Partial discharge identification method and system for switch cabinet based on wavelet spectrum and autoencoder

Abstract

The present invention provides a method and system for identifying partial discharge of a switch cabinet based on wavelet spectrum and autoencoder, and belongs to the technical field of partial discharge identification. The method includes: building a defect-free model and an actual defect model of the switch cabinet, creating a database based on the background noise signal during the operation of the defect-free model and the original partial discharge signal during the operation of each actual defect model; extracting N segmented pulse signals from a single partial discharge original signal according to a fixed time window, performing continuous wavelet transform on the N segmented pulse signals, and obtaining a two-dimensional defect time-frequency energy spectrum; performing the same processing on the background noise signal to obtain N' noise time-frequency energy spectra; inputting all defect time-frequency energy spectra and noise time-frequency energy spectra into a convolution autoencoder for model training; obtaining the partial discharge signal of the switch cabinet under test and converting it into a time-frequency energy spectrum, and analyzing the time-frequency energy spectrum based on the convolution autoencoder to obtain the defect type of the switch cabinet under test.

Description

Technical Field

[0001] The present invention relates to the technical field of partial discharge identification, and in particular to a method and system for partial discharge identification of a switch cabinet based on wavelet spectrum and autoencoder. Background Art

[0002] Switchgear is one of the main equipment in the distribution network. Improving the power supply reliability of the distribution network is inseparable from the timely and effective monitoring of the insulation status of the switchgear. Statistics show that 85% of insulation faults are caused by partial discharge. If these insulation faults cannot be prevented and monitored in time, serious power accidents will occur, causing economic losses and personal harm. The occurrence of partial discharge is accompanied by electron avalanches and repetitive high-frequency transient current pulses. The pulse width and rise time depend on the location and type of the partial discharge source. Since different types of partial discharge sources in the switchgear have different degrees of harm to the switchgear, it is necessary to monitor the switchgear in real time to see if there is a partial discharge problem, and to diagnose the type of partial discharge source without power outages for maintenance, so as to avoid excessive maintenance or underestimation of the degree of harm of the partial discharge source, resulting in failure to eliminate hidden dangers in time. Summary of the invention

[0003] In view of this, the present invention provides a switch cabinet partial discharge identification method and system based on wavelet spectrum and autoencoder, which are used to monitor in real time whether there is a partial discharge problem in the switch cabinet, understand the defect type in the switch cabinet and the location of the partial discharge source under the premise of non-power-off maintenance, and provide a basis for hazard level assessment.

[0004] The technical solution adopted by the embodiment of the present invention to solve the technical problem is:

[0005] A method for identifying partial discharge of a switch cabinet based on wavelet spectrum and autoencoder, characterized by comprising:

[0006] Step S1, building a defect-free model of the switch cabinet and a typical actual defect model in the switch cabinet, and collecting background noise signals during the operation of the defect-free model and original partial discharge signals during the operation of each of the actual defect models;

[0007] Step S2, extracting N segmented pulse signals from a single partial discharge original signal according to a fixed time window, performing continuous wavelet transform on the N segmented pulse signals, and obtaining N two-dimensional defect time-frequency energy spectra, wherein the defect time-frequency energy spectra are marked with defect categories c, where c=1, 2, 3...C; performing the same fixed time window extraction and continuous wavelet transform on the background noise signal, and obtaining N' noise time-frequency energy spectra, wherein the noise time-frequency energy spectra are marked with the defect category c, where c=0;

[0008] Step S3, creating a database, and adding the defect time-frequency energy spectrum corresponding to each of the actual defect models and the noise time-frequency energy spectrum corresponding to the defect-free model to the database;

[0009] Step S4, performing model training on a CAE-PD internal defect diagnosis model based on the defect time-frequency energy spectrum and the noise time-frequency energy spectrum in the database;

[0010] Step S5, obtaining the partial discharge signal of the switch cabinet under test and converting it into a time-frequency energy spectrum, analyzing the time-frequency energy spectrum based on the trained CAE-PD internal defect diagnosis model to obtain the defect type c of the switch cabinet under test, when the result obtained is c=C, it means that the switch cabinet under test has an internal defect of type C, and when the result obtained is c=0, it means that the switch cabinet under test does not have a defect type within the recognition range of the CAE-PD internal defect diagnosis model.

[0011] Preferably, each of the actual defect models is a model with a single internal defect, and the form of the internal defect at least includes the presence of a sharp point between the busbar and the insulator, dirt on the inner surface of the bushing, metal debris at the static contact, and a defect in the cable terminal head.

