An automatic identification method for high-frequency pulse current waveform polarity based on deep learning

Abstract

This invention discloses an automatic identification method for the polarity of high-frequency pulse current waveforms based on deep learning, belonging to the field of power equipment insulation fault detection technology. This method deeply utilizes the characteristics of the pulse signal waveform to identify the first wave and its polarity. Firstly, under laboratory conditions, the propagation characteristics of the first wave at different typical discharge locations and types are obtained by injecting steep pulses. The response signal waveforms at each outgoing line coupling end are measured. Then, using the waveform sequence as input vectors, a deep learning network is constructed. Considering convolutional neural networks, a sample library of typical response waveforms for each injection method and location is established using a digital image matrix as input. The sample library is continuously expanded through adversarial learning. An artificial neural network is used to train the first wave waveform and polarity on the input waveform sequence, and then uses the artificial neural network to identify waveform details. This achieves fully automatic first wave polarity identification, suitable for real-time algorithm applications in online monitoring.

Description

Technical Field

[0001] This invention belongs to the field of power equipment insulation fault detection technology, and specifically relates to an automatic identification method for the polarity of high-frequency pulse current waveforms based on deep learning. Background Technology

[0002] Partial discharge is an important indicator of internal insulation defects in power equipment and a crucial basis for diagnosing and locating equipment insulation faults. High-frequency partial discharge detection offers wide bandwidth and rich information, and is easily implemented through equipment grounding wires. The polarity information of high-frequency pulse current signals is essential for identifying interference pulses, determining discharge location, and classifying discharge patterns.

[0003] Under energized operating conditions, the polarity of the measured pulse current waveform is affected by background noise and interference signals, causing the initial wave of the discharge waveform to be "contaminated" and difficult to identify. Furthermore, the high-frequency components of high-frequency signals attenuate and distort after propagation through windings and long conductors, further complicating initial wave identification. Currently, in on-site partial discharge detection, testers primarily rely on experience to manually confirm waveform polarity and estimate time delays to analyze and determine the discharge type and location. Given the unavoidable personal safety risks to on-site personnel during defective equipment testing, machine identification through online monitoring or intensive care systems is essential. Existing methods for pulse waveform polarity determination include the following:

[0004] (1) First, various filtering methods are used to filter out interference signals, reduce background noise levels, and improve the signal-to-noise ratio of the detection, thereby increasing the recognizability of the first wave signal. The threshold method sets a threshold based on the background noise level, and the polarity of the pulse can be determined based on the level value of the first wave exceeding the threshold. Figure 1 The threshold method shown is used to determine the waveform polarity (transformer conduit end). Figure 1 The two parallel horizontal lines in the middle serve as a threshold; the first wave passing through the threshold is considered positive. This method is simple and intuitive, but when the background noise level is high or there are interfering signals, the signal-to-noise ratio of the detected signal decreases, making it very difficult to determine the polarity of the first wave.

[0005] (2) Directly determining the polarity of the first wave by analyzing the waveform and using the polarity of the first wave crossing the threshold is a basic method, but it is affected by threshold setting, background noise, and interference signals (such as...). Figure 2 (The diagram shows the influence of background noise and interference signals on the determination of waveform polarity.) Alternatively, using correlation analysis, a representative waveform X is selected as a reference, whose polarity is known. The similarity coefficient between the waveform u2(j) whose polarity is to be determined and the template file Y is calculated using the following formula.

[0006]

[0007] The similarity coefficient ρ≥k, where k is the determination threshold, and ρ ranges from 0 to 1. The closer ρ is to 1, the higher the similarity between the two waveforms and the higher the consistency between the polarity of the measured waveform and the reference waveform. This method is also the basic algorithm for waveform polarity judgment and time delay estimation. However, this method is also greatly affected by the background noise level. At the same time, due to the attenuation and distortion of the signal propagation process, especially the superposition effect of the reflected and refracted signals, the effectiveness of the correlation method will be seriously reduced.

