Blind source separation method for multi-source fault spectrum of oil-immersed power transformer

Abstract

The blind source separation method of multi-source fault graphs of oil-immersed power transformers includes the following steps: inputting the multi-source partial discharge fault graph into the CNN encoder, adding a ResNet network to the generated network, and introducing a gradient penalty term based on the adversarial loss function of the traditional CGAN network; using a deep learning dedicated server for multiple training and verification, continuously optimizing and adjusting the model training parameters, and obtaining a network weight data file that meets the requirements of the partial discharge fault diagnosis task of the oil-immersed power transformer; using the trained YOLOv8 network as the diagnosis network, identifying the fault type of each generated single-source graph, clarifying the fault type of the multi-source graph, and deducing and diagnosing the fault state of the oil-immersed power transformer; deploying the obtained Encoder‑CGAN network and YOLOv8 network weight data files to the power inspection equipment and platform to complete the fault monitoring and deduction diagnosis tasks of the oil-immersed power transformer. Break through the difficulty of deducing and diagnosing the fault type of the partial discharge fault graph of the oil-immersed power transformer.

Description

Technical Field

[0001] The invention belongs to the field of state monitoring and fault deduction diagnosis of oil-immersed power transformers, and particularly relates to a blind source separation method for multi-source fault graphs of oil-immersed power transformers. Background Art

[0002] The key to the stable operation of the power system lies in the timely and effective diagnosis of power equipment faults. Power equipment such as mutual inductors, circuit breakers, transformers and bushings are exposed to the natural environment for a long time and are susceptible to failures under harsh conditions, which seriously affects the normal operation of the power system. In recent years, deep learning technology has been widely used in the field of power equipment fault diagnosis. Compared with traditional machine learning methods, deep learning has end-to-end learning capabilities. It uses graph feature extraction and target detection, and learns the deep features of the graph through multi-level nonlinear transformation, thereby improving the robustness of power equipment fault detection.

[0003] Most of the existing graph detection algorithms are based on deep neural networks, such as Faster R-CNN, YOLO, Transformer and other networks. These network models perform relatively well in terms of graph detection accuracy and robustness, but usually have a large number of parameters and computational complexity, resulting in slow model operation speed, making it difficult to meet the real-time requirements of power transmission and transformation equipment fault monitoring and diagnosis, and difficult to deploy on mobile terminals. In addition, these network models have low direct detection accuracy for multi-fault source fusion graphs (such as multi-source partial discharge fault graphs), which makes it difficult to meet the needs of oil-immersed power transformer fault monitoring and deduction diagnosis tasks. Therefore, the research on network models and technical means that can achieve both lightweight and accurate detection of multi-fault source fusion graphs has become an important research direction in the field of power equipment fault monitoring and diagnosis.

[0004] Single-source fault maps can well reflect certain faults of oil-immersed power transformers, and the detection difficulty is relatively low. However, in actual engineering applications, once a certain type of fault (such as partial discharge) occurs in an oil-immersed power transformer, it is often accompanied by multiple faults, such as surface discharge accompanied by corona discharge. Therefore, the fault maps collected by the edge detection equipment usually present a fusion state of two or more fault feature information, accompanied by noise interference feature information generated by the equipment itself and the environment, and even spectrum distortion and feature area overlap, which makes it difficult for the spectrum detection algorithm to accurately extract the feature information of the spectrum and accurately diagnose the fault type of the oil-immersed power transformer.

[0005] The partial discharge phase distribution (PRPD) spectrum is mainly collected by pulse current method, ultra-high frequency detection method, and ultrasonic detection method, and processed by the power equipment status comprehensive monitoring device (partial discharge instrument) or high-frequency oscilloscope. The PRPD spectrum is a two-dimensional scatter plot. Generally, the x-axis represents the phase, the y-axis represents the signal strength or amplitude, and the accumulated color depth and density of the points represent the density of the discharge pulse here. The type of partial discharge can be determined by the PRPD spectrum characteristics, and the severity and location of partial discharge can be diagnosed by combining the internal structure characteristics of the electrical equipment and the ultrasonic detection and positioning results.

[0006] At present, a large number of deep learning methods and technologies have been used for fault spectrum diagnosis of power equipment. For example: Wu M, Jiang W, Shen D, et al. Multi-source partial discharge pattern recognition algorithm based on D-conditional generative adversarial network-Yolov5[J]. IEEE Transactions on Power Delivery, 2022. doi: 10.1109 / TPWRD.2022.3222317. This paper uses D-conditional generative adversarial network to generate and expand partial discharge fault spectrum, and uses improved YOLOv5 network to realize multi-source partial discharge fault spectrum diagnosis, thus solving the problems of scarcity of partial discharge sample spectrum and difficulty in multi-source spectrum recognition to a certain extent; Zheng H, Sun Y, LiuX, et al. Infrared image detection of substation insulators using an improved fusion single shot multibox detector[J]. IEEE transactions on power delivery, 2020, 36(6): 3351-3359. This document designs a new feature fusion module in the shallow part of the SSD model to improve the model's feature extraction capability for the infrared spectrum of the transformer, thereby improving the detection accuracy of the model.

[0007] Although the above methods have achieved good detection results, they have not completely solved the problem of poor diagnosis effect of multi-source fault maps of oil-immersed power transformers under complex backgrounds, and the model robustness is not high enough, which further affects the model's diagnosis of faults. Therefore, after research, the present invention proposes a blind source separation method for multi-source fault maps of oil-immersed power transformers. Summary of the invention

[0008] The technical problem to be solved by the present invention is to provide a blind source separation method for multi-source fault spectra of oil-immersed power transformers. The present invention can help power inspection departments to efficiently diagnose hidden dangers of oil-immersed power transformers and effectively reduce the failure risks of oil-immersed power transformers, which is of great significance to maintaining the stability of the power system.