[0012] Preferably, the step S2 of extracting N segmented pulse signals from the partial discharge original signal according to a fixed time window, performing continuous wavelet transform on the N segmented pulse signals, and obtaining N two-dimensional time-frequency energy spectra comprises:

[0013] Step S21, identifying a single local discharge original signal x c The position of the local discharge pulse in the middle, the segmented pulse signal x with a duration of t is intercepted from each of the N local discharge pulses i,c , where i represents the extraction order of the segmented pulse signals, and all the segmented pulse signals x i,c The maximum value of the pulse is located at a%t, a%<0.5;

[0014] Step S22: the segmented pulse signal x i,c Perform continuous wavelet transform, the transformation formula is:

[0015]

[0016] Among them, τ is the translation parameter, α is the scale parameter, Used for energy normalization, the * sign indicates taking the conjugate complex number, g(τ,α) is the mother wavelet, and the Morse wavelet is used. The frequency domain of the Morse wavelet is defined as:

[0017]

[0018] Among them, U(ω) is a step function, a P,γ is the normalization constant, P 2 =βγ is the time-bandwidth product, γ represents the symmetry of the Morse wavelet;

[0019] Step S23, calculate each CWT i,c The mean square modulus is taken as the time-frequency energy spectrum y i,c , the time-frequency energy spectrum y i,c It is a pseudo-color RGB image;

[0020] Step S24, processing all the partial discharge original signals in sequence, and obtaining N segmented pulse signals corresponding to each of the partial discharge original signals.

[0021] Preferably, the specific steps of performing model training on the CAE-PD internal defect diagnosis model based on the defect time-frequency energy spectrum and the noise time-frequency energy spectrum in the database in step S4 include:

[0022] Step S41, building the convolutional autoencoder, the convolutional autoencoder consists of an encoder and a decoder, the encoder has an input layer and a four-layer convolutional layer structure, the decoder has a three-layer deconvolution structure, and the convolutional autoencoder uses zero padding and SELU activation function and uses the mean square error of the adam optimizer as the loss function;

[0023] Step S42, remove the decoder in the convolutional autoencoder, add a fully connected network, and form the CAE-PD internal defect diagnosis model. The overall structure of the convolutional neural network of the CAE-PD internal defect diagnosis model is set as follows: input layer → convolutional layer 1 → convolutional layer 2 → convolutional layer 3 → convolutional layer 4 → flat layer → exit layer → output layer. The number of nodes in the output layer is C, and each output node corresponds to the defect type c, where c = 0, 1, 2, 3 ..., C. The loss function used by the fully connected network is classification cross entropy and Adam optimization function.

[0024] Step S43, dividing all signals in the database into a training set, a validation set and a test set according to a preset ratio, wherein the background noise signal is in the training set;

[0025] Step S44, using the signals in the training set to train the CAE-PD internal defect diagnosis model, using the validation set to correct the CAE-PD internal defect diagnosis model, using the test set to evaluate the accuracy of the CAE-PD internal defect diagnosis model, and the training of the CAE-PD internal defect diagnosis model is completed.

[0026] Preferably, the step S5 acquires the partial discharge signal of the switch cabinet under test and converts it into a time-frequency energy spectrum, and the step of analyzing the time-frequency energy spectrum based on the trained CAE-PD internal defect diagnosis model to obtain the defect type of the switch cabinet under test includes:

[0027] Step S51, using an online detection device to measure a partial discharge signal of partial discharge in the switch cabinet under test;

[0028] Step S52, obtaining a set of time-frequency energy spectra based on the partial discharge signal of the switch cabinet under test through fixed time window extraction processing and continuous wavelet transform processing;

[0029] Step S53: analyzing the time-frequency energy spectrum based on the trained CAE-PD internal defect diagnosis model to obtain the defect type c of the switch cabinet under test.

[0030] The present invention provides a switch cabinet partial discharge identification system based on wavelet spectrum and autoencoder, comprising:

[0031] An acquisition module is used to build a defect-free model of the switch cabinet and a typical actual defect model in the switch cabinet, and to collect background noise signals during the operation of the defect-free model and original partial discharge signals during the operation of each actual defect model;

[0032] A transformation module extracts N segmented pulse signals from a single partial discharge original signal according to a fixed time window, performs continuous wavelet transformation on the N segmented pulse signals, and obtains N two-dimensional defect time-frequency energy spectra, wherein the defect time-frequency energy spectra are marked with defect categories c, where c=1, 2, 3…C; performs the same fixed time window extraction and continuous wavelet transformation on the background noise signal, and obtains N' noise time-frequency energy spectra, wherein the noise time-frequency energy spectra are marked with the defect category c, where c=0;

[0033] A creation module, used for creating a database, and adding the defect time-frequency energy spectrum corresponding to each of the actual defect models and the noise time-frequency energy spectrum corresponding to the defect-free model to the database;

[0034] A training module, performing model training on a CAE-PD internal defect diagnosis model based on the defect time-frequency energy spectrum and the noise time-frequency energy spectrum in the database;

[0035] The test module obtains the partial discharge signal of the switch cabinet under test and converts it into a time-frequency energy spectrum. Based on the trained CAE-PD internal defect diagnosis model, the time-frequency energy spectrum is analyzed to obtain the defect type c of the switch cabinet under test. When the result obtained is c=C, it means that the switch cabinet under test has an internal defect of type C. When the result obtained is c=0, it means that the switch cabinet under test does not have a defect type within the recognition range of the CAE-PD internal defect diagnosis model.