[0008] (3) The method of reading the first wave of the pulse signal through the energy accumulation method and the cross-correlation method. This method can improve the consistency and stability of identification compared with directly judging through the threshold method, but its effect is still restricted by the background noise and interference signals to a great extent and degrades significantly as the level of the interference signal increases; the energy accumulation method is essentially a second-order statistic method, which observes the starting position of the first wave of the signal after squaring the signal. Considering that the signal energy is proportional to the square of the voltage, the voltage waveform of the pulse signal can be transformed into an energy-related value accumulation curve. When the partial discharge signal is much larger than the background noise, an obvious inflection point will appear on this curve, and this inflection point can be regarded as the starting moment of the partial discharge. Un is the voltage value of the nth point on the signal waveform, and h (h < N) is the number of points for the signal cumulative calculation, then the cumulative energy is

[0009] In the formula: ti is the starting moment of signal acquisition;

[0010] u(t’) is the amplitude of the UHF signal at time t’;

[0011] R is the input impedance of the acquisition system.

[0012] Thus, an energy accumulation curve is formed, and its inflection point is considered to be the starting moment of the signal. Finding the starting moment of the original signal is transformed into finding the inflection point of the energy accumulation curve (as shown in the energy accumulation curve of the pulse waveform in Figure 3 ). It can be seen from Figure 3 that the transition region of the energy accumulation curve is very gentle, and the position of the arrival of its first wave is not easy to distinguish, and the square transformation makes the polarity information of the waveform disappear. Obviously, the background noise and interference signals will seriously degrade the inflection information of this method.

[0013] (4) The fractional lower order statistics theory (FLOS) is developed from the second-order statistics theory. Introducing FLOS into the high-resolution multipath time delay estimation algorithm can improve the ability of the algorithm to resist impulse noise and solve the problem of performance degradation of the classical algorithm in the distributed noise environment, but it has a high dependence on the prior knowledge of the noise signal characteristics.

[0014] In summary, existing methods for identifying the polarity of the first wave and waveform perform well under high signal-to-noise ratio (SNR) conditions. However, as background noise levels and interference signal strength increase, and as the SNR decreases, the accuracy of identifying the polarity of the first wave and waveform drops significantly, failing to meet the requirements for diagnosing and evaluating discharge types, locations, and states in engineering applications. The lack of exploration and utilization of waveform features is a major reason for the poor performance of these methods at low SNR levels. Therefore, based on the observation and analysis of a large amount of experimental data, a deep learning-based automatic identification method for the polarity of high-frequency pulse current waveforms is proposed. This method is an automatic identification method for the polarity of the first wave of pulse signals based on neural networks. Tests show that this method has high accuracy and strong adaptability in determining the polarity of the first wave of high-frequency partial discharge signals, and is suitable for fault analysis and diagnosis in online monitoring and intensive care processes without manual intervention. Summary of the Invention

[0015] The purpose of this invention is to propose an automatic identification method for the polarity of high-frequency pulse current waveforms based on deep learning. The method is characterized by its deep utilization of the first wave and its polarity identification of pulse signal waveform features. Firstly, under laboratory conditions, the propagation characteristics of the first wave at different typical discharge locations and types are obtained by injecting steep pulses. The response signal waveforms at each outgoing coupling terminal are measured, and then a deep learning network is built using the waveform sequence as input vectors. A sample library of typical response waveforms for each injection method and location is established. The method uses an artificial neural network to train the first wave waveform and polarity on the input waveform sequence, and then uses the artificial neural network to identify waveform details. The specific steps include:

[0016] (1) Build a transformer physical model platform and simulate the internal discharge of the equipment by injecting signals into the physical transformer model through a steep pulse generator. The injection methods include simulating winding to ground discharge, winding inter-turn or inter-pane discharge, winding external discharge and winding phase-to-phase discharge.

[0017] (2) Install high-frequency CT sensors at the transformer bushing end screen, neutral point, core, clamp grounding wire, oil tank grounding wire, etc., and synchronously sample high-frequency pulse response waveforms through the acquisition device; at this time, since the equipment is not energized under the injected signal, there are basically no external interference signals entering the test circuit, and the background noise level is very low.