[0009] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0010] The blind source separation method of multi-source fault spectrum of oil-immersed power transformer includes the following steps:

[0011] S1. Collect multi-source partial discharge fault maps, perform background denoising, grayscale processing and map compression on the collected partial discharge fault maps, obtain partial discharge fault map data with background and pixel interference removed and uniform map size, and then screen, annotate and expand the partial discharge fault map data to obtain a sample data set suitable for subsequent deep learning network model training, verification and testing;

[0012] S2. Input the multi-source partial discharge fault map in the sample data set into the Convolutional Neural Networks (CNN) encoder: extract the core feature vector of the multi-source partial discharge fault map through multiple convolutional layers, pooling layers, batch normalization layers and fully connected layers, and use the core feature vector and the single-source feature conditions corresponding to five types of partial discharge as conditional generative adversarial network inputs; the five types of partial discharge include tip discharge, suspended discharge, surface discharge, air gap discharge and particle discharge;

[0013] S3. Add a deep neural network architecture ResNet (Residual Network) to the generative network in the conditional generative adversarial network, combine the input single-source feature conditions, extract the single-source feature vector in the partial discharge fault map feature vector, and input the single-source feature vector into the multi-layer ResNet, convolution layer, upsampling layer and activation function to achieve single-source map generation;

[0014] S4. Using the standard PatchGAN structure of the identification network in the conditional generative adversarial network, the generated single-source partial discharge spectrum is identified according to the real single-source spectrum. If the single-source partial discharge spectrum is true, the generated single-source spectrum is output, otherwise it is discarded;

[0015] S5. In order to alleviate the problems of gradient vanishing and mode collapse of traditional conditional generative adversarial networks, a gradient penalty term is introduced based on the adversarial loss function of traditional conditional generative adversarial networks to improve the stability of training, alleviate the mode collapse problem, and improve the quality of generated samples;

[0016] S6. Using the sample data set constructed in step S1 as the input of the Encoder-CGAN network (convolutional neural network encoder and conditional generative adversarial network) and YOLOv8 (You Only Look Once version 8) network constructed in steps S2 to S5, multiple training and verification are performed using a deep learning dedicated server, and the model training parameters are continuously optimized and adjusted to obtain a network weight data file that meets the requirements of the partial discharge fault deduction and diagnosis task of oil-immersed power transformers; wherein the English abbreviation of the conditional generative adversarial network is CGAN (conditional generative adversarial network);

[0017] S7. In order to accurately identify the fault type of the partial discharge fault spectrum, the YOLOv8 network trained in step S6 is used as a diagnostic network to identify the single source spectrum type, determine the fault type of the partial discharge fault spectrum, and then diagnose the fault state of the oil-immersed power transformer;

[0018] S8. Deploy the Encoder-CGAN network and YOLOv8 network weight data files obtained in step S6 to the corresponding power inspection equipment and platform to complete the oil-immersed power transformer fault monitoring and deduction diagnosis tasks.

[0019] In a preferred solution, step S1 is divided into the following steps:

[0020] S11. Experiments were conducted using a self-designed partial discharge model of oil-immersed power transformers to collect single-source and multi-source partial discharge fault maps;

[0021] S12, performing background denoising on the collected partial discharge fault spectrum (hereinafter referred to as PRPD spectrum), setting a threshold to filter out grid lines and sine waves, and converting the black background into a white background;

[0022] S13, performing grayscale processing on the PRPD spectrum obtained in step S12 using a perceptual weighting method based on the ITU-R BT.601 standard to obtain a PRPD grayscale spectrum;

[0023] S14. In order to better input the deep learning network, the PRPD grayscale map is compressed, and then the map is pre-processed by screening and expansion to construct a single-source and multi-source partial discharge data set suitable for the deep learning network;

[0024] S15. Construct a YOLOv8 network model training data set, use the Labelme image annotation tool to annotate the single-source partial discharge fault data set, and organize the annotated data set according to the YOLO data set format.

[0025] In a preferred solution, step S2 is divided into the following steps:

[0026] S21, inputting the partial discharge fault map into the CNN encoder, and extracting the feature vector of the partial discharge fault map through multiple convolutional layers, pooling layers, and batch normalization layers;

[0027] S22. At the same time, the extracted partial discharge fault spectrum feature vector and the single-source feature conditions of the five partial discharge types are used as conditional generative adversarial network inputs.

[0028] In a preferred solution, in step S21, the partial discharge fault map is input into a CNN encoder, and a feature vector of the partial discharge fault map is extracted through multiple convolutional layers, pooling layers, and batch normalization layers. The implementation steps are as follows:

[0029] S211. Design CNN encoder, optimize the structure of convolution layer, pooling layer, batch normalization layer, and ensure effective extraction of feature maps of input graphs;

[0030] S212, input the partial discharge fault map into the CNN encoder, extract it through the convolution layer, the pooling layer, and the batch normalization, and obtain the feature vector of the partial discharge fault map.

[0031] In a preferred solution, in step S22, the extracted feature vector of the partial discharge fault map and the five single-source feature conditions of the partial discharge fault map are used as inputs of the conditional generative adversarial network, and the implementation steps are as follows:

[0032] S221. Obtain single-source characteristic conditions of five types of partial discharge (tip discharge, suspension discharge, surface discharge, air gap discharge and particle discharge) spectra through convolutional neural network;

[0033] S222, optimizing the conditional generative adversarial network to generate network convolutional layers, pooling layers and other structures, so as to facilitate the conditional input of multi-source feature vectors and single-source feature vectors;

[0034] S223, inputting the partial discharge fault spectrum feature vector and the single source feature condition extracted in step S21b into the generation network.

[0035] Preferably, in step S3, the generation network extracts the single-source feature vectors that meet the single-source feature vector conditions from the partial discharge fault spectrum feature vectors, and obtains the single-source spectrum through operations such as convolution, and the implementation steps are as follows:

[0036] S31. Add ResNet network to the generative network to achieve efficient extraction of single-source feature vectors while alleviating the problem of gradient vanishing that may occur in the generative network;

[0037] S32, extracting a single source feature vector from a partial discharge fault spectrum feature vector in combination with the input single source feature condition;

[0038] S33, the extracted single-source feature vector is passed through multiple layers of ResNet layers, convolution layers, upsampling layers and activation functions to generate a single-source map.

[0039] In the preferred solution, in step S31, a ResNet network is added to each conditional generative adversarial network generation network to achieve efficient extraction of single-source feature maps that meet the single-source feature conditions in the partial discharge fault spectrum feature vector and alleviate the problem of gradient vanishing that may occur in the generation network.

