS equipment operation and maintenance. Identification of the partial discharge type is an important part of partial discharge detection in GIS, mainly based on partial discharge PRPD diagram. At present, for the identification of GIS partial discharge type, in addition to relying on the maintenance personnel to identify manually based on experience, many research institutions have researched and developed artificial intelligence identification methods based on image identification technology to realize the automation of partial discharge identification. This type of method has the following problem: for a partial discharge diagram data, it can only identify based on the overall feature of the diagram image, and only output one partial discharge type as the final identification result. However, in practice, there are often multiple partial discharge signals coexisting within the GIS, as well as internal partial discharge signals coexisting with external noise signals, the PRPD diagram presents a mixed overlay image of multiple partial discharge signal diagrams, resulting in the diagram being "neither fish nor fowl" and causing identification errors. Even if the type of one of these sources is correctly identified, other coexisting sources are bound to be missed. SUMMARY OF THE INVENTION The purpose of the present application is to provide a separation and identification method for multi-source hybrid ultra-high frequency partial discharge diagram, which is used to solve the problem that existing methods are not able to detect and identify each of the multiple signal sources one-by-one when the diagram is coexisting with multiple signal sources. A separation and identification method for a multi-source hybrid ultra-high frequency partial discharge diagram, comprising the following steps: step 1: obtaining PRPD diagrams for GIS partial discharge signals by an ultra-high frequency partial discharge detection device, performing process to convert the PRPD diagrams into standardization grayscale images, and constructing a partial discharge sample database; step 2: labeling the partial discharge sample database, boxing the diagram region corresponding to each signal source in each sample diagram one-by-one, labeling its corresponding signal type, and constructing a PRPD diagram dataset; step 3: determining a deep learning network as detection model and inputting detection model parameters; step 4: dividing the PRPD diagram dataset and inputting to the deep learning network detection model to perform training, adjusting the detection model and training parameters, and retaining the trained models; step 5: evaluating the trained models and selecting a model with optimal effect as a PRPD diagram detection model for partial discharge signals; step 6: performing discharge signal detection on a diagram to be tested and detecting various targets of partial discharge and noise interference within a corresponding grayscale image; step 7: analyzing and confirming the detection targets one-by-one, retaining a final detection and multi-source separation result of the detection targets that meet standards. In one embodiment, in the step 1, the PRPD diagrams are obtained by an ultra-high frequency partial discharge detection device, the PRPD diagrams comprise: independent partial discharge signals and interference noise signals, as well as several mixed partial discharge and interference noise signals; converting the PRPD diagrams into grayscale images, and the range of normalized grayscale values for each pixel's pulse count at that position is [0,255], that is, the maximum number of pulses is normalized to 255, and other pulse numbers are normalized to integers within the range of [0,254] in the same proportion. In one embodiment, in the step 2, using an image annotation tool "labelme" to label the partial discharge sample database, wherein for each diagram in the partial discharge sample database, using a rectangular box to label the positions of the partial discharge signals and the interference noise signals in the PRPD diagram, and signal types thereof; the suspended discharge signals, the solid insulation discharge signals, the particle discharge signals are two clusters with a phase deviation of 180 degrees in the PRPD diagram which are labeled separately using two rectangular boxes; The PRPD diagram dataset comprises five types of targets to be tested of suspended discharge, solid insulation discharge, particle discharge, tip discharge, and interference noise. In one embodiment, in the step 3, the deep learning detection model is based on YOLOv3 detection algorithm. In one embodiment, in the step 3, darknet-53 is used as backbone network of the Y0L0v3 detection algorithm. In one embodiment, in the step 3, the deep learning network detection model comprises feature maps of 13 * 13, 26 * 26, and 52 * 52 pixels, and the size used for the feature maps described is the average size of 9 anchor boxes clustered by the K-means algorithm. In one embodiment, in the step 4, the PRPD diagram dataset comprises a training set, a validation set, and a test set, wherein the ratio of the training set, the validation set, and the test set is 6:2:2. In one embodiment, in the step 4, setting parameters of batch (the number of diagrams for each input