[0018] (3) By injecting at different locations and in different ways, a waveform sample library is established, and the polarity of the first wave is marked;

[0019] (4) Extract subsequences from the waveforms in the sample library. Use k times the root mean square value of the sequence as the threshold. Take a pulse subsequence 1 μs before and 2 μs after the first point that crosses the threshold as the input to the artificial neural network. Use the polarity of the first wave as the output to train the network parameter matrix of the neural network until the set recognition accuracy is reached. The initial stage is set to 100%. If the convergence is not achieved after reaching the set number of iterations limit, the accuracy limit can be reduced by 0.1% and retraining can be performed. Among them, k is 1.3 to 1.5.

[0020] (5) For waveform truncation subsequences in the sample library, in order to ensure that both ends obtain the waveforms from the sample library, the subsequences are windowed, including:

[0021] 1) Enlarge the sub-waveforms captured above to convert them into images, i.e., multiply them point by point with the corresponding terms of the window function so that the two ends of the waveform are 0; then perform maximum value normalization to obtain the subsequence waveforms.

[0022] 2) By artificially adding noise signals to waveform sequences in the sample library and adjusting the amplitude level of the noise signals, the signal-to-noise ratio (SNR) can be controlled, achieving a specified SNR. th Above, SNR th A signal-to-noise ratio (SNR) threshold is set, which is not less than 10 dB; the neural network is trained using image data of waveform subsequences;

[0023] 3) Build a deep learning network; Consider that convolutional neural networks (CNNs) take digital image matrices as input, and can autonomously discover and extract the color, texture, shape and topological information contained therein, which can have a certain degree of robustness and higher computational efficiency.

[0024] 4) A convolutional neural network consists of an image matrix input layer, multiple stacked convolutional and pooling layers, a fully connected layer, and a softmax output layer, forming a multi-layered independent network structure. Each layer consists of multiple parallel convolutional kernels, which are responsible for performing convolution or pooling operations on the feature map calculated by the previous layer and extracting the features contained in the data during this process. After these previous layers and the fully connected layer are connected in series, the calculated feature values ​​are fed into the softmax layer to give the classification result, so as to realize transfer learning.

[0025] (6) Apply the network parameters after training to further strengthen the training of the neural network to improve its adaptability. Once the set recognition accuracy is achieved, the network parameters can be output. Extract the collected pulse waveform and use the extracted subsequence as the input vector of the neural network to realize the automatic recognition of the polarity of the high-frequency pulse current waveform.

[0026] The structure of the Convolutional Neural Network (CNN) model is as follows: an image matrix input layer, multiple stacked convolutional and pooling layers, a fully connected layer, and a softmax output layer. The network takes a two-dimensional image as input, converts the image into a pixel matrix, preprocesses it, and then feeds it into the convolutional layers to extract feature parameters. Feature compression is performed by the pooling layers, and finally, the classification result is calculated in the fully connected layer and the softmax layer. Specifically, it includes:

[0027] (1) Input layer

[0028] The input layer of a convolutional neural network takes a two-dimensional matrix of the image after pixel value normalization preprocessing as input;

[0029] (2) Convolutional layer

[0030] Each network layer contains multiple parallel convolutional kernels. Convolution is the foundation of CNN models. This layer performs convolution operations between the input feature matrix data and the convolutional kernels to extract features. Deeper convolutional kernels extract deeper feature information. The output of this convolution operation can be calculated according to formula (1):

[0031]

[0032] In the formula, conv represents convolution calculation, and w i x is the convolution kernel parameter. i is the feature map calculated by the previous convolutional layer, b is the bias constant, and f(conv+b) is the non-linear activation function, which is a commonly used activation function, including ReLU, Sigmid and tanh functions;

[0033] The common parameters included in the convolution calculation are the kernel size, number, and stride. The calculation steps for a single convolution kernel are as follows: the kernel multiplies the parameters at corresponding positions in each convolution kernel region of the input feature image matrix and sums them to obtain the parameters at the corresponding positions in the output feature map matrix. The distance the convolution kernel moves in a single pass on the feature matrix is ​​called the stride. The convolution kernel completes one convolution operation after traversing the original feature map once. The size of the convolution kernel and the stride together determine the size of the output feature matrix. The more convolution kernels in each network layer, the richer the extracted feature parameters, but this also increases the number of computational parameters.

[0034] (3) Pooling layer

[0035] Pooling is used after convolution to compress image data, reduce the dimensionality of the feature matrix output by the previous layer, and reduce the number of network parameters while retaining the effective features in the matrix. Pooling can be regarded as performing convolution operation on the feature map with a special convolution kernel, and the convolution kernel will not act on adjacent overlapping regions in the image when traversing the feature map during the pooling operation.

