Fault identification method based on time domain waveform image of partial discharge of GIL metal particles

Abstract

A fault identification method based on the time-domain waveform image of partial discharge of GIL metal particles includes the following steps: collecting historical images; preprocessing the historical images and dividing them into training sets and test sets; constructing a partial discharge time-domain waveform image recognition model and training it according to the sample set to determine the classifier model; inputting the test set into the classifier model, and optimizing it based on TensorFlow deep learning to obtain the final classifier model; collecting real-time images to determine the fault type. The present invention realizes the selection of a support vector machine model for the image containing only partial discharge pulse voltage obtained by preprocessing, using radial basis function as the kernel function, and training a DAG classifier through a deep learning algorithm, directly performing pattern recognition without the need for manual feature extraction, having a higher recognition rate, and providing favorable support for the monitoring of metal particle discharge faults inside GIL equipment.

Description

Technical Field

[0001] The present invention relates to the field of numerical calculation and analysis of electric power equipment, and in particular to a fault identification method based on time-domain waveform images of partial discharge of GIL metal particles. Background Art

[0002] Gas insulated transmission lines (GIL) are widely used in power systems. However, the generation of metal particles is inevitable during the production, assembly and operation of GIL equipment. For example, metal shavings remaining from poor internal cleaning during the production process, debris generated by extrusion and friction during transportation, mechanical friction caused by contact plugging during operation, and metal shavings and particles generated by spark discharge and ablation. The generated metal particles will be induced to be charged in the electric field, and move between electrodes to cause discharge, destroying the insulation performance of the GIL equipment. Therefore, in view of the above situation, it is of great significance to identify the internal discharge mode of the GIL equipment for its subsequent fault analysis and maintenance.

[0003] Traditional partial discharge pattern recognition requires the use of partial discharge live detection to obtain partial discharge data based on time or phase, and artificially extract statistical features. The existing stored partial discharge data mostly exist in the form of images. For instruments commonly used for on-site partial discharge detection, such as oscilloscopes, due to the limitations of storage speed and storage depth, the data generally stored in on-site detection are mostly time domain waveform image data. Traditional partial discharge pattern recognition methods cannot be applied to this type of unstructured data. By establishing an image data set of partial discharge defect types, for the images obtained by preprocessing that only contain partial discharge pulse voltage, a support vector machine model is selected, and the radial basis function is used as the kernel function. The directed acyclic graph (DAG) classifier is trained by the SMO optimization method to directly perform pattern recognition, which can provide the necessary data support for the recognition of metal particle discharge images of GIL equipment, thereby truly preventing the metal particle discharge phenomenon of GIL equipment from happening. Summary of the invention

[0004] In view of the problem in the background technology that the traditional partial discharge pattern recognition method cannot be applied to unstructured data such as partial discharge of metal particles in GIL equipment, the present invention provides a fault recognition method based on the time domain waveform image of partial discharge of GIL metal particles.

[0005] In order to solve the above technical problems, the present invention specifically adopts the following technical solutions:

[0006] A fault identification method based on a time domain waveform image of a partial discharge of GIL metal particles is characterized in that the method comprises the following steps:

[0007] Step 1: Collect the time domain waveform history images of GIL metal particle partial discharge under different fault types;

[0008] Step 2: preprocessing the time domain waveform history image of the partial discharge of metal particles inside the GIL in step 1 to construct a training data sample set and a test sample data set, wherein the preprocessing refers to performing image enhancement processing on the time domain waveform image of the partial discharge of metal particles inside the GIL using a histogram equalization algorithm;

[0009] Step 3: Use the SVM binary classifier to train the preprocessed training data sample set to determine the classifier model;

[0010] Step 4: Input the test data sample set in step 2 into the classifier model determined after training in step 3, optimize each node in the classifier, and obtain the final DAG classifier model;

[0011] Step 5: Collect the time domain waveform image of the partial discharge of the GIL metal particles in real time, input it into the classifier model obtained in step 4, and output the corresponding fault identification result.

[0012] The present invention further adopts the following preferred technical solutions:

[0013] In step 1, the GIL metal particle partial discharge fault types include suspension discharge fault, surface discharge fault, air gap discharge fault, tip discharge fault and other interferences.