detection model for training) and an initial learning rate, and dynamically adjust the learning rate according to the change of loss function; The condition of the trained model is that the value of the loss function fluctuates within a small range and does not decrease. In one embodiment, in the step 5, the model with optimal effect is tire trained model with the highest mAP on the test set. In one embodiment, in the step 7, the analyzing and confirming the detection targets one-by-one further comprises: for the targets to be tested of the suspended discharge, the solid insulation discharge, and the particle discharge, in the area where phase offset is 180 degrees and amplitude difference is within lOdB, if there are the same type of targets, they will be paired and used as the same detected partial discharge signals; if there are no the same type of targets, discarding the targets. The present application has the following technical effects: 1. Solve the problem of not being able to detect multiple partial discharge signals separately when using image identification algorithms to identify discharge types in GIS ultra-high frequency PRPD diagrams; 2. Not only accurately perform multi-source separation from multiple partial discharge signals mixed diagram, but also identify the types of separated signals separately; 3. Change the traditional GIS discharge signal detection method, achieve automated detection and multi-source separation, and improve the initial candidate box parameters using K-means clustering algorithm, effectively improving identification speed. Moreover, the mAP on the test set has reached over 95%, enabling more accurate and efficient real-time monitoring of GIS discharge signals. BRIEF DESCRIPTION OF THE FIGURES Figure 1 is a flowchart of the present application; Figure 2 is a schematic diagram of the standardization of the PRPD diagrams and dataset annotation of the present application; In the figure: (a) - (e) are the grayscale images of the standardized PRPD diagrams, and (f) - (j) are the annotated diagrams of the dataset; Figure 3 is a schematic diagram of the YOLOv3 target detection algorithm network structure of the present application; Figure 4 is a schematic diagram of the detection process for the tested diagram of the present application. DETAILED DESCRIPTION The following will further elaborate on the technical solution of the present application in conjunction with embodiments and accompanying drawings. The PRPD diagram, also known as the Phase Resolved Partial Discharge, is a statistical diagram formed by the partial discharge detection device collecting multiple partial discharge ultra-high frequency pulses over a period of time, and simultaneously collecting and recording the grid frequency phase (0-360 °) at the time of each pulse collection. The diagram is a two-dimensional coordinate plane with scatter points. The horizontal axis represents the grid frequency phase of the pulse (0-360 °), and the vertical axis represents the power of the pulse (generally ranging from -80 to 0 dBm). The value of each scatter point represents the number of pulses that occur at that coordinate during the collecting time period. This embodiment provides a separation and identification method for a multi-source hybrid ultra-high frequency partial discharge diagram, comprising the following steps: step 1: obtaining PRPD diagrams for GIS partial discharge signals by an ultra-high frequency partial discharge detection device, performing process to convert the PRPD diagrams into standardization grayscale images, and constructing a partial discharge sample database; step 2: labeling the partial discharge sample database, boxing the diagram region corresponding to each signal source in each sample diagram one-by-one, labeling its corresponding signal type, and constructing a PRPD diagram dataset; step 3: determining a deep learning network as detection model and inputting detection model parameters; step 4: dividing the PRPD diagram dataset and inputting to the deep learning network detection model to perform training, adjusting the detection model and training parameters, and retaining the trained models; step 5: evaluating the trained models and selecting a model with optimal effect as a PRPD diagram detection model for partial discharge signals; step 6: performing discharge signal detection on a diagram to be tested and detecting various targets of partial discharge and noise interference within a corresponding grayscale image; step 7: analyzing and confirming the detection targets one-by-one, retaining a final detection and multi-source separation result of the detection targets that meet standards. In the further implementation of this embodiment, in the step 1, the PRPD diagrams are obtained by an ultra-high frequency partial discharge detection device. The PRPD diagrams includes four types of independent partial discharge signal diagrams, various types of external interference signal diagrams, and various types of mixed partial discharge and external interference signal diagrams; The PRPD diagrams are converted into grayscale images and obtain a total of 50000 grayscale images corresponding to the PRPD diagrams. Among them, 10000 grayscale images are obtained for each of the four types of independent partial