[0040] In a preferred solution, step S4 is divided into the following steps:

[0041] S41, the identification network uses the standard PatchGAN structure as a binary classifier to distinguish the true and false generated graphs;

[0042] S42, taking the real single-source partial discharge spectrum and the generated single-source spectrum as input to the identification network for identification, if true, output the generated single-source spectrum, otherwise discard it, so as to achieve the purpose of blind source separation of the partial discharge fault spectrum into a single-source spectrum.

[0043] In the preferred solution, in step S5, in order to alleviate the problems of gradient vanishing and mode collapse of the conditional generative adversarial network, a gradient penalty term is introduced on the basis of the adversarial loss function to improve the stability of training, alleviate the mode collapse problem, and improve the quality of generated samples, which is divided into the following three parts:

[0044] S51, known generation network G The goal is to generate high-quality single-source partial discharge spectrum samples G(z|y) , so that these samples tend to be real in the identification network, so the loss function of the generated network is expressed as:

[0045] (1);

[0046] In the formula, L G is the loss value of the generated network, D To identify the network, y is the characteristic condition of single-source partial discharge, G(z|y) Give the generator a condition y and the noise vector z The generated samples, z is the characteristic vector of partial discharge fault spectrum, p z （·) is the characteristic data distribution of partial discharge fault spectrum, p(·) is the conditional feature data distribution, For p z (·) expectations;

[0047] S52. In order to avoid the problem of unstable training of conditional generative adversarial networks or adversarial imbalance between the generator network and the discriminator network, a gradient penalty term is introduced, which is defined as:

[0048] (2);

[0049] In the formula, is generated from linear interpolation between the generated image and the real image, λ is a hyperparameter that controls the weight of the gradient penalty term. For the discriminator D On input (Given the conditions y ), represents the gradient;

[0050] S53, known identification network D The goal is to identify the real single-source partial discharge samples and generate single-source partial discharge samples. The identification process also needs to consider the single-source partial discharge characteristic conditions. y The loss function of the identification network is usually expressed as:

[0051] (3);

[0052] In the formula, L D To identify the loss value of the network, x A sample of a real single-source partial discharge spectrum. D(x|y) Give conditions to the discriminator y Next, we consider the real sample x is the true probability, p data (x,y) is the distribution of real data, For P data (x,y) expectations.

[0053] In a preferred embodiment, step S6 is divided into the following steps:

[0054] S61, using the data set constructed in step S1 as input to the Encoder-CGAN network and the YOLOv8 network constructed in steps S2 to S5;

[0055] S62. Use a dedicated deep learning server to perform multiple training and verifications, continuously optimize and adjust the network parameters of the Encoder-conditional generative adversarial network, and obtain a network weight data file that meets the requirements of the multi-source fault map blind source separation task;

[0056] S63. Use a dedicated deep learning server to conduct multiple training and verifications, continuously optimize and adjust the network parameters of YOLOv8, and obtain a network weight data file that meets the requirements of the partial discharge fault diagnosis task for oil-immersed power transformers.

[0057] Preferably, in step S62 and step S63, the implementation steps are: using Pytorch as the environment for model training, verification and testing, training 16 samples per batch, and training 300 batches in total; the initial learning rate is set to 0.01.

[0058] In the preferred solution, in step S7, in order to accurately identify the fault type of the partial discharge fault spectrum, the trained YOLOv8 network is used as the diagnostic network to identify the fault type of the single source spectrum, determine the fault type of the partial discharge fault spectrum, and deduce and diagnose the fault state of the oil-immersed power transformer.

[0059] In the preferred solution, in step S8, the Encoder-CGAN network and YOLOv8 network weight data files obtained in step S6 are deployed to the corresponding inspection platform to complete the oil-immersed power transformer fault monitoring and deduction diagnosis tasks.

[0060] A blind source separation system for multi-source fault spectrum of oil-immersed power transformer, the system comprising:

[0061] Acquisition unit: used to collect multi-source partial discharge fault maps, perform background denoising, grayscale processing and map compression on the collected partial discharge fault maps, obtain partial discharge fault map data with background and pixel interference removed and uniform map size, and then screen, annotate and expand the partial discharge fault map data to obtain a sample data set suitable for subsequent deep learning network model training, verification and testing;

[0062] Processing unit: used to input the multi-source partial discharge fault map in the sample data set into the convolutional neural network encoder: extract the core feature vector of the multi-source partial discharge fault map through multiple convolution layers, pooling layers, batch normalization layers and fully connected layers, and use the core feature vector and the single-source feature conditions corresponding to five types of partial discharge as conditional generative adversarial network input; the five types of partial discharge include tip discharge, suspension discharge, surface discharge, air gap discharge and particle discharge;

[0063] Single-source graph generation unit: used to add a deep neural network architecture ResNet network to the generation network in the conditional generative adversarial network, combine the input single-source feature conditions, extract the single-source feature vector in the partial discharge fault graph feature vector, and input the single-source feature vector into the multi-layer ResNet, convolution layer, upsampling layer and activation function to achieve single-source graph generation.

[0064] The system also includes:

[0065] Identification unit: used to identify the single-source partial discharge spectrum using the PatchGAN structure in the conditional generative adversarial network. The identification process is: identify the generated single-source partial discharge spectrum according to the real single-source spectrum. If the single-source partial discharge spectrum is true, the generated single-source spectrum is output, otherwise it is discarded.

[0066] Deduction and diagnosis unit: used to use the sample data set as the input of the convolutional neural network encoder and the conditional generative adversarial network, and as the input of the conditional generative adversarial network YOLOv8 network, and use the deep learning dedicated server to adjust the model training parameters to obtain the network weight data file that meets the requirements of the partial discharge fault diagnosis task of the oil-immersed power transformer; use the trained YOLOv8 network as the diagnosis network to identify the fault type of each generated single-source map, clarify the fault type of the multi-source map, and deduce and diagnose the fault state of the oil-immersed power transformer; deploy the network weight data file described in the Encoder-CGAN network to the power inspection equipment and platform to perform oil-immersed power transformer fault monitoring and deduction diagnosis.

[0067] A computer device comprising:

[0068] one or more processors;

[0069] The processor is used to store one or more programs;

[0070] When the one or more programs are executed by the one or more processors, the method for blind source separation of multi-source fault graphs of oil-immersed power transformers is implemented.