[0014] The step 2 comprises the following steps:

[0015] Step 201: establishing an image histogram according to the grayscale distribution of the image, wherein the abscissa and ordinate represent the brightness range of each pixel in the image and the variation of each grayscale in the whole image respectively;

[0016] Step 202: Assuming that the original image contains N gray levels, the distribution p(r k ):

[0017]

[0018] Among them, n is the sum of the number of pixels in the entire image, n k Represents the number of pixels distributed at the kth gray level, k = 0, 1, 2, L N-1;

[0019] Step 203: discretize the distribution function of step 202 to obtain the discretized cumulative distribution function:

[0020]

[0021] The above formula means that the gray level is r k The pixels are transformed into grayscale S k Pixels, T(r k ) is the transformation function, P r (r i ) indicates the gray level is r i The probability of a pixel appearing in the image, k=0,1,2,L N-1.

[0022] In step 3, according to the number n of different fault types set in step 1, n(n-1) / 2 SVM binary classifiers are constructed, and all binary classifiers form a binary tree structure. Each node of the binary tree structure corresponds to a binary classifier, and each leaf represents a fault category.

[0023] The SVM binary classifier is used to perform pairwise training on the training data sample sets corresponding to different faults to obtain the trained classifier model.

[0024] In the binary tree structure, the SVM binary classifier with the highest accuracy after training is used as the root node of the binary tree structure, and the SVM binary classifiers with the second highest accuracy are used as the lower nodes in turn.

[0025] When the SVM binary classifier is used to train the training data sample set, the input historical image is convolved using the following formula:

[0026]

[0027] Where: f( ) represents the activation function; F conv2D represents the two-dimensional convolution formula; b j Indicates the bias value, Represents the bias value of convolution layer l; ω ij represents the weights between neurons; represents the net activation of the jth channel of the convolutional layer l, which is obtained by filtering the feature map output by the previous layer It is obtained after convolution summation and bias; i and k represent the number of accumulated eigenvalues ​​respectively.

[0028] In step 4, the test data sample set is subjected to one-dimensional data normalization processing by the following formula:

[0029]

[0030] Among them, x * is the one-dimensional result of the data; x is the average grayscale value of the sample data; x1 represents the minimum grayscale value of the sample data; x2 represents the maximum grayscale value of the sample data.

[0031] The step 4 comprises the following steps:

[0032] Step 401: constructing a training model by stacking multiple layers of restricted Boltzmann machine networks and one layer of BP neural network;

[0033] Step 402: Divide the processed data into a training set and a test set, and input the training set into the bottom-level restricted Boltzmann machine network layer of the training model;

[0034] Step 403: Start unsupervised pre-training, randomly initialize the parameters ε=(a, b, ω) of the restricted Boltzmann machine network, the weight ω, the visible layer bias b and the hidden layer bias a, set the number of layer nodes of the network, and the maximum number of layers m;

[0035] Step 404: The probability of neurons in the hidden layer and the visible layer of a restricted Boltzmann machine network layer being activated is calculated by the following formula, and if the calculated result is greater than a preset value, step 405 is executed, otherwise the training of the restricted Boltzmann machine network layer is terminated:

[0036]

[0037] Among them, h j represents the state of the jth hidden layer neuron, v j represents the state of the jth visual layer neuron; v represents the state of the visual layer neuron, h represents the state of the hidden layer neuron; b j represents the bias of the jth visual layer, a j represents the bias of the jth hidden layer, ω ij is the weight between neurons;

[0038] Step 405: Update the network weights and the biases of the visible layer and the hidden layer using the following formula:

[0039]

[0040] Where: Δ[ω ij ] n+1 is the updated weight from the i-th visual layer neuron to the j-th hidden layer neuron, n represents the n-th iteration, λ is the momentum factor, η is the learning factor, and v i represents the state of the i-th visual layer neuron, h j represents the state of the jth hidden layer neuron, (v i h j ) d represents the expected data distribution, (v i h j ) m represents the expectation of the model distribution; Δ[b j ] n+1 is the updated bias of the jth visual layer, Δ[a i ]n+1 is the bias of the i-th hidden layer after update;

[0041] Step 406: Repeat steps 404-405 until all restricted Boltzmann machine network layers are trained, use the top BP network layer to perform reverse parameter optimization, and output the trained DAG model.

[0042] The present invention has the following beneficial effects:

[0043] The present invention realizes the image containing only partial discharge pulse voltage obtained by preprocessing, selects the support vector machine model, uses the radial basis function as the kernel function, and trains a directed acyclic graph (DAG) classifier through a deep learning algorithm to directly perform pattern recognition without manual feature extraction, with a higher recognition rate, which provides favorable support for the internal metal particle discharge fault monitoring of GIL equipment BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flow chart of the present invention.