discharge signals, and 10000 grayscale images are obtained for mixed partial discharge signals. The horizontal and vertical coordinates are quantized into 416 units respectively, and the grayscale values of each pixel are normalized to the range of [0,255] for the number of pulses at that position, that is, the maximum number of pulses is normalized to 255, and other pulse numbers are normalized to integers within the range of [0,254] in the same proportion. In the further implementation of this embodiment, in the step 2, the partial discharge sample data is labeled using the image annotation tool "labelme" to generate an XML file containing annotation information. For each diagram in the partial discharge sample database, a rectangular box is used to separately label the regions of each partial discharge signal and each noise signal in the PRPD diagram, as well as their respective signal types; the signals of suspension discharge, the solid insulation discharge, and the particle discharge are two clusters with a phase deviation of 180 degrees in the PRPD diagram, and are labeled separately using two rectangular boxes; The PRPD diagram dataset includes five types of targets to be tested: suspended discharge, solid insulation discharge, particle discharge, tip discharge, and interference noise. In the further implementation of this embodiment, in the step 3, the deep learning detection model is based on the YOLOv3 detection algorithm, with an input image size parameter of 416 * 416 pixels. Yolo is an object detection algorithm, and the task of object detection is to find objects from the diagram and provide their categories and positions. It applies a single convolutional neural network (CNN) to the entire diagram, divides the diagram into grids, and predicts the class probability and bounding boxes of each grid. In the further implementation of this embodiment, in the step 3, darknet-53 is used as the backbone network of the YOLOv3 detection algorithm. The darknet-53 includes 53 layers of convolutions, with the first 52 layers set as the main network and the last layer as the fully connected layer. In the further implementation of this embodiment, in the step 3, the deep learning network detection model includes feature maps of 13 * 13, 26 * 26, and 52 * 52 pixels. The size of the feature maps used is the average size of 9 anchor boxes clustered by the K-means algorithm. The number of prior boxes is 13 * 13 * 3+26 * 26 * 3+52 * 52 * 3=10647, and each prediction is a 10 dimensional vector (4+1+5=10). The 10 dimensional vector includes: border coordinates (4 values), border confidence (1 value), probability of object type (4 discharge signals+1 noise signal); The final output depth of each feature map is: 3 * (4+l+5)=30; The loss function includes: center coordinate error, width and height coordinate errors, confidence error, and type error. In the further implementation of this embodiment, in the step 4, the PRPD diagram dataset includes a training set, a validation set, and a test set. The ratio of the training set, the validation set, and the test set is 6:2:2, with quantities of 30000, 10000, and 10000, respectively. In the further implementation of this embodiment, in the step 4, the batch is set to 64, which means that a forward propagation is performed after every 64 samples are trained to update the network parameters, the total number of iterations is set to 400000, the initial learning rate is 0.001, and the learning rate is dynamically adjusted according to the change of the loss function, when iterating to 200000 times, the learning rate is decayed by ten times, and when iterating to 300000 times, the learning rate is decayed ten times more on the basis of the previous learning rate; The condition for the trained models is that the value of the loss function fluctuates within a small range and does not decrease, and the number of retained models is 5. In the further implementation of this embodiment, in the step 5, the five trained models are tested on the test set, mAP (average accuracy of five types of targets) is calculated, and the model with the highest mAP is selected as the final detection model. In the further implementation of this embodiment, in the step 7, the detection targets are analyzed and confinned one-by-one. Specifically, for the test targets of the tip discharge signals and interference noise signals, only two types of targets are detected, and the detection results are retained; For the targets to be tested of suspended discharge, the solid insulation discharge, and the particle discharge, in the area where phase offset is 180 degrees and amplitude difference is within lOdB (corresponding to a horizontal translation of 208 pixels and a vertical difference of 52 pixels on the grayscale image), if there are the same type of targets, they will be paired and used as the same detected partial discharge signals; if there are no the same type of targets, discarding the targets. Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present application, and not to limit it; Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the aforementioned embodiments or equivalently replace some of the technical features thereof. And these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present application.