[0071] A computer-readable storage medium stores a computer program, which, when executed, implements the blind source separation method for multi-source fault spectrum of an oil-immersed power transformer.

[0072] The present invention can achieve the following beneficial effects:

[0073] The present invention proposes an Encoder-CGAN network, which extracts a multi-source fault spectrum feature vector, combines the input single-source feature conditions, and generates a partial discharge spectrum of the single-source partial discharge type contained therein; wherein, a ResNet network is added to the generation network, so as to realize the efficient extraction of the single-source feature vector in the partial discharge fault spectrum feature vector, and reduce the possibility of gradient vanishing; the identification network adopts a standard PatchGAN structure, and inputs the generated single-source partial discharge spectrum and the real single-source partial discharge spectrum for comparison and identification, and if it is true, the generated single-source spectrum is output, otherwise it is discarded, thereby achieving the purpose of blind source separation of multi-source fault spectra; on the basis of the adversarial loss function, a gradient penalty term is introduced to further avoid the problem of gradient vanishing, so that the Encoder-CGAN network approaches Nash equilibrium; in addition, a YOLOv8 network is adopted as a diagnostic identification network to identify the fault type of the generated single-source partial discharge spectrum, clarify the fault type of the partial discharge fault spectrum, and break through the difficult problem of accurate diagnosis of the partial discharge fault spectrum. Finally, the improved Encoder-conditional generative adversarial network model network, YOLOv8 model network and the obtained weight data file are deployed to the power inspection equipment and platform to complete the oil-immersed power transformer fault monitoring and deduction diagnosis tasks. The present invention can help the power inspection department to efficiently diagnose the hidden dangers of oil-immersed power transformers and effectively reduce the failure risk of oil-immersed power transformers, which is of great significance to maintaining the stability of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0075] Figure 1 It is a schematic diagram of the implementation process of an embodiment of the present invention;

[0076] Figure 2 is the oil-immersed power transformer fault model of this embodiment;

[0077] Figure 3 It is a partial discharge fault spectrum of some oil-immersed power transformers according to an embodiment of the present invention;

[0078] Figure 4 This is the process of background denoising and grayscale processing of partial discharge spectrum of oil-immersed power transformer in an embodiment of the present invention;

[0079] Figure 5 The CNN encoder structure of the embodiment of the present invention;

[0080] Figure 6 A network structure is generated for an embodiment of the present invention;

[0081] Figure 7 The ResNet network module structure of an embodiment of the present invention;

[0082] Figure 8 An identification network structure according to an embodiment of the present invention;

[0083] Fig. 9 The overall structure of the Encoder-CGAN network of an embodiment of the present invention;

[0084] Fig.10 It is the YOLOv8 network structure of the embodiment of the present invention;

[0085] Fig.11 This is a diagram showing the effects of testing and deploying a multi-source partial discharge fault spectrum of some oil-immersed power transformers according to an embodiment of the present invention. DETAILED DESCRIPTION

[0086] The preferred solution is Figures 1 to 11 As shown, the blind source separation method of multi-source fault spectrum of oil-immersed power transformer includes the following steps:

[0087] S1. Using the self-designed partial discharge model of oil-immersed power transformer, the partial discharge fault maps of single-source and multi-source oil-immersed power transformers are collected experimentally, such as Figure 2 As shown in the figure, the collected partial discharge fault map is subjected to background denoising, grayscale processing and map compression, and then the map is pre-processed by screening, labeling and expansion to establish a special sample data set suitable for deep learning network;

[0088] S2. Input the partial discharge fault map into the CNN encoder, extract the feature vector of the partial discharge fault map through multiple convolutional layers, pooling layers, and batch normalization layers, and use it and the single-source feature conditions of five partial discharge types (tip discharge, suspension discharge, surface discharge, air gap discharge, and particle discharge) as the input of the conditional generative adversarial network;

[0089] S3, adding a ResNet network to the generation network, combining the input single-source feature conditions, extracting the single-source feature vector in the partial discharge fault spectrum feature vector, inputting it into the transposed convolution layer and the deconvolution layer, and generating a single-source spectrum;

[0090] S4, the identification network uses the standard PatchGAN structure, takes the generated single-source partial discharge spectrum and the generated single-source spectrum as input for identification, and outputs the generated single-source spectrum if it is true, otherwise it is discarded, so as to achieve the purpose of blind source separation of multi-source spectrum into single-source spectrum;

[0091] S5. In order to alleviate the problems of gradient vanishing and mode collapse in traditional conditional generative adversarial networks, a gradient penalty term is introduced based on the adversarial loss function to improve the stability of training, alleviate the mode collapse problem, and improve the quality of generated samples;

[0092] S6. Using the data set constructed in step S1 as the input of the Encoder-CGAN network and the YOLOv8 network constructed in steps S2 to S5, multiple training and verification are performed using a deep learning dedicated server, and the model training parameters are continuously optimized and adjusted to obtain a network weight data file that meets the requirements of the partial discharge deduction and diagnosis task of the oil-immersed power transformer;

[0093] S7. In order to accurately identify the generated single-source spectrum fault type, the YOLOv8 network model trained in step S6 is used as the diagnosis network to identify the single-source spectrum fault type, determine the partial discharge fault spectrum fault type, and deduce and diagnose the fault state of the oil-immersed power transformer;

[0094] S8, deploying the Encoder-CGAN network and YOLOv8 network weight data files obtained in step S6 to the corresponding inspection platform to complete the oil-immersed power transformer fault monitoring and deduction diagnosis tasks;

[0095] The above steps complete the blind source separation and diagnosis tasks of oil-immersed power transformer spectrum.

[0096] In a preferred embodiment, step S1 includes the following steps:

[0097] S11. Experiments were conducted using a self-designed partial discharge model of oil-immersed power transformers to collect single-source and multi-source partial discharge fault maps. Partial discharge fault maps of oil-immersed power transformers are shown in Figure 1. Figure 3 As shown;

[0098] S12, performing background denoising on the collected PRPD spectrum, setting a threshold to filter out grid lines and sine waves, and converting the black background into a white background;

[0099] S13, the PRPD spectrum obtained in step S12 is gray-processed by using the perceptual weighting method based on the ITU-R BT.601 standard to obtain a PRPD gray-scale spectrum. Steps S12 and S13 are as follows: Figure 4 As shown;

[0100] S14. In order to better input the deep learning network, the PRPD grayscale map is compressed, and then the map is pre-processed by screening and expansion to construct a single source and partial discharge fault map dataset suitable for the deep learning network;

[0101] S15. Construct a YOLOv8 network model training data set, use the Labelme image annotation tool to annotate the single-source partial discharge fault data set, and organize the annotated data set according to the YOLO data set format to construct a training set, a validation set, and a test set.