[0036] Given the input feature image and the size of the pooling convolution kernel, with each parameter of the convolution kernel having a weight of 1 / 4, in the mean pooling operation, all features of the data are preserved while the image size parameter is reduced to 1 / 4; in the max pooling operation, the value of the convolution kernel at the maximum position in the corresponding original input feature map matrix is ​​1, and the other parameters are 0, and the stride of each movement on the feature map is 2, preserving the most prominent texture features in the image and reducing the size to 1 / 4 of the original image;

[0037] (4) Fully connected layer

[0038] In a convolutional neural network, the image matrix abstracted by convolutional pooling operations is converted into a column vector of a certain dimension and used as the input of a fully connected layer. This layer then extracts features again and uses them as input data for the softmax layer, which calculates and provides the classification result.

[0039] In a fully connected layer, each neuron in the network is connected to neurons in the neighboring layer, so the layer contains more parameters.

[0040] (5) Softmax layer

[0041] The number of neurons in the input layer is the dimension of the column vector obtained by stretching the feature map calculated by convolution in the previous layer into one dimension; followed by a softmax layer containing a specified number of neurons for each category to give the classification result; after the output of the last fully connected layer, the feature map is transformed into a one-dimensional feature sequence; the neurons in the softmax layer calculate the one-dimensional sequence according to equation (2) to obtain the probability of the sample to be identified belonging to each type, and take the classification output with the highest probability as the recognition result.

[0042]

[0043] Where z j This represents the output of each neuron in the fully connected layer, where M is the number of neurons in the fully connected layer.

[0044] The specific steps for training the top layer using image data of waveform subsequences are as follows:

[0045] 1) Normalize the pixels of the waveform image of the input network.

[0046] 2) Remove the top-level structure of the pre-trained CNN network and build a new fully connected layer for feature extraction and dimensionality reduction; the softmax layer outputs the recognition result of waveform polarity.

[0047] 3) Import the model parameters of each network after training on the ImageNet image set, and prohibit the updating of network parameters except before the creation of a new fully connected layer during the parameter update process.

[0048] 4) Train the top-level weight parameters of the network using the waveform diagram of the training set, calculate the output results of the training set, calculate the error between the output category and the corresponding true category label, and update the network parameters using the backpropagation algorithm and performing a specified number of iterations.

[0049] The beneficial effect of this invention is that it applies CNN transfer learning based on the ImageNet dataset in deep learning to the identification of high-frequency discharge waveform polarity. Several classic networks are used: VGG16 with sequential convolutional layers, InceptionV3 with a modular "network-within-a-network" structure, and ResNet50 with residual modules. The network parameters of the feature extraction part of each model after training on the ImageNet image set are retained. The number of neurons in the top fully connected layer of the network is modified to adapt to the weight parameter calculation when the dataset is small. The softmax layer structure is modified for the classification of two types of discharge waveform polarity. This invention utilizes artificial neural networks to identify waveform details, effectively accumulating previously acquired experience and continuously expanding the sample through adversarial learning, achieving fully automatic first-wave polarity identification. It is efficient and concise in application, suitable for real-time algorithm applications in online monitoring. Attached Figure Description

[0050] Figure 1 Determining waveform polarity using the threshold method;

[0051] Figure 2 This is a schematic diagram illustrating the influence of background noise and interference signals on the determination of waveform polarity.

[0052] Figure 3 The energy accumulation curve of the pulse waveform;

[0053] Figure 4 To simulate external interference and injected signals of different discharge forms, where a) external interference signal; b) winding to ground; c) winding turn / panel inter-turn;

[0054] Figure 5 This is a schematic diagram showing the location of the signal coupling point;

[0055] Figure 6 The typical waveform sample library includes: a) multi-terminal detection waveform when injected at the beginning of the winding, and b) multi-terminal detection waveform when injected between the pancake turns in the middle of the winding.

[0056] Figure 7 This is a flowchart for automatic polarity identification of high-frequency pulse current waveforms.

[0057] Figure 8 This is a schematic diagram of sub-waveform extraction from the original waveform.