[0036] Preferably, each of the actual defect models is a model with a single internal defect, and the form of the internal defect at least includes the presence of a sharp point between the busbar and the insulator, dirt on the inner surface of the bushing, metal debris at the static contact, and a defect in the cable terminal head.

[0037] Preferably, the overall structure of the convolutional neural network of the CAE-PD internal defect diagnosis model is set as: input layer→convolution layer 1→convolution layer 2→convolution layer 3→convolution layer 4→flat layer→exit layer→output layer, the number of nodes in the output layer is C, and each output node corresponds to the defect type c, where c=0, 1, 2, 3…, C.

[0038] Preferably, the test module uses an online detection device to measure the partial discharge signal of the partial discharge in the switch cabinet under test; through fixed time window extraction processing and continuous wavelet transform processing, a set of time-frequency energy spectra are obtained based on the partial discharge signal of the switch cabinet under test; based on the trained CAE-PD internal defect diagnosis model, the time-frequency energy spectrum is analyzed to obtain the defect type c of the switch cabinet under test.

[0039] It can be seen from the above technical scheme that the switch cabinet partial discharge identification method and system based on wavelet spectrum and autoencoder provided in the embodiment of the present invention build a defect-free model of the switch cabinet and a typical actual defect model in the switch cabinet, collect the background noise signal during the operation of the defect-free model and the original partial discharge signal during the operation of each actual defect model; extract N segmented pulse signals from a single partial discharge original signal according to a fixed time window, perform continuous wavelet transform on the N segmented pulse signals, and obtain a two-dimensional defect time-frequency energy spectrum; perform the same fixed time window extraction and continuous wavelet transform on the background noise signal to obtain N' noise time-frequency energy spectra; create a database, and add the defect time-frequency energy spectrum corresponding to each actual defect model and the noise time-frequency energy spectrum corresponding to the defect-free model to the database; perform model training on the CAE-PD internal defect diagnosis model based on the defect time-frequency energy spectrum and the noise time-frequency energy spectrum in the database; obtain the partial discharge signal of the switch cabinet under test and convert it into a time-frequency energy spectrum, and obtain the defect type of the switch cabinet under test based on the analyzed time-frequency energy spectrum after training. The method of the present invention can be used to monitor in real time whether a switch cabinet has a partial discharge problem, and to understand the defect type and partial discharge source in the switch cabinet without power outages for maintenance, thereby providing a basis for hazard assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a flow chart of the switchgear partial discharge identification method based on wavelet spectrum and autoencoder.

[0041] Figure 2 This is an example diagram of the actual defect model design of the constructed switch cabinet.

[0042] Figure 3 It is a schematic diagram of the database signal acquisition and processing process in the present invention.

[0043] Figure 4 This is a simplified model of the CAE structure in the present invention.

[0044] Figure 5 Schematic diagram of the CAE-PD internal defect diagnosis model training results.

[0045] Figure 6 This is an example diagram of the classification and identification results of the CAE-PD internal defect diagnosis model. DETAILED DESCRIPTION

[0046] The technical scheme and technical effects of the present invention are further elaborated in detail below in conjunction with the accompanying drawings of the present invention.

[0047] The present invention provides a switch cabinet partial discharge identification method based on wavelet spectrum and autoencoder. First, a database is constructed, and then a convolutional autoencoder classification model is trained using the database. When identifying a signal, the model is used to classify and judge the pre-processed signal. Figure 1 As shown, the steps include:

[0048] Step S1, building a defect-free model of the switch cabinet and a typical actual defect model in the switch cabinet, collecting background noise signals during the operation of the defect-free model and original partial discharge signals during the operation of each actual defect model;

[0049] Step S2, extracting N segmented pulse signals from a single partial discharge original signal according to a fixed time window, performing continuous wavelet transform on the N segmented pulse signals, and obtaining N two-dimensional defect time-frequency energy spectra, wherein the defect time-frequency energy spectra are marked with defect categories c, where c=1,2,3…C; performing the same fixed time window extraction and continuous wavelet transform on the background noise signal, and obtaining N' noise time-frequency energy spectra, wherein the noise time-frequency energy spectra are marked with defect categories c, where c=0;