[0045] Figure 2 This is the structural diagram of the time domain waveform image measurement system.

[0046] Figure 3 This is the bifurcation diagram of the DAG training process. DETAILED DESCRIPTION

[0047] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0048] like Figure 1 As shown, a fault identification method based on the time domain waveform image of GIL metal particle partial discharge of the present invention comprises the following steps:

[0049] Step 1: Collect the time domain waveform history images of GIL metal particle partial discharge under different fault types. Specifically, in one embodiment of the present invention, the faults of GIL metal particle partial discharge are divided into suspension discharge fault, surface discharge fault, air gap discharge fault, tip discharge fault and other interference. Figure 2 As shown in the figure, the UHF detection system of GIL equipment is usually composed of UHF sensors, signal conditioning units, data acquisition cards, control units, etc. The detection uses a UHF-based substation partial discharge live detection system. The detection host is a high-speed acquisition oscilloscope, the system simulation bandwidth is 1-2GHz, the oscilloscope sampling rate is 4GHz, and the cache is 8M. When the oscilloscope stores the original waveform signal data file, the file is larger than its storage depth, resulting in incomplete data and slow storage speed. Therefore, the on-site detection is saved as a continuous multi-cycle time domain waveform image.

[0050] Step 2: Preprocess the time domain waveform history image of the partial discharge of metal particles inside the GIL in step 1 to construct a training data sample set and a test sample data set, wherein the preprocessing refers to using a histogram equalization algorithm to perform image enhancement processing on the time domain waveform image of the partial discharge of metal particles inside the GIL.

[0051] Specifically, step 2 includes the following steps:

[0052] Step 201: A discrete function is formulated according to the distribution law of each gray level of the image, and then the image is enhanced. In two-dimensional space, it can be represented by a histogram in a Cartesian coordinate system, where the horizontal and vertical axes represent the brightness range of each pixel in the image and the change of each gray level in the entire image.

[0053] Step 202: Assuming that the original image contains N gray levels, the distribution of gray levels in the image histogram p(r k ) can be determined by the following transformation:

[0054]

[0055] Where n is the sum of the number of pixels in the entire image, n k Represents the number of pixels distributed at the kth gray level.

[0056] Step 203: The discretized cumulative distribution function is:

[0057]

[0058] The above formula means that the gray level is r k The pixels are transformed into grayscale S k Pixels, T(r k ) is the transformation function, P r (r i ) indicates the gray level is r k The probability of a pixel appearing in the image.

[0059] When the grayscale values ​​are concentrated in a certain area, it means that the image contrast is small. At this time, the distribution of each grayscale level of the image can be uniformly adjusted through histogram equalization, thereby achieving the effect of image enhancement.

[0060] Step 3: Use the SVM binary classifier to train the training data sample set obtained by preprocessing and determine the classifier model. Specifically, the training data sample set obtained by preprocessing is sent as the target sample to the constructed partial discharge time domain waveform image recognition model for learning, and the one-dimensional convolutional neural network structure of deep learning, the TensorFlow deep learning framework and the Python language are used to build the network model to obtain the corresponding feature parameters and determine the classifier model.

[0061] like Figure 3 As shown in the figure, according to the number of different fault types n set in step 1, n(n-1) / 2 SVM binary classifiers are constructed. All binary classifiers form a binary tree structure. Each node of the binary tree structure corresponds to a binary classifier, and each leaf represents a fault category. The training data sample sets corresponding to different faults are trained in pairs using the SVM binary classifier to obtain the trained classifier model.

[0062] In the binary tree structure, the SVM binary classifier with the highest accuracy after training is used as the root node of the binary tree structure, and the SVM binary classifiers with the second highest accuracy are used as the lower nodes in turn.

[0063] In addition, when the SVM binary classifier is used to train the training data sample set, the input historical image is convolved using the following formula:

[0064]

[0065] Where: f( ) represents the activation function; F conv2D represents the two-dimensional convolution formula; b j Indicates the bias value, Represents the bias value of convolution layer l; ω ij represents the weights between neurons; represents the net activation of the jth channel of the convolutional layer l, which is obtained by filtering the feature map output by the previous layer It is obtained after convolution summation and bias; i and k respectively indicate the number of accumulated eigenvalues.

[0066] Step 4: Input the test data sample set in step 2 into the classifier model determined after training in step 3, optimize each node in the classifier, and obtain the final DAG classifier model;

[0067] In step 4, before inputting the test sample data into the classifier model in step 3, the test data sample set is subjected to one-dimensional normalization processing by the following formula:

[0068]

[0069] Among them, x * is the one-dimensional result of the data; x is the average grayscale value of the sample data; x1 represents the minimum grayscale value of the sample data; x2 represents the maximum grayscale value of the sample data.