[0058] Figure 9 for Figure 8A magnified view of a portion of the extracted sub-waveform.

[0059] Figure 10 This is a schematic diagram of windowing processing on a sub-waveform.

[0060] Figure 11 To normalize the sub-waveforms,

[0061] Figure 12 This is a classic convolutional neural network structure.

[0062] Figure 13 This is a schematic diagram of convolution calculation.

[0063] Figure 14 This is a schematic diagram of pooling operations, where a is mean pooling and b is maximum pooling;

[0064] Figure 15 This is a schematic diagram of a fully connected layer.

[0065] Figure 16 Transfer learning flowchart. Detailed Implementation

[0066] This invention proposes an automatic identification method for the polarity of high-frequency pulse current waveforms. This method deeply utilizes the characteristics of the first wave and its polarity in pulse signal waveforms. First, under laboratory conditions, the propagation characteristics of the first wave at different typical discharge locations and types are obtained by injecting steep pulses. The response signal waveforms at each outgoing line coupling terminal are measured to establish a sample library of typical response waveforms for each injection method and location. Then, using the waveform sequence as the input vector, an artificial neural network is used to train the first wave waveform and polarity on the input waveform sequence. The invention will be further described below with reference to the accompanying drawings.

[0067] The specific steps for identifying the first wave and polarity of the pulse signal waveform features are as follows: Figure 7 The flowchart for automatic polarity identification of high-frequency pulse current waveforms is shown below:

[0068] (1) Construct a physical transformer model platform. Simulate internal discharge of the equipment by injecting signals into the physical transformer model using a steep pulse generator. The injection methods include simulating winding-to-ground discharge, winding-turn or inter-turn discharge, winding-external discharge, and winding-phase-to-phase discharge, respectively, as follows: Figure 4 The diagram shows the injected signals under simulated external interference and different discharge modes, where a) is the external interference signal; b) is the winding to ground; and c) is the winding turn / panel signal.

[0069] (2) High-frequency CT sensors are installed at the transformer bushing end screen, neutral point, core, clamp grounding wire, oil tank grounding wire, etc., and high-frequency pulse response waveforms are synchronously sampled through a data acquisition device (e.g., Figure 5(As shown in the schematic diagram of the signal coupling point) Since the device is not powered under the injected signal, there are basically no external interference signals entering the test circuit, and the background noise level is very low.

[0070] (3) By injecting data at different locations and in different ways, a waveform sample library is established, and the polarity of the first wave is labeled. For example... Figure 6 Typical waveform sample library is shown. Among them, a) multi-end detection waveform when injected at the beginning of the winding, b) multi-end detection waveform when injected between the disc turns in the middle of the winding;

[0071] (4) For waveform truncation subsequences in the sample library, the method is to take k times the root mean square value of the sequence (k is 1.3 to 1.5) as the threshold, take 1 μs forward and 2 μs backward from the first point of the threshold in the sequence as the input of the artificial neural network, and use the polarity of the first wave as the output to train the network parameter matrix of the neural network until the set recognition accuracy is reached. The initial stage is set to 100%. If it still does not converge after reaching the set iteration limit, the accuracy limit can be reduced by 0.1% and re-iterate training is performed.

[0072] (5) By artificially adding noise signals to the waveform sequences in the sample library, the signal-to-noise ratio (SNR) is controlled by adjusting the amplitude level of the noise signals. th The above further enhances the neural network through training to improve its adaptability and achieve the set recognition accuracy, at which point the network parameters can be output.

[0073] (6) By applying the trained network parameters, the collected pulse waveform is truncated, and the truncated subsequence is used as the input vector of the neural network, so as to realize the automatic identification of the polarity of the first wave of high-frequency partial discharge pulse.