[0050] Step S3, creating a database, adding the defect time-frequency energy spectrum corresponding to each actual defect model and the noise time-frequency energy spectrum corresponding to the defect-free model to the database;

[0051] Step S4, training the CAE-PD internal defect diagnosis model based on the defect time-frequency energy spectrum and noise time-frequency energy spectrum in the database;

[0052] Step S5, obtaining the partial discharge signal of the switch cabinet under test and converting it into a time-frequency energy spectrum, analyzing the time-frequency energy spectrum based on the trained CAE-PD internal defect diagnosis model to obtain the defect type c of the switch cabinet under test, when the result obtained is c=C, it means that the switch cabinet under test has an internal defect of type C, and when the result obtained is c=0, it means that the switch cabinet under test does not have a defect type within the recognition range of the CAE-PD internal defect diagnosis model.

[0053] Each actual defect model is a model with a single internal defect, such as Figure 2 As shown, the forms of internal defects include:

[0054] (1) There are sharp points between the busbar and the insulator: The busbar of the actual switch cabinet may have loose screws, cracked insulation, burrs, etc., resulting in insulation defects between the busbar and the insulator;

[0055] (2) The bushing surface is dirty: Due to the humidity of the environment and problems in the assembly process, the actual switch cabinet may have condensation and dirt, resulting in surface discharge;

[0056] (3) Metal debris exists at the static contact: Suspended discharge is prone to occur at the static contact of the actual switch cabinet, so the static contact of the design has metal debris defects;

[0057] (4) Cable terminal defects: The cable terminal in the actual switch cabinet is prone to defects, mainly because the process of the cable terminal does not meet the requirements during the manufacturing process. Therefore, in this example, conductive particles and cut marks are set at the cable terminal to simulate the insulation defects of the cable terminal caused by improper operation in actual production.

[0058] Step S2 extracts N segmented pulse signals from the partial discharge original signal according to a fixed time window, performs continuous wavelet transform on the N segmented pulse signals, and obtains N two-dimensional time-frequency energy spectra. The implementation steps include:

[0059] Step S21, identifying a single partial discharge original signal x c The position of the local discharge pulse in the middle, the segmented pulse signal x with a duration of t is intercepted from each of the N local discharge pulses i,c , where i represents the extraction order of the segmented pulse signal, all segmented pulse signals x i,c The maximum value of the pulse is located at a%t, a%<0.5;

[0060] Step S22: segment pulse signal x i,c Perform continuous wavelet transform, the transformation formula is:

[0061]

[0062] Among them, τ is the translation parameter, α is the scale parameter, Used for energy normalization, the * sign indicates taking the conjugate complex number. g(τ,α) is the mother wavelet, using the Morse wavelet, the frequency domain of the Morse wavelet is defined as:

[0063]

[0064] Among them, U(ω) is a step function, a P,γ is the normalization constant, P 2 =βγ is the time-bandwidth product, γ represents the symmetry of the Morse wavelet, Morse (3,60) means γ = 3, P 2 =60.

[0065] Step S23, calculate each CWT i,c The mean square modulus is taken as the time-frequency energy spectrum y i,c , time-frequency energy spectrum y i,c It is a pseudo-color RGB image;

[0066] Step S24, processing all partial discharge original signals in sequence, and obtaining N segmented pulse signals corresponding to each partial discharge original signal.

[0067] like Figure 3In the example shown, for each string of signals, the position of the partial discharge pulse is identified, the time window is defined as 6.4 μs, and a segmented pulse signal of 6.4 μs is intercepted at each pulse. There are 512 data points in each time window. The position of the pulse maximum value is automatically adjusted according to the signal characteristics, so that the pulse maximum value of each segmented pulse signal is all located at 1.25 μs. According to the original signal and extraction order of each segmented pulse signal, each segmented pulse signal intercepted is named x i,c , where c represents the discharge source type or the defect type of the response, i is the pulse number, i = 0, 1, 2, ..., N, representing the i-th pulse of this type; the wavelet used to calculate the CWT filter bank is morse (3, 60), and each segmented pulse signal x is calculated i,c The mean square modulus of the CWT is used as the time-frequency energy spectrum and converted into a pseudo-color RGB image. Each image is an array of size 224×224×3. Each image is also marked as y according to the above method. i,c .