[0102] Preferably, the features in step S12 include:

[0103] S121, after classifying and arranging the collected 5 types of single-source partial discharge maps of oil-immersed power transformers and 10 types of partial discharge fault maps, set the three primary color RGB thresholds, and filter out interference including grid lines, sine waves, and noise scatter points;

[0104] S122, setting the color adaptive threshold inversion to change the black background to a white background.

[0105] Preferably, the features in step S13 include:

[0106] The PRPD spectrum obtained in step S12 is gray-processed using a perceptual weighting method based on the ITU-R BT.601 standard to obtain a PRPD gray-scale spectrum. The gray-scale processing formula is:

[0107] (1);

[0108] In the formula, R , G , B Respectively represent the values ​​of the red, green, and blue primary color channels of the spectrum. Gray This is the image after grayscale processing.

[0109] Preferably, the features in step S14 include:

[0110] S141. Compress the PRPD grayscale map, and then perform preprocessing operations including screening and expansion on the map. The data maps include: 4,000 single-source fault grayscale maps, including 800 tip discharge maps, 640 suspension discharge maps, 800 surface discharge maps, 800 air gap discharge maps, and 800 particle discharge maps; 2,000 multi-source fault grayscale maps, mainly multi-source fault maps of two random combinations of the above five single-source partial discharge signals, a total of 10 types, 200 maps each;

[0111] S142. Construct a single-source and multi-source partial discharge fault sample data set suitable for a deep learning network, and construct a training set and a validation set in a ratio of 8:2;

[0112] Preferably, the features in step S15 include:

[0113] S151. Labelme image annotation tool is used to annotate the single-source partial discharge fault data set.

[0114] S152. Arrange the data set annotated according to the YOLO data set format, and construct the training set, validation set, and test set in a ratio of 8:1:1.

[0115] In a preferred embodiment, step S2 includes the following steps:

[0116] S21, input the partial discharge fault map into the CNN encoder, and extract the core feature vector of the partial discharge fault map through multiple convolution layers, pooling layers, and batch normalization layers. The structure of the CNN encoder is as follows: Figure 5 As shown;

[0117] S22. It and the single-source characteristic conditions of five local discharge types (tip discharge, suspension discharge, surface discharge, air gap discharge and particle discharge) are used as the input of the conditional generative adversarial network generation network.

[0118] In a preferred embodiment, step S3 includes the following steps:

[0119] S31. Add ResNet network to the generated network, and generate a network structure like Figure 6 As shown, the ResNet network module structure is as follows Figure 7 As shown;

[0120] S32. The generation network combines the input single-source feature conditions to extract the single-source feature vector in the partial discharge fault spectrum feature vector, and inputs it into the transposed convolution layer and the deconvolution layer to realize single-source spectrum generation.

[0121] In a preferred embodiment, step S4 includes the following steps:

[0122] S41, the identification network uses the standard PatchGAN structure, and the identification network structure is as follows Figure 8 As shown;

[0123] S42, using the single-source partial discharge spectrum in the data set and the generated single-source spectrum as input for identification, if true, outputting the generated single-source spectrum, otherwise discarding it, so as to achieve the purpose of blind source separation of the partial discharge fault spectrum into a single-source spectrum;

[0124] In the preferred solution, step S5 is characterized by the loss function part of the conditional generative adversarial network. In order to alleviate the problems of gradient vanishing and mode collapse of the conditional generative adversarial network, a gradient penalty term is introduced on the basis of the adversarial loss function to improve the stability of training, alleviate the mode collapse problem, and improve the quality of generated samples. It is specifically divided into the following three parts:

[0125] S51, known generation network G The goal is to generate high-quality single-source partial discharge spectrum samples G(z|y) , so that these samples tend to be real in the identification network, so the loss function of the generated network is expressed as:

[0126] (2);

[0127] In the formula, L Gis the loss value of the generated network, D To identify the network, y is the characteristic condition of single-source partial discharge, G(z|y) Give the generator a condition y and the noise vector z The generated samples, z is the characteristic vector of partial discharge fault spectrum, p z （·) is the characteristic data distribution of partial discharge fault spectrum, p(·) is the conditional feature data distribution, For p z (·) expectations;

[0128] S52. In order to avoid the problem of unstable training of conditional generative adversarial networks or adversarial imbalance between the generator network and the discriminator network, a gradient penalty term is introduced, which is defined as:

[0129] (3);

[0130] In the formula, is generated from linear interpolation between the generated image and the real image, λ is a hyperparameter that controls the weight of the gradient penalty term. For the discriminator D On input (Given the conditions y ), represents the gradient;

[0131] S53, known identification network D The goal is to identify the real single-source partial discharge samples and generate single-source partial discharge samples. The identification process also needs to consider the single-source partial discharge characteristic conditions. y The loss function of the identification network is usually expressed as:

[0132] (4);

[0133] In the formula, L D To identify the loss value of the network, x A sample of a real single-source partial discharge spectrum. D(x|y) Give conditions to the discriminator y Next, we consider the real sample x is the true probability, p data (x,y) is the distribution of real data, For P data (x,y)expectations.

[0134] In the formula, L D To identify the loss value of the network, x A sample of a real single-source partial discharge spectrum. D(x|y) It is the sample characteristic of the real single-source partial discharge spectrum.

[0135] In a preferred embodiment, step S6 includes the following steps:

[0136] S61, using the data set constructed in step S1 as the input of the Encoder-CGAN network and the YOLOv8 network constructed in steps S2 to S5, as follows: Fig. 9 The Encoder-CGAN network structure of the embodiment is shown as follows: Fig.10 The YOLOv8 network structure of the embodiment is shown;

[0137] S62, using a dedicated deep learning server to perform multiple training and verification, continuously optimizing and adjusting the network parameters of the Encoder-conditional generative adversarial network, and obtaining a network weight data file that meets the requirements of the multi-source graph blind source separation task;

[0138] S63. Use a dedicated deep learning server to conduct multiple training and verifications, continuously optimize and adjust the network parameters of YOLOv8, and obtain a network weight data file that meets the requirements of the partial discharge fault diagnosis task for oil-immersed power transformers.