[0074] For waveform subsequence extraction from the sample library, the method is to use k times the root mean square value of the sequence (k ranges from 1.3 to 1.5) as a threshold, and extract pulse subsequences 1 μs before and 2 μs after the first point exceeding the threshold. This is done to ensure that both ends of the subsequence sample library waveform are obtained (e.g.,...). Figure 8 The diagram within the box shows a sub-waveform extracted from the original waveform. Windowing is applied to the sub-sequence to locally enlarge the extracted sub-waveform. Figure 9 The sub-waveform shown in the box is magnified. The sub-waveform is converted into an image by multiplying it point by point with the corresponding terms of the window function until both ends of the waveform are 0 (e.g.). Figure 10 The diagram shows a windowing process applied to the sub-waveform (the dashed lines in the diagram represent the window function waveform); then, maximum value normalization is performed to obtain the sub-sequence waveform (e.g., Figure 11 (The sub-waveform normalization process shown)

[0075] The reinforcement training of the neural network to improve its adaptability includes:

[0076] (1) Build a deep learning network; Consider that the Convolutional Neural Network (CNN) takes a digital image matrix as input, and can autonomously discover and extract the color, texture, shape and topological information contained therein, which can have a certain robustness and higher computational efficiency.

[0077] The convolutional neural network is composed of multiple independent networks connected in series. Each layer consists of multiple parallel convolutional kernels, which are responsible for performing convolution or pooling operations on the feature map calculated by the previous layer, and extracting the features contained in the data during this process. After these pre-layer and fully connected layers are connected in series, the calculated feature values ​​are fed into the softmax layer to give the classification result (e.g., ...). Figure 12 The classic convolutional neural network structure shown is used to implement transfer learning.

[0078] The structure of a CNN model is as follows: an image matrix input layer, multiple convolutional and pooling layers (or downsampling layers) stacked together, a fully connected layer, and a softmax output layer. The network uses... Figure 10 The image is taken as input, converted into a pixel matrix and preprocessed, then fed into a convolutional layer to extract feature parameters. Feature compression is performed via a pooling layer, and finally, the classification result is calculated in a fully connected layer and a softmax layer. Specifically, this includes:

[0079] (1) Input layer

[0080] The input layer of a convolutional neural network takes a two-dimensional matrix of the image after pixel value normalization as input.

[0081] (2) Convolutional layer

[0082] Each network layer contains multiple parallel convolutional kernels, and convolution is the foundation of CNN models. This layer performs convolution operations with the input feature matrix data and extracts features; deeper convolutional kernels extract deeper-level feature information. The output of a general convolution operation can be calculated according to formula (1):

[0083]

[0084] In the formula, conv represents convolution calculation, and w i x is the convolution kernel parameter. i Let be the feature map calculated from the previous convolutional layer, b be the bias constant, and f() be the non-linear activation function. Commonly used activation functions include ReLU, Sigmid, and tanh.

[0085] Common parameters involved in convolution calculations include kernel size, number of kernels, and stride. For example... Figure 13 The diagram illustrates convolution calculation, using a 5×5 feature map matrix as an example to explain the calculation steps of a single convolution kernel. The kernel size is 3×3. The kernel multiplies and sums the parameters at corresponding positions within each 3×3 region of the input feature map matrix to obtain the parameters at the corresponding positions in the output feature map matrix. The distance the kernel moves across the feature matrix in a single pass is called the stride. The kernel completes one convolution operation after traversing the original feature map once. The kernel size and stride together determine the size of the output feature matrix. The more convolution kernels in each network layer, the richer the extracted feature parameters, but this also increases the number of computational parameters.

[0086] (3) Pooling layer

[0087] Pooling is used after convolution to compress image data, reduce the dimensionality of the feature matrix output by the previous layer, and reduce the number of network parameters while retaining the effective features in the matrix.

[0088] Pooling can be viewed as performing convolution operations on feature maps using special convolution kernels. In pooling operations, the convolution kernel does not act on adjacent overlapping regions in the feature map when traversing it.

[0089] by Figure 14 Taking the pooling operation diagram (where a is mean pooling and b is max pooling) as an example, the given input feature image size is 4×4, the pooling convolution kernel size is 2×2, the weights of each parameter of the convolution kernel are all 1 / 4, and the stride of each convolution kernel movement is 2. In the mean pooling operation, all features of the data are preserved while the image size (parameters) is reduced to 1 / 4. In the max pooling operation, the value of the convolution kernel at the maximum value position in the corresponding original input feature map matrix is ​​1, and the other parameters are 0. The stride of each movement on the feature map is 2, which preserves the most prominent texture features in the image and reduces the size to 1 / 4 of the original image.