[0070] Specifically, step 4 includes the following steps:

[0071] The step 4 comprises the following steps:

[0072] Step 401: constructing a training model by stacking multiple layers of restricted Boltzmann machine networks and one layer of BP neural network;

[0073] Step 402: Divide the processed data into a training set and a test set, and input the training set into the bottom-level restricted Boltzmann machine network layer of the training model;

[0074] Step 403: Start unsupervised pre-training, randomly initialize the parameters ε=(a, b, ω) of the restricted Boltzmann machine network, the weight ω, the visible layer bias b and the hidden layer bias a, set the number of layer nodes of the network, and the maximum number of layers m;

[0075] Step 404: The probability of neurons in the hidden layer and the visible layer of a restricted Boltzmann machine network layer being activated is calculated by the following formula, and if the calculated result is greater than a preset value, step 405 is executed, otherwise the training of the restricted Boltzmann machine network layer is terminated. Preferably, in the present invention, step 405 is executed when the calculated result is greater than 73.5%.

[0076]

[0077] Among them, h j represents the state of the jth hidden layer neuron, v j represents the state of the jth visual layer neuron; v represents the state of the visual layer neuron, h represents the state of the hidden layer neuron; b j represents the bias of the jth visual layer, a j represents the bias of the jth hidden layer, ω ij is the weight between neurons;

[0078] Step 405: Update the network weights and the biases of the visible layer and the hidden layer using the following formula:

[0079]

[0080] Where: Δ[ω ij ] n+1 is the updated weight from the i-th visual layer neuron to the j-th hidden layer neuron, n represents the n-th iteration, λ is the momentum factor, η is the learning factor, and v i represents the state of the i-th visual layer neuron, h j represents the state of the jth hidden layer neuron, (v i h j ) d represents the expected data distribution, (v i h j ) m represents the expectation of the model distribution; Δ[b j ] n+1is the updated bias of the jth visual layer, Δ[a i ] n+1 is the bias of the i-th hidden layer after update;

[0081] Step 406: Repeat steps 404-405 until all restricted Boltzmann machine network layers are trained, use the top BP network layer to perform reverse parameter optimization, and output the trained DAG model.

[0082] TensorFlow is an open source software library that can use numerical calculations. It can run stably in the Python language environment. The program takes into account certain accuracy and stability based on the implementation. TensorFlow is also completely open source and can be run on smartphones or data service center clusters. What is important is that it is based on data flow graphs and can be widely used in machine learning and deep learning fields such as speech or image classification. TensorFlow represents the calculation process required for a tensor to move one part of a flowchart to the other end. In terms of decomposition, Tensor represents a tensor (composed of image dimensions), and Flow represents operations based on data flow graphs.

[0083] A deep learning network is a neural network designed based on brain neuron reasoning. It consists of a multi-layer Restricted Boltzmann Machine (RBM) and a layer of BP neural network. Data is input through the bottom layer and then input to the hidden layer through the RBM. The output of the low-layer RBM is used as the input of the high-layer RBM. The training method of the deep learning network first uses a bottom-up unsupervised learning method to initialize the parameters of the entire model layer by layer, and then uses a top-down supervised learning method to fine-tune the network parameters.

[0084] Step 5: Collect the time domain waveform image of the partial discharge of the GIL metal particles in real time, input it into the classifier model obtained in step 4, and output the corresponding fault identification result.

[0085] The present invention realizes the image containing only partial discharge pulse voltage obtained by preprocessing, selects the support vector machine model, uses the radial basis function as the kernel function, and trains the directed acyclic graph (DAG) classifier through the deep learning algorithm to directly perform pattern recognition without manual feature extraction, with a higher recognition rate, which provides favorable support for the internal metal particle discharge fault monitoring of GIL equipment

[0086] The above embodiments are only for illustrating the technical idea of ​​the present invention, and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims

1. A fault identification method based on the time domain waveform image of partial discharge of GIL metal particles, characterized in that: The method comprises the following steps: Step 1: Collect the time domain waveform history images of GIL metal particle partial discharge under different fault types; Step 2: preprocessing the time domain waveform history image of the partial discharge of metal particles inside the GIL in step 1 to construct a training data sample set and a test data sample set, wherein the preprocessing refers to performing image enhancement processing on the time domain waveform image of the partial discharge of metal particles inside the GIL using a histogram equalization algorithm; Step 3: Use the SVM binary classifier to train the preprocessed training data sample set to determine the classifier model; Step 4: Input the test data sample set in step 2 into the classifier model determined after training in step 3, optimize each node in the classifier, and obtain the final DAG classifier model; Before the test data samples are input into the classifier model in step 3, the data are normalized into one dimension, the processed data are divided into a training set and a test set, and the training set is input into the training model; The step 4 optimizes each node in the classifier, including the following steps: Step 401: constructing a training model by stacking multiple layers of restricted Boltzmann machine networks and one layer of BP neural network; Step 402: normalize the test data samples into one dimension, divide the processed data into a training set and a test set, and input the training set into the bottommost restricted Boltzmann machine network layer of the training model; Step 403: Start unsupervised pre-training, randomly initialize the parameters ε=(a, b, ω) of the restricted Boltzmann machine network, the weight ω, the visible layer bias b and the hidden layer bias a, set the number of layer nodes of the network, and the maximum number of layers m; Step 404: The probability of neurons in the hidden layer and the visible layer of a restricted Boltzmann machine network layer being activated is calculated by the following formula, and if the calculated result is greater than a preset value, step 405 is executed, otherwise the training of the restricted Boltzmann machine network layer is terminated: Among them, h j represents the state of the jth hidden layer neuron, v j represents the state of the jth visual layer neuron; v represents the state of the visual layer neuron, h represents the state of the hidden layer neuron; b j represents the bias of the jth visual layer, a j represents the bias of the jth hidden layer, ω ij is the weight between neurons; Step 405: Update the network weights and the biases of the visible layer and the hidden layer using the following formula: Where: Δ[ω ij ] n+1 is the updated weight from the i-th visual layer neuron to the j-th hidden layer neuron, n represents the n-th iteration, λ is the momentum factor, η is the learning factor, and v i represents the state of the i-th visual layer neuron, h j represents the state of the jth hidden layer neuron, (v i h j ) d represents the expected data distribution, (v i h j ) m represents the expectation of the model distribution; Δ[b j ] n+1 is the updated bias of the jth visual layer, Δ[a i ] n+1 is the bias of the i-th hidden layer after update; Step 406: Repeat steps 404-405 until all restricted Boltzmann machine network layers are trained, use the top BP network layer to perform reverse parameter optimization, and output the trained DAG model; Step 5: Collect the time domain waveform image of the partial discharge of the GIL metal particles in real time, input it into the DAG classifier model obtained in step 4, and output the corresponding fault identification result.

2. The fault identification method based on the time domain waveform image of GIL metal particle partial discharge according to claim 1 is characterized in that: In step 1, the GIL metal particle partial discharge fault types include suspension discharge fault, surface discharge fault, air gap discharge fault, tip discharge fault and other interferences.

3. The fault identification method based on the time domain waveform image of GIL metal particle partial discharge according to claim 1 or 2, characterized in that: The step 2 comprises the following steps: Step 201: establishing an image histogram according to the grayscale distribution of the image, wherein the abscissa and ordinate represent the brightness range of each pixel in the image and the variation of each grayscale in the whole image respectively; Step 202: Assuming that the original image contains N gray levels, the distribution p(r k ): Among them, n is the sum of the number of pixels in the entire image, n k Represents the number of pixels distributed at the kth gray level, k = 0, 1, 2, ..., N-1; Step 203: discretize the distribution function of step 202 to obtain the discretized cumulative distribution function: The above formula means that the gray level is r k The pixels are transformed into grayscale S k Pixels, T(r k ) is the transformation function, P r (r i ) indicates the gray level is r i The probability of a pixel appearing in the image, k = 0, 1, 2, ..., N-1.

4. The fault identification method based on the time domain waveform image of GIL metal particle partial discharge according to claim 1 is characterized in that: In step 3, according to the number n of different fault types set in step 1, n(n-1) / 2 SVM binary classifiers are constructed, and all binary classifiers form a binary tree structure. Each node of the binary tree structure corresponds to a binary classifier, and each leaf represents a fault category. The SVM binary classifier is used to perform pairwise training on the training data sample sets corresponding to different faults to obtain the trained classifier model.

5. The fault identification method based on the time domain waveform image of GIL metal particle partial discharge according to claim 4 is characterized in that: In the binary tree structure, the SVM binary classifier with the highest accuracy after training is used as the root node of the binary tree structure, and the SVM binary classifiers with the second highest accuracy are used as the lower nodes in turn.

Images