[0139] Furthermore, in step S62 and step S63, the implementation steps are to use Pytorch as the environment for model training, verification and testing; wherein, the training set includes 3200 single-source fault maps and 1600 multi-source fault maps, 16 samples are trained in each batch, and a total of 200 batches are trained; the initial learning rate is set to 0.01.

[0140] In the preferred solution, step S7 is characterized by: using the YOLOv8 network model trained in step S6 as the diagnostic network, identifying the single-source spectrum type, determining the partial discharge fault spectrum type input by the original encoder in step S21, and then determining the fault state of the oil-immersed power transformer.

[0141] In the preferred solution, step S8 is characterized by: deploying the Encoder-CGAN network and YOLOv8 network weight data files obtained in step S6 to the corresponding inspection platform / equipment to complete the oil-immersed power transformer fault monitoring and deduction diagnosis tasks;

[0142] Furthermore, the proposed blind source separation method of multi-source fault spectrum of oil-immersed power transformer is used to diagnose the partial discharge spectrum of oil-immersed power transformer on site, detect the fault characteristics of oil-immersed power transformer, and judge the corresponding fault type, such as Fig.11 shown.

[0143] Preferably, Example 1 shows that the method of the present invention can well complete the task of oil-immersed power transformer fault diagnosis, solves the problem of multi-source fault spectrum diagnosis, and the detection speed can theoretically meet the requirements of on-site oil-immersed power transformer real-time monitoring. At the same time, the model size meets the requirements of power inspection equipment and platform deployment.

[0144] A blind source separation system for multi-source fault spectrum of oil-immersed power transformer, the system comprising:

[0145] Acquisition unit: used to collect multi-source partial discharge fault maps, perform background denoising, grayscale processing and map compression on the collected partial discharge fault maps, obtain partial discharge fault map data with background and pixel interference removed and uniform map size, and then screen, annotate and expand the partial discharge fault map data to obtain a sample data set suitable for subsequent deep learning network model training, verification and testing;

[0146] Processing unit: used to input the multi-source partial discharge fault map in the sample data set into the convolutional neural network encoder: extract the core feature vector of the multi-source partial discharge fault map through multiple convolution layers, pooling layers, batch normalization layers and fully connected layers, and use the core feature vector and the single-source feature conditions corresponding to five types of partial discharge as conditional generative adversarial network input; the five types of partial discharge include tip discharge, suspension discharge, surface discharge, air gap discharge and particle discharge;

[0147] Single-source graph generation unit: used to add a deep neural network architecture ResNet network to the generation network in the conditional generative adversarial network, combine the input single-source feature conditions, extract the single-source feature vector in the partial discharge fault graph feature vector, and input the single-source feature vector into the multi-layer ResNet, convolution layer, upsampling layer and activation function to achieve single-source graph generation.

[0148] The system also includes:

[0149] Identification unit: used to identify the single-source partial discharge spectrum using the PatchGAN structure in the conditional generative adversarial network. The identification process is: identify the generated single-source partial discharge spectrum according to the real single-source spectrum. If the single-source partial discharge spectrum is true, the generated single-source spectrum is output, otherwise it is discarded.

[0150] Deduction and diagnosis unit: used to use the sample data set as the input of the convolutional neural network encoder and the conditional generative adversarial network, and as the input of the conditional generative adversarial network YOLOv8 network, and use the deep learning dedicated server to adjust the model training parameters to obtain the network weight data file that meets the requirements of the partial discharge fault diagnosis task of the oil-immersed power transformer; use the trained YOLOv8 network as the diagnosis network to identify the fault type of each generated single-source map, clarify the fault type of the multi-source map, and deduce and diagnose the fault state of the oil-immersed power transformer; deploy the network weight data file described in the Encoder-CGAN network to the power inspection equipment and platform to perform oil-immersed power transformer fault monitoring and deduction diagnosis.

[0151] A computer device comprising:

[0152] one or more processors;

[0153] The processor is used to store one or more programs;

[0154] When the one or more programs are executed by the one or more processors, the method for blind source separation of multi-source fault graphs of oil-immersed power transformers is implemented.

[0155] A computer-readable storage medium stores a computer program, which, when executed, implements the blind source separation method for multi-source fault spectrum of an oil-immersed power transformer.

[0156] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A blind source separation method for multi-source fault spectrum of oil-immersed power transformer, characterized in that The following steps are involved: Collect multi-source partial discharge fault maps, perform background denoising, grayscale processing and map compression on the collected partial discharge fault maps to obtain partial discharge fault map data with background and pixel interference removed and uniform map size. Then, pre-process the partial discharge fault map data by screening, labeling and expansion to obtain a sample data set suitable for subsequent deep learning network model training, verification and testing. Input the multi-source partial discharge fault map in the sample data set into the convolutional neural network encoder: extract the core feature vector of the multi-source partial discharge fault map through multiple convolutional layers, pooling layers, batch normalization layers and fully connected layers, and use the core feature vector and the single-source feature conditions corresponding to five types of partial discharge as conditional generative adversarial network inputs; the five types of partial discharge include tip discharge, suspension discharge, surface discharge, air gap discharge and particle discharge; A deep neural network architecture ResNet network is added to the generative network in the conditional generative adversarial network. The single-source feature vector in the partial discharge fault spectrum feature vector is extracted by combining the input single-source feature conditions. The single-source feature vector is input into the multi-layer ResNet, convolution layer, upsampling layer and activation function to realize single-source spectrum generation.

2. The blind source separation method for multi-source fault spectrum of oil-immersed power transformer according to claim 1 is characterized in that: After the single-source spectrum is generated, the PatchGAN structure in the conditional generative adversarial network is used to identify the single-source partial discharge spectrum. The identification process is: the generated single-source partial discharge spectrum is identified according to the real single-source spectrum. If the single-source partial discharge spectrum is true, the generated single-source spectrum is output, otherwise it is discarded.