Claims

1. A separation and identification method for multi-source hybrid ultra-high frequency partial discharge diagram, comprising the following steps:step 1: obtaining PRPD diagrams for GIS partial discharge signals, performing standardization process on the PRPD diagrams, and constructing a partial discharge sample database;step 2: labeling the partial discharge sample database and constructing a PRPD diagram dataset for discharge signals and noise signals;step 3: determining a deep learning network to serve as detection model and inputting detection model parameters;step 4: dividing up the PRPD diagram dataset and inputting it into deep learning network detection models for training, adjusting the detection models and training parameters, and retaining the trained models;step 5: evaluating the trained models and selecting a model with optimal effect as a PRPD diagram detection model for partial discharge signals;step 6: performing discharge signal detection on a diagram to be tested and detecting various targets of partial discharge and noise interference within a corresponding grayscale image;step 7: analyzing and confirming the detection targets one-by-one, retaining a final detection and multi-source separation result of the detection targets that meet standards;the PRPD diagram dataset comprises: tested targets of suspended discharge, solid insulation discharge, particle discharge, tip discharge, and interference noise; in the step 2, using an image annotation tool to label the partial discharge sample database, wherein for each diagram in the partial discharge sample database, using a rectangular box to label the positions of the partial discharge signals and the interference noise signals in the PRPD diagram, and signal types thereof; wherein the suspended discharge signals, the solid insulation discharge signals, the particle discharge signals are two clusters with a phase deviation of 180 degrees in the PRPDdiagram which are labeled separately using two rectangular boxes;in the step 7, the analyzing and confirming the detection targets one-by-one further comprises: for the tested targets of the suspended discharge, the solid insulation discharge, and the particle discharge, in the area where phase offset is 180 degrees and amplitude difference is within lOdB, if there are the same type of targets, they will be paired and used as the same detected partial discharge signals; if there are no the same type of targets, discarding the targets.

2. The separation and identification method for multi-source hybrid ultra-high frequency partial discharge diagram of claim 1, wherein in the step 1, the PRPD diagrams are obtained by an ultra-high frequency partial discharge detection device, the PRPD diagrams comprise: independent partial discharge signals and interference noise signals, as well as several mixed partial discharge and interference noise signals;converting the PRPD diagrams into grayscale images, and the range of normalized grayscale values for each pixel's pulse count at that position is [0,255],3. The separation and identification method for multi-source hybrid ultra-high frequency partial discharge diagram of claim 1, wherein in the step 3, the deep learning detection model is based on YOLOv3 detection algorithm.

4. The separation and identification method for multi-source hybrid ultra-high frequency partial discharge diagram of claim 3, wherein in the step 3, darknet-53 is used as backbone network of the YOLOv3 detection algorithm.

5. The separation and identification method for multi-source hybrid ultra-high frequency partial discharge diagram of claim 4, wherein in the step 3, the deep learning network detection model comprises feature maps of 3 types of pixels, and the size used for the feature maps described is the average size of 9 anchor boxes clustered by K-means algorithm.

6. The separation and identification method for multi-source hybrid ultra-high frequency partial discharge diagram of claim 1, wherein in the step 4, the PRPD diagram dataset comprises a training set, a validation set, and a test set, wherein the ratio of the training set, the validation set, and the test set is 6:2:2.

7. The separation and identification method for multi-source hybrid ultra-high frequency partial discharge diagram of claim 1,wherein in the step 4, setting the number of diagrams and an initial learning rate for each input detection model for training, and dynamically adjust the learning rate according to the change of loss function;the condition for the trained models is that the value of the loss function fluctuates within a small range and does not decrease.