[0068] Furthermore, the specific steps of step S4 for training the CAE-PD internal defect diagnosis model based on the defect time-frequency energy spectrum and the noise time-frequency energy spectrum in the database include:

[0069] Step S41, building a convolutional autoencoder and confirming its validity, wherein the convolutional autoencoder is composed of an encoder and a decoder, the encoder has an input layer and a four-layer convolutional layer structure, the decoder has a three-layer deconvolution structure, and the convolutional autoencoder uses zero padding and SELU activation function, and uses the mean square error of the adam optimizer as the loss function;

[0070] Step S42, remove the decoder in the convolutional autoencoder, add a fully connected network, and form a CAE-PD internal defect diagnosis model. The overall structure of the convolutional neural network of the CAE-PD internal defect diagnosis model is set as follows: input layer → convolutional layer 1 → convolutional layer 2 → convolutional layer 3 → convolutional layer 4 → flat layer → exit layer → output layer. The number of nodes in the output layer is C, and each output node corresponds to a defect type c, where c = 0, 1, 2, 3 ..., C. The loss function used by the fully connected network is classification cross entropy and Adam optimization function.

[0071] Step S43, dividing all signals in the database into a training set, a validation set and a test set according to a preset ratio;

[0072] Step S44, use the signals in the training set to train the CAE-PD internal defect diagnosis model, use the validation set to correct the CAE-PD internal defect diagnosis model, use the test set to evaluate the accuracy of the CAE-PD internal defect diagnosis model, and the training of the CAE-PD internal defect diagnosis model is completed.

[0073] like Figure 4 As shown, there are two options for the convolutional autoencoder structure, and their structures are shown in the figure as structure 1 and structure 2. Structure 1 uses a pooling layer plus a convolutional layer in the encoder, and uses upsampling plus a convolutional layer in the decoder; while structure 2 uses a convolutional layer in the encoder and a deconvolutional layer in the decoder. Considering that the deeper the number of layers of the neural network, the better the fitting ability, but too deep will cause overfitting and increase the difficulty of training, and the same is true for the selection of the number of nodes in each layer. In this example, after several attempts, it can be found that four layers of convolution are used in the encoder and three layers of convolution are used in the decoder to achieve a good image restoration effect. In order to compare the performance of these two structures, the number of parameters to be trained is roughly the same. By comparison, the recognition accuracy of structure 2 is not much different from that of structure 1, and the required training time is halved, so the present invention finally chooses to use the convolutional autoencoder in structure 2.

[0074] Therefore, the model training process in this example includes: First, unsupervised training of the convolutional autoencoder model: using Figure 4 The structure of the second convolutional autoencoder model is input with unlabeled images for training, and the decoder outputs the reconstructed image to confirm that the encoder of the current convolutional autoencoder model is valid. Zero padding and SELU activation functions are used, and the mean square error of the adam optimizer is used as the loss function. Second, the encoder is connected to the deep neural network to form a classifier for supervised training: the decoder in the model built in step (1) is removed from the network, and the output of the encoder is connected to the fully connected network. While keeping the encoder part unchanged, the time-frequency energy spectrum with classification labels in the database is input. The structure of the fully connected network is a flat layer for converting a two-dimensional tensor into a one-dimensional tensor, a dropout layer with a dropout rate of 50%, a fully connected layer, and an output layer corresponding to the number of classifications. The loss functions used are classification cross entropy and Adam optimization. 80% of the images in the dataset are used as training sets, 10% are used for verification, and 10% are used as test sets.

[0075] Preferably, step S5 acquires the partial discharge signal of the switch cabinet under test and converts it into a time-frequency energy spectrum, and the step of analyzing the time-frequency energy spectrum based on the trained CAE-PD internal defect diagnosis model to obtain the defect type of the switch cabinet under test includes:

[0076] Step S51, using an online detection device to measure a partial discharge signal of partial discharge in the switch cabinet under test;

[0077] Step S52, obtaining a set of time-frequency energy spectra based on the partial discharge signal of the switch cabinet under test through fixed time window extraction processing and continuous wavelet transform processing;

[0078] Step S53: Analyze the time-frequency energy spectrum based on the trained CAE-PD internal defect diagnosis model to obtain the defect type c of the switch cabinet under test.

[0079] like Figure 5 The model training results shown in the figure show that for the time-frequency energy spectrum in the upper left figure, the trained convolutional autoencoder model extracts the features of the original image in the encoder and reconstructs the image with the decoder to obtain the upper right figure. Through the CAE-PD internal defect diagnosis model of the present invention, the lower left figure shows the changes in the training and verification accuracy during its training process, and the lower right figure shows the confusion matrix of the CAE-PD internal defect diagnosis model of the present invention. It can be seen that each image to be identified has been correctly classified, verifying the accuracy of the CAE structure selected by the present invention.

[0080] like Figure 6 As shown in the figure, the newly received signal is classified using the trained model. The classification result is shown in the figure. As can be seen from the right figure, the recognition accuracy is still about 87% when there is a certain amount of background noise.