[0090] (4) Fully connected layer

[0091] In a convolutional neural network, the image matrix abstracted through operations such as convolution and pooling is converted into a column vector of a certain dimension. This vector is then used as the input to a fully connected layer. The network extracts features again through this layer, which then serves as the input data for the softmax layer, which calculates and provides the classification result.

[0092] In a fully connected layer, every neuron in the network is connected to neurons in the nearest neighboring layer. Therefore, this layer contains a relatively large number of parameters, and its structure is as follows: Figure 15 Fully connected layer diagram

[0093] (5) Softmax layer

[0094] like Figure 15 The diagram of the fully connected layer is shown, with circles representing neurons. The number of neurons in the input layer is equal to the dimension of the column vector obtained by stretching the feature map calculated by convolution in the previous layer in one dimension; followed by a softmax layer containing the specified number of neurons for each category to provide the classification result.

[0095] (5) Softmax layer

[0096] After the output of the last fully connected layer, the feature map is transformed into a one-dimensional feature sequence. The neurons in the Softmax layer calculate the one-dimensional sequence according to Equation (2) to obtain the probability of the sample to be identified belonging to each type, and take the classification output with the highest probability as the recognition result.

[0097]

[0098] Where z j This represents the output of each neuron in the fully connected layer, where M is the number of neurons in the fully connected layer.

[0099] The process of training the top layer using image data of waveform subsequences is as follows (e.g.) Figure 16 (As shown in the transfer learning flowchart)

[0100] The specific implementation steps are as follows:

[0101] 1) Normalize the pixels of the waveform image of the input network.

[0102] 2) Remove the top-level structure of the pre-trained CNN network and build a new fully connected layer for feature extraction and dimensionality reduction; the softmax layer outputs the recognition result of waveform polarity.

[0103] 3) Import the model parameters of each network after training on the ImageNet image set, and prohibit the updating of network parameters except before the creation of a new fully connected layer during the parameter update process.

[0104] 4) Train the top-level weight parameters of the network using the waveform diagram of the training set, calculate the output results of the training set, calculate the error between the output category and the corresponding true category label, and update the network parameters using the backpropagation algorithm and performing a specified number of iterations.

[0105] In summary, this invention applies CNN transfer learning based on the ImageNet dataset in deep learning to the recognition of high-frequency discharge waveform polarity. Several classic networks were used: VGG16 with sequential convolutional layers, InceptionV3 with a modular "network-within-a-network" structure, and ResNet50 with residual modules. The network parameters of the feature extraction portion of each model after training on the ImageNet image set were retained. The number of neurons in the top fully connected layer was modified to adapt to the weight parameter calculation when the dataset is small. The softmax layer structure was modified for the classification of two types of discharge waveform polarity. This invention has the following characteristics:

[0106] 1. By artificially adding noise signals to waveform sequences in the sample library, the signal-to-noise ratio (SNR) is controlled by adjusting the amplitude level of the noise signals, achieving a specified SNR. th In addition, the neural network is further trained to enhance its adaptability and achieve the set recognition accuracy.

[0107] 2. Import the saved model parameters into the network, and send the test set data after pixel matrix normalization into the network to obtain the waveform polarity recognition results.

[0108] 3. By applying the trained network parameters to truncate the acquired pulse waveform and using the truncated subsequence as the input vector of the neural network, the polarity of the first wave of high-frequency partial discharge pulses can be automatically identified. This improves the accuracy of the first wave and its polarity identification, and enhances the robustness and automation of the diagnosis.