8. The separation and identification method for multi-source hybrid ultra-high frequency partial discharge diagram of claim 1, wherein in the step 5, the model with optimal effect is the trained model with the highest mAP on the test set.A. CLASSIFICATION OF SUBJECT MATTERG06T 7 / 00(2017.01)i; G06V 10 / 764(2022.01)1According to International Patent Classification (IPC) or to both national classification and IPCB. FIELDS SEARCHEDMinimum documentation searched (classification system followed by classification symbols) G06T, G06V , G06FDocumentation searched other than minimum documentation to the extent that such documents are included in the fields searchedElectronic data base consulted during the international search (name of data base and, where practicable, search terms used)CNTXT, ENTXT, DWPI, CNKI, IEEE: WIW< HB, if 15, M iOl, ffi, Hit. W, fig, fiff, partialw discharge, PRPD, graph, map, spectrogram, detection, identification, recognize, noise, grayscale, overlay, phase, bias, offsetDOCUMENTS CONSIDERED TO BE RELEVANTCategory* Citation of document, with indication, where appropriate, of the relevant passages Relevant to claim No. PX CN 115187527 A (SHANGHAI GELUBU TECHNOLOGY CO., LTD.) 14 October 2022 (2022-10-14) claims 1-10 1-8 A CN 112149549 A (TAIAN POWER SUPPLY COMPANY OF STATE GRID SHANDONG ELECTRIC POWER COMPANY et al.) 29 December 2020 (2020-12-29) description, paragraphs [0056]-[0071] 1-8 A CN 112198399 A (RED PHASE INC.) 08 January 2021 (2021-01-08) entire document 1-8 A CN 114186589 A (SHANGHAI ELECTRIC POWER COMPANY OF STATE GRID et al.) 15 March 2022 (2022-03-15) entire document 1-8 A CN 106556781 A (PDSTARS TECHNOLOGY (SHANGHAI) CO., LTD.) 05 April 2017 (2017-04-05) entire document 1-8 A CN 114417926 A (SHANDONG UNIVERSITY) 29 April 2022 (2022-04-29) entire document 1-8| | Further documents are listed in the continuation of Box C.annex.* Special categories of cited documents:“A” document defining the general state of the art which is not considered to be of particular relevance“D” document cited by the applicant in tire international application“E” earlier application orpatent but published on or after the international“T”“O”“P”filing datedocument which may throw doubts on priority claim(s) or which is cited to establish the publication date of another citation or other special reason (as specified)document referring to an oral disclosure, use, exhibition or other meansdocument published prior to the international filing date but later than the priority date claimed‘Y’later document published after the international filing date or priority date and not in conflict with the application but cited to understand the principle or theory underlying the inventiondocument of particular relevance; the claimed invention cannot be considered novel or cannot be considered to involve an inventive step when the document is taken alonedocument of particular relevance; the claimed invention cannot be considered to involve an inventive step when the document is combined with one or more other such documents, such combination being obvious to a person skilled in the artdocument member of the same patent familyDate of the actual completion of the international searchDate of mailing of the international search report11 September 2023Name and mailing address of the ISA / CNChina National Intellectual Property Administration (ISA / CN)China No. 6, Xitucheng Road, Jimenqiao, Haidian District,Beijing 100088Authorized officer15 September 2023Telephone No.C. DOCUMENTS CONSIDERED TO BE RELEVANTCategory* Citation of document, with indication, where appropriate, of the relevant passages Relevant to claim No. A WO 2015070513 Al (STATE GRID CORPORATION OF CHINA et al.) 21 May 2015 (2015-05-21) entire document 1-8 A KR 20200116425 A (MYONGJI UNIVERSITY INDUSTRY AND ACADEMIA COOPERATION FOUNDATION) 12 October 2020 (2020-10-12) entire document 1-8International application No.PCT / CN2023 / 101087Patent document cited in search report Publication date (day / month / year) Patent family member)s) Publication date (day / month / year) CN 115187527 A 14 October 2022 None CN 112149549 A 29 December 2020 None CN 112198399 A 08 January 2021 None CN 114186589 A 15 March 2022 None CN 106556781 A 05 April 2017 None CN 114417926 A 29 April 2022 None WO 2015070513 Al 21 May 2015 CA 2918679 Al 21 May 2015 CN 103558529 A 05 February 2014 KR 20200116425 A 12 October 2020 None

Images