[0081] Furthermore, the present invention provides a switch cabinet partial discharge identification system based on wavelet spectrum and autoencoder, which is used to implement Figure 1 The switch cabinet partial discharge identification method shown in the figure includes:

[0082] The acquisition module builds a defect-free model of the switch cabinet and a typical actual defect model in the switch cabinet, and collects the background noise signal during the operation of the defect-free model and the original partial discharge signal during the operation of each actual defect model;

[0083] The transformation module extracts N segmented pulse signals from a single partial discharge original signal according to a fixed time window, performs continuous wavelet transformation on the N segmented pulse signals, and obtains N two-dimensional defect time-frequency energy spectra, which are marked with defect categories c, where c = 1, 2, 3 ... C; performs the same fixed time window extraction and continuous wavelet transformation on the background noise signal, and obtains N' noise time-frequency energy spectra, which are marked with defect categories c, where c = 0;

[0084] A creation module is used to create a database, and add the defect time-frequency energy spectrum corresponding to each actual defect model and the noise time-frequency energy spectrum corresponding to the defect-free model to the database;

[0085] The training module trains the CAE-PD internal defect diagnosis model based on the defect time-frequency energy spectrum and noise time-frequency energy spectrum in the database;

[0086] The test module obtains the partial discharge signal of the switch cabinet under test and converts it into a time-frequency energy spectrum. Based on the trained CAE-PD internal defect diagnosis model, the time-frequency energy spectrum is analyzed to obtain the defect type c of the switch cabinet under test. When the result is c=C, it means that the switch cabinet under test has an internal defect of type C. When the result is c=0, it means that the switch cabinet under test does not have a defect type within the recognition range of the CAE-PD internal defect diagnosis model.

[0087] The test module uses an online detection device to measure the partial discharge signal of the partial discharge in the switch cabinet under test; through fixed time window extraction processing and continuous wavelet transform processing, a set of time-frequency energy spectra are obtained based on the partial discharge signal of the switch cabinet under test; based on the trained CAE-PD internal defect diagnosis model, the time-frequency energy spectrum is analyzed to obtain the defect type c of the switch cabinet under test.

[0088] Among them, each actual defect model is a model with a single internal defect. The form of the internal defect can be referred to Figure 2 The examples shown include at least the presence of sharp points between the busbar and the insulator, dirt on the inner surface of the bushing, metal debris at the stationary contact, and defects in the cable terminal head.

[0089] Preferably, if Figure 3 As shown in the figure, the overall structure of the convolutional neural network of the CAE-PD internal defect diagnosis model is set as: input layer → convolution layer 1 → convolution layer 2 → convolution layer 3 → convolution layer 4 → flat layer → exit layer → output layer, where the number of nodes in the output layer is C, and each output node corresponds to a defect type c, where c = 0, 1, 2, 3…, C.

[0090] The present invention provides a method and system for identifying partial discharge of a switch cabinet based on wavelet spectrum and autoencoder, which collects partial discharge pulse signals from typical partial discharge sources in actual switch cabinets, extracts each partial discharge pulse in each type of partial discharge, and converts it into a time-frequency energy spectrum through continuous wavelet transform, and then stores it in a database in the form of a picture; then, a convolutional autoencoder is used to extract the features of the picture in the database, and then the encoder is connected to a convolutional fully connected network to form a classification model, and the classification model is trained using a picture database; finally, when partial discharge diagnosis is carried out on a new switch cabinet, the trained classification model is used to perform diagnostic classification on the new detection data. The present invention simplifies the structure of the convolutional autoencoder, shortens the model training time while ensuring the recognition accuracy, can accurately identify the type of partial discharge defect source under interference conditions, and is suitable for on-site partial discharge diagnosis of high-voltage switch cabinets in power grids.

[0091] The present invention has at least the following advantages: to ensure the effectiveness of the classification model, the data used to train and test the convolutional autoencoder model is generated by typical real defects found in the switch device, so that the trained model is closer to reality; the present invention uses the continuous wavelet transform method (CWT) to calculate the time-frequency energy spectrum of the partial discharge signal as the database of the classification and recognition model. Due to the advantages of CWT in analyzing signals of different frequencies and different resolutions, it has the advantage of being almost unaffected by noise interference.