3. The blind source separation method for multi-source fault spectrum of oil-immersed power transformer according to claim 2 is characterized in that: After the single-source partial discharge spectrum is identified, the single-source partial discharge spectrum is deduced and diagnosed. The process is as follows: the sample data set is used as the input of the convolutional neural network encoder and the conditional generative adversarial network, and as the input of the conditional generative adversarial network YOLOv8 network. The model training parameters are adjusted using a deep learning dedicated server to obtain a network weight data file that meets the requirements of the partial discharge fault diagnosis task of the oil-immersed power transformer; The trained YOLOv8 network is used as the diagnostic network to identify the fault type of each generated single-source graph, clarify the fault type of the multi-source graph, and deduce and diagnose the fault state of the oil-immersed power transformer; The network weight data file described in the Encoder-CGAN network is deployed to the power inspection equipment and platform to perform fault monitoring and deduction diagnosis of oil-immersed power transformers.

4. The blind source separation method for multi-source fault spectrum of oil-immersed power transformer according to claim 1 is characterized in that: The method of collecting multi-source partial discharge fault maps, performing background denoising, grayscale processing and map compression processing on the collected partial discharge fault maps to obtain partial discharge fault map data with background and pixel interference removed and uniform map size, and then screening, labeling and expanding the partial discharge fault map data to obtain a sample data set suitable for subsequent deep learning network model training, verification and testing includes the following steps: Experiments were conducted on the fault model of oil-immersed power transformers to collect single-source and multi-source partial discharge fault maps; For the collected partial discharge fault maps, background denoising is performed, and the threshold is set to filter out grid lines and sine waves, and the black background is converted into a white background; The obtained partial discharge fault map is gray-processed by using the perceptual weighting method to obtain the PRPD gray-scale map; The PRPD grayscale map is compressed and then preprocessed including screening and expansion to construct single-source and multi-source partial discharge sample data sets suitable for deep learning networks. A YOLOv8 network training dataset was constructed, and the Labelme image annotation tool was used to annotate the single-source partial discharge fault dataset. The annotated dataset was then sorted according to the YOLO dataset format.

5. The blind source separation method for multi-source fault spectrum of oil-immersed power transformer according to claim 1 is characterized in that: The multi-source partial discharge fault map in the sample data set is input into the convolutional neural network encoder: the core feature vector of the multi-source partial discharge fault map is extracted through multiple convolutional layers, pooling layers, batch normalization layers and fully connected layers, and the core feature vector and the single-source feature conditions corresponding to the five types of partial discharge are used as conditional generative adversarial network inputs; the five types of partial discharge include tip discharge, suspension discharge, surface discharge, air gap discharge and particle discharge, and the following steps are included: The partial discharge fault map is input into the convolutional neural network encoder, and the feature vector of the partial discharge fault map is extracted through multiple convolutional layers, pooling layers, batch normalization layers and fully connected layers; At the same time, the extracted feature vectors of partial discharge fault maps and the single-source feature vector conditions corresponding to the five types of partial discharge are used as the input of the conditional generative adversarial network.

6. The blind source separation method for multi-source fault spectrum of oil-immersed power transformer according to claim 1 is characterized in that: The method comprises the following steps: adding a deep neural network architecture ResNet network to the generative network in the conditional generative adversarial network, combining the input single-source feature conditions, extracting the single-source feature vector in the partial discharge fault map feature vector, and inputting the single-source feature vector into the multi-layer ResNet, convolution layer, upsampling layer and activation function to realize single-source map generation; and Add a ResNet network to the generative network in the conditional generative adversarial network; Combined with the input single-source characteristic conditions, the single-source characteristic information in the characteristic vector of the partial discharge fault spectrum is extracted; The extracted single-source feature vector is passed through multiple ResNet layers, convolutional layers, upsampling layers and activation functions to generate a single-source map.

7. The blind source separation method for multi-source fault spectrum of oil-immersed power transformer according to claim 2 is characterized by: The PatchGAN structure in the conditional generative adversarial network is used to identify the single-source partial discharge spectrum. The identification process is: the generated single-source partial discharge spectrum is identified according to the real single-source spectrum. If the single-source partial discharge spectrum is true, the generated single-source spectrum is output, otherwise it is discarded; including the following steps: The discriminator network in the conditional generative adversarial network uses the standard PatchGAN structure as a binary classifier to distinguish true from false images; The generated single-source partial discharge spectrum is identified with the real single-source spectrum. If it is true, the generated single-source spectrum is output, otherwise it is discarded, so as to achieve the purpose of blind source separation of multi-source spectrum into single-source spectrum.

8. The blind source separation method for multi-source fault spectrum of oil-immersed power transformer according to claim 1 is characterized by: In order to alleviate the gradient vanishing and mode collapse problems of the traditional conditional generative adversarial network, a gradient penalty term is introduced on the basis of the adversarial loss function to improve the stability of training, alleviate the mode collapse problem, and improve the quality of generated samples; the sub-steps are as follows: Known Generative Network G The goal is to generate high-quality single-source partial discharge spectrum samples G(z|y) , so that these samples tend to be real in the identification network, so the loss function of the generated network is expressed as: (1)； In the formula, L G is the loss value of the generated network, D To identify the network, y is the characteristic condition of single-source partial discharge, G(z|y) Give the generator a condition y and the noise vector z The generated samples, z is the characteristic vector of partial discharge fault spectrum, p z （·) is the characteristic data distribution of partial discharge fault spectrum, p(·) is the conditional feature data distribution, For p z (·) expectations; In order to avoid the problem of unstable training of conditional generative adversarial networks or adversarial imbalance between the generator and the discriminator, a gradient penalty term is introduced, which is defined as: (2)； In the formula, is generated from linear interpolation between the generated image and the real image, λ is a hyperparameter that controls the weight of the gradient penalty term. For the discriminator D On input , and given the condition y The gradient at represents the gradient; Known Identification Network D The goal is to identify the real single-source partial discharge samples and generate single-source partial discharge samples. The identification process also needs to consider the single-source partial discharge characteristic conditions. y The loss function of the identification network is usually expressed as: (3)； In the formula, L D To identify the loss value of the network, x A sample of a real single-source partial discharge spectrum. D(x|y) Give conditions to the discriminator y Next, we consider the real sample x is the true probability, p data (x,y) is the distribution of real data, For P data (x,y) expectations.