[0092] What is disclosed above is only a preferred embodiment of the present invention, which certainly cannot be used to limit the scope of rights of the present invention. A person skilled in the art can understand that all or part of the processes of the above embodiments and equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A switch cabinet partial discharge identification method based on wavelet spectrum and autoencoder, characterized in that: include: Step S1, building a defect-free model of the switch cabinet and a typical actual defect model in the switch cabinet, and collecting background noise signals during the operation of the defect-free model and original partial discharge signals during the operation of each of the actual defect models; Step S2, extracting N segmented pulse signals from a single partial discharge original signal according to a fixed time window, performing continuous wavelet transform on the N segmented pulse signals, and obtaining N two-dimensional defect time-frequency energy spectra, wherein the defect time-frequency energy spectra are marked with defect categories c, where c=1, 2, 3...C; performing the same fixed time window extraction and continuous wavelet transform on the background noise signal, and obtaining N' noise time-frequency energy spectra, wherein the noise time-frequency energy spectra are marked with the defect category c, where c=0; Step S3, creating a database, and adding the defect time-frequency energy spectrum corresponding to each of the actual defect models and the noise time-frequency energy spectrum corresponding to the defect-free model to the database; Step S4, performing model training on a CAE-PD internal defect diagnosis model based on the defect time-frequency energy spectrum and the noise time-frequency energy spectrum in the database; Step S5, obtaining the partial discharge signal of the switch cabinet under test and converting it into a time-frequency energy spectrum, analyzing the time-frequency energy spectrum based on the trained CAE-PD internal defect diagnosis model to obtain the defect type c of the switch cabinet under test, when the result obtained is c=C, it means that the switch cabinet under test has an internal defect of type C, and when the result obtained is c=0, it means that the switch cabinet under test does not have a defect type within the recognition range of the CAE-PD internal defect diagnosis model; The specific steps of performing model training on the CAE-PD internal defect diagnosis model based on the defect time-frequency energy spectrum and the noise time-frequency energy spectrum in the database in step S4 include: Step S41, building a convolutional autoencoder, the convolutional autoencoder consists of an encoder and a decoder, the encoder has an input layer and a four-layer convolutional layer structure, the decoder has a three-layer deconvolution structure, the convolutional autoencoder uses zero padding and SELU activation function, and uses the mean square error of the adam optimizer as the loss function; Step S42, remove the decoder in the convolutional autoencoder, add a fully connected network, and form the CAE-PD internal defect diagnosis model. The overall structure of the convolutional neural network of the CAE-PD internal defect diagnosis model is set as follows: input layer → convolutional layer 1 → convolutional layer 2 → convolutional layer 3 → convolutional layer 4 → flat layer → exit layer → output layer. The number of nodes in the output layer is C, and each output node corresponds to the defect type c, where c = 0, 1, 2, 3 ..., C. The loss function used by the fully connected network is classification cross entropy and Adam optimization function. Step S43, dividing all signals in the database into a training set, a validation set and a test set according to a preset ratio; Step S44, using the signals in the training set to train the CAE-PD internal defect diagnosis model, using the validation set to correct the CAE-PD internal defect diagnosis model, using the test set to evaluate the accuracy of the CAE-PD internal defect diagnosis model, and the training of the CAE-PD internal defect diagnosis model is completed.

2. The switch cabinet partial discharge identification method based on wavelet spectrum and autoencoder according to claim 1 is characterized in that: Each of the actual defect models is a model with a single internal defect, and the form of the internal defect at least includes the presence of a sharp point between the busbar and the insulator, dirt on the inner surface of the bushing, metal debris at the static contact, and a defect in the cable terminal head.

3. The switch cabinet partial discharge identification method based on wavelet spectrum and autoencoder according to claim 2 is characterized in that: The step S2 extracts N segmented pulse signals from the partial discharge original signal according to a fixed time window, and performs continuous wavelet transform on the N segmented pulse signals to obtain N two-dimensional time-frequency energy spectra, including: Step S21, identifying a single local discharge original signal x c The position of the local discharge pulse in the middle, the segmented pulse signal x with a duration of t is intercepted from each of the N local discharge pulses i,c , where i represents the extraction order of the segmented pulse signals, and all the segmented pulse signals x i,c The maximum value of the pulse is located at a%t, a%<0.5; Step S22: the segmented pulse signal x i,c Perform continuous wavelet transform, the transformation formula is: Among them, τ is the translation parameter, α is the scale parameter, Used for energy normalization, the * sign indicates taking the conjugate complex number, g(τ,α) is the mother wavelet, and the Morse wavelet is used. The frequency domain of the Morse wavelet is defined as: Among them, U(ω) is a step function, a P,γ is the normalization constant, P 2 =βγ is the time-bandwidth product, γ represents the symmetry of the Morse wavelet; Step S23, calculate each CWT i,c The mean square modulus is taken as the time-frequency energy spectrum y i,c , the time-frequency energy spectrum y i,c It is a pseudo-color RGB image; Step S24, processing all the partial discharge original signals in sequence, and obtaining N segmented pulse signals corresponding to each of the partial discharge original signals.

4. The switch cabinet partial discharge identification method based on wavelet spectrum and autoencoder according to claim 3 is characterized in that: The step S5 of acquiring the partial discharge signal of the switch cabinet under test and converting it into a time-frequency energy spectrum, and analyzing the time-frequency energy spectrum based on the trained CAE-PD internal defect diagnosis model to obtain the defect type of the switch cabinet under test includes: Step S51, using an online detection device to measure a partial discharge signal of partial discharge in the switch cabinet under test; Step S52, obtaining a set of time-frequency energy spectra based on the partial discharge signal of the switch cabinet under test through fixed time window extraction processing and continuous wavelet transform processing; Step S53: analyzing the time-frequency energy spectrum based on the trained CAE-PD internal defect diagnosis model to obtain the defect type c of the switch cabinet under test.

5. A switch cabinet partial discharge identification system based on wavelet spectrum and autoencoder, characterized in that: include: An acquisition module is used to build a defect-free model of the switch cabinet and a typical actual defect model in the switch cabinet, and to collect background noise signals during the operation of the defect-free model and original partial discharge signals during the operation of each actual defect model; A transformation module extracts N segmented pulse signals from a single partial discharge original signal according to a fixed time window, performs continuous wavelet transformation on the N segmented pulse signals, and obtains N two-dimensional defect time-frequency energy spectra, wherein the defect time-frequency energy spectra are marked with defect categories c, where c=1, 2, 3…C; performs the same fixed time window extraction and continuous wavelet transformation on the background noise signal, and obtains N' noise time-frequency energy spectra, wherein the noise time-frequency energy spectra are marked with the defect category c, where c=0; A creation module, used for creating a database, and adding the defect time-frequency energy spectrum corresponding to each of the actual defect models and the noise time-frequency energy spectrum corresponding to the defect-free model to the database; A training module, performing model training on a CAE-PD internal defect diagnosis model based on the defect time-frequency energy spectrum and the noise time-frequency energy spectrum in the database; The test module obtains the partial discharge signal of the switch cabinet under test and converts it into a time-frequency energy spectrum, and analyzes the time-frequency energy spectrum based on the trained CAE-PD internal defect diagnosis model to obtain the defect type c of the switch cabinet under test. When the result obtained is c=C, it means that the switch cabinet under test has an internal defect of type C. When the result obtained is c=0, it means that the switch cabinet under test does not have a defect type within the recognition range of the CAE-PD internal defect diagnosis model. The training module is used to: Build a convolutional autoencoder, which consists of an encoder and a decoder. The encoder has an input layer and a four-layer convolutional layer structure, and the decoder has a three-layer deconvolution structure. The convolutional autoencoder uses zero padding and SELU activation functions and uses the mean square error of the adam optimizer as the loss function; The decoder in the convolutional autoencoder is removed, and a fully connected network is added to form the CAE-PD internal defect diagnosis model. The overall structure of the convolutional neural network of the CAE-PD internal defect diagnosis model is set as follows: input layer→convolutional layer 1→convolutional layer 2→convolutional layer 3→convolutional layer 4→flat layer→exit layer→output layer. The number of nodes in the output layer is C, and each output node corresponds to the defect type c, where c=0, 1, 2, 3…, C. The loss function used by the fully connected network is classification cross entropy and Adam optimization function. Dividing all signals in the database into a training set, a validation set and a test set according to a preset ratio; The CAE-PD internal defect diagnosis model is trained using the signals in the training set, the CAE-PD internal defect diagnosis model is corrected using the validation set, and the accuracy of the CAE-PD internal defect diagnosis model is evaluated using the test set. The training of the CAE-PD internal defect diagnosis model is completed.

6. The switch cabinet partial discharge identification system based on wavelet spectrum and autoencoder according to claim 5, characterized in that: Each of the actual defect models is a model with a single internal defect, and the form of the internal defect at least includes the presence of a sharp point between the busbar and the insulator, dirt on the inner surface of the bushing, metal debris at the static contact, and a defect in the cable terminal head.

7. The switch cabinet partial discharge identification system based on wavelet spectrum and autoencoder according to claim 6, characterized in that: The overall structure of the convolutional neural network of the CAE-PD internal defect diagnosis model is set as: input layer→convolution layer 1→convolution layer 2→convolution layer 3→convolution layer 4→flat layer→exit layer→output layer, the number of nodes in the output layer is C, and each output node corresponds to the defect type c, where c=0, 1, 2, 3…, C.

8. The switch cabinet partial discharge identification system based on wavelet spectrum and autoencoder according to claim 7, characterized in that: The test module uses an online detection device to measure the partial discharge signal of the partial discharge in the switch cabinet under test; through fixed time window extraction processing and continuous wavelet transform processing, a set of time-frequency energy spectra corresponding to the partial discharge signal of the switch cabinet under test is obtained; based on the CAE-PD internal defect diagnosis model after training, the time-frequency energy spectrum is analyzed to obtain the defect type c of the switch cabinet under test.

Images