9. The blind source separation method for multi-source fault spectrum of oil-immersed power transformer according to claim 3 is characterized by: The method uses the sample data set as the input of the convolutional neural network encoder and the conditional generative adversarial network, and as the input of the YOLOv8 network, and uses a deep learning dedicated server to adjust the model training parameters to obtain a network weight data file that meets the requirements of the partial discharge fault diagnosis task of the oil-immersed power transformer; the method includes the following steps: Use the sample dataset as the input of the convolutional neural network encoder and the conditional generative adversarial network, and as the input of the YOLOv8 network Encoder-CGAN network; Use a dedicated deep learning server to train and verify the model, continuously optimize and adjust the network parameters of the Encoder-conditional generative adversarial network, and obtain the network weight data file that meets the requirements of the partial discharge fault spectrum blind source separation task; A deep learning dedicated server was used to conduct multiple training and verifications, and the network parameters of YOLOv8 were continuously optimized and adjusted to obtain the network weight data file that meets the requirements of the partial discharge fault diagnosis task of oil-immersed power transformers.

10. The blind source separation method for multi-source fault spectrum of oil-immersed power transformer according to claim 3 is characterized by: In order to accurately identify the type of the generated single-source graph, the trained YOLOv8 network model is used as the diagnostic network to identify the type of each generated single-source graph, determine the type of the original multi-source fault graph, and deduce and diagnose the fault state of the oil-immersed power transformer.

11. The blind source separation method for multi-source fault spectrum of oil-immersed power transformer according to claim 9 is characterized in that: The convolutional neural network encoder and the Encoder-CGAN network weight data files of the conditional generative adversarial network are deployed to the corresponding inspection platform to complete the fault monitoring and deduction diagnosis tasks of oil-immersed power transformers.

12. The blind source separation method for multi-source fault spectrum of oil-immersed power transformer according to claim 5 is characterized by: The partial discharge fault map is input into the convolutional neural network encoder, and the core feature vector of the partial discharge fault map is extracted through multiple convolutional layers, pooling layers, and batch normalization layers. Its characteristics are divided into the following steps: Design the structure of the convolutional neural network encoder and optimize the structures of the convolutional layer, pooling layer, and batch normalization layer; The partial discharge fault map is input into the convolutional neural network encoder, and the feature vector of the partial discharge fault map is obtained through convolution layer, pooling layer and batch normalization feature extraction.

13. The blind source separation method for multi-source fault spectrum of oil-immersed power transformer according to claim 5 is characterized by: The extracted partial discharge fault spectrum feature vector and the single-source feature conditions of five partial discharge types are used as the input of the conditional generative adversarial network; Includes the following The single-source characteristic conditions of five types of partial discharge are obtained through convolutional neural networks; Optimize the convolutional layer and pooling layer structure of the generative adversarial network to facilitate the conditional input of multi-source feature vectors and single-source feature vectors; The partial discharge fault spectrum feature vector and the single source feature conditions are input into the generative network of the conditional generative adversarial network.

14. The blind source separation method for multi-source fault spectrum of oil-immersed power transformer according to claim 5 is characterized in that: The generated single-source partial discharge spectrum will be identified according to the real single-source spectrum. If the single-source partial discharge spectrum is true, the generated single-source spectrum will be output, otherwise it will be discarded; the sub-steps are: The discriminant network of the standard PatchGAN structure is used as a binary classifier to distinguish the authenticity of the generated images; Input the constructed real single-source PD dataset map and the generated single-source map into the discrimination network; If the identification result is true, the generated single-source graph is output, otherwise it is discarded.

15. A blind source separation system for multi-source fault spectrum of oil-immersed power transformer, characterized by: The system includes: Acquisition unit: used to collect multi-source partial discharge fault maps, perform background denoising, grayscale processing and map compression on the collected partial discharge fault maps, obtain partial discharge fault map data with background and pixel interference removed and uniform map size, and then screen, annotate and expand the partial discharge fault map data to obtain a sample data set suitable for subsequent deep learning network model training, verification and testing; Processing unit: used to input the multi-source partial discharge fault map in the sample data set into the convolutional neural network encoder: extract the core feature vector of the multi-source partial discharge fault map through multiple convolution layers, pooling layers, batch normalization layers and fully connected layers, and use the core feature vector and the single-source feature conditions corresponding to five types of partial discharge as conditional generative adversarial network input; the five types of partial discharge include tip discharge, suspension discharge, surface discharge, air gap discharge and particle discharge; Single-source graph generation unit: used to add a deep neural network architecture ResNet network to the generative network in the conditional generative adversarial network, combine the input single-source feature conditions, extract the single-source feature vector in the partial discharge fault graph feature vector, and input the single-source feature vector into the multi-layer ResNet, convolution layer, upsampling layer and activation function to achieve single-source graph generation.

16. The blind source separation system for multi-source fault spectrum of oil-immersed power transformer according to claim 15, characterized in that: The system also includes: Identification unit: used to identify the single-source partial discharge spectrum using the PatchGAN structure in the conditional generative adversarial network. The identification process is: identify the generated single-source partial discharge spectrum according to the real single-source spectrum. If the single-source partial discharge spectrum is true, the generated single-source spectrum is output, otherwise it is discarded; Deduction and diagnosis unit: used to use the sample data set as the input of the convolutional neural network encoder and the conditional generative adversarial network, and as the input of the conditional generative adversarial network YOLOv8 network, and use the deep learning dedicated server to adjust the model training parameters to obtain the network weight data file that meets the requirements of the partial discharge fault diagnosis task of the oil-immersed power transformer; use the trained YOLOv8 network as the diagnosis network to identify the fault type of each generated single-source map, clarify the fault type of the multi-source map, and deduce and diagnose the fault state of the oil-immersed power transformer; deploy the network weight data file described in the Encoder-CGAN network to the power inspection equipment and platform to perform oil-immersed power transformer fault monitoring and deduction diagnosis.

17. A computer device, characterized in that: include: one or more processors; The processor is used to store one or more programs; When the one or more programs are executed by the one or more processors, the method for blind source separation of multi-source fault graphs of oil-immersed power transformers as described in any one of claims 1 to 14 is implemented.

18. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed, a blind source separation method for multi-source fault graphs of oil-immersed power transformers according to any one of claims 1-14 is implemented.

Images