Claims

1. A deep learning-based automatic recognition method of high-frequency pulse current waveform polarity, characterized in that, The method is a first wave and polarity recognition method deeply utilizing the waveform characteristics of the pulse signal; first, under laboratory conditions, the first wave propagation characteristics of different discharge positions and types are obtained by injecting an abrupt pulse, the response signal waveforms are measured at each coupling end, and then the waveform sequence is taken as an input vector to build a deep learning network; a typical response waveform sample library of each injection method and position is established; the first wave waveform and polarity are trained by an artificial neural network for the input waveform sequence, and the waveform details are identified by the artificial neural network, including the following steps: (1) A transformer entity model platform is built, and an abrupt pulse generator is used to simulate internal discharge of the equipment by injecting signals into the entity transformer model, and the injection methods include simulating winding-to-ground discharge, winding interturn discharge, pie interdischarge, winding external discharge and winding interphase discharge; (2) High-frequency CT sensors are installed at the end screen, neutral point, core, clamp ground wire and oil tank ground wire of the transformer bushing, and the high-frequency pulse response waveforms are synchronously sampled by a sampling device; at this time, since the equipment is not charged under the condition of signal injection, there is no external interference signal entering the test loop, and the background noise level is very low; (3) By injecting in different positions and different ways, a waveform sample library is established, and the first wave polarity is labeled; (4) The root mean square value of the sequence is taken as the threshold, the first threshold point is taken as the starting point, and the pulse subsequence of 1us forward and 2us backward is taken as the input of the artificial neural network, and the first wave polarity is taken as the output to train the network parameter matrix of the neural network until the set recognition accuracy is reached, and then stop, the initial stage is set to 100%, if the set iteration limit is reached and the accuracy is not converged, the accuracy limit can be reduced by 0.1% and the iteration training is restarted; wherein k is 1.3-1.5; (5) For the subsequence of the waveform in the sample library, in order to ensure that the subsequence sample library waveforms are obtained at both ends, the subsequence is windowed, including: 1) The above-mentioned sub-waveform is enlarged locally, the sub-waveform is converted into a picture, that is, each point is multiplied by the corresponding term of the window function to make the waveforms at both ends 0; then maximum normalization processing is performed to obtain the subsequence waveform; 2) by artificially adding noise signals to the waveform sequence in the sample library, adjusting the amplitude level of the noise signals to control the signal-to-noise ratio, and at a specified signal-to-noise ratio SNR th The SNR th is a set signal-to-noise ratio threshold, which is not less than 10dB; the image data of the waveform subsequence is used to train the top layer of the neural network; 3) A deep learning network is built; considering the convolutional neural network (CNN), the picture of the subsequence waveform is taken as the input, the color, texture, shape and topological structure information contained therein can be autonomously extracted, and the robustness and calculation efficiency are higher; 4) The convolutional neural network is composed of an image matrix input layer, a plurality of convolutional layers and pooling layers stacked, a full connection layer and a softmax output layer, which form a multi-layer independent network string connection structure, each layer is composed of a plurality of parallel convolution kernels, responsible for convolution or pooling operation on the feature map calculated by the previous layer network, and extracts the features contained in the data in the process, and after connecting the front layers and full connection layers in series, the feature quantity calculated and output is sent to the softmax layer to give the classification result, so as to realize transfer learning; (6) Using the trained network parameters, further reinforcing the neural network to improve adaptability, and outputting the network parameters when the set recognition accuracy is reached; intercepting the collected pulse waveform, using the intercepted sub-sequence as the input vector of the neural network, and thus realizing automatic recognition of the polarity of the high-frequency pulse current waveform.

2. The method of claim 1, wherein the method is a deep learning-based automatic recognition method of high-frequency pulse current waveform polarity. The structure of the convolutional neural network (CNN) model is sequentially: an image matrix input layer, a plurality of convolutional layers and pooling layer stacks, a fully connected layer, and a softmax output layer; the network takes a two-dimensional image as input, converts the image into a pixel matrix and pre-processes it, then inputs it into the convolutional layer to extract feature parameters, performs feature compression through the pooling layer, and finally calculates the classification result in the fully connected layer and the softmax layer; specifically comprising: (1) Input layer The input layer of the convolutional neural network takes the two-dimensional matrix of the image after pixel value normalization preprocessing as input; (2) Convolutional layer Each network layer contains multiple parallel convolution kernels, and convolution operation is the basis of the CNN model; this layer performs convolution operation on the input feature matrix data and the convolution kernel to extract features, and the deeper the convolution kernel, the deeper the feature information extracted; the output of this convolution operation can be calculated according to formula (1): where conv represents a convolution calculation, w i is a convolution kernel parameter, x i is a feature map obtained by a convolution calculation of a previous layer, b is a bias constant, and f(conv+b) is a nonlinear activation function, which is a commonly used activation function including a Relu, a Sigmid, and a tanh function. (3) Pooling layer The pooling operation is used after the convolution operation to compress image data and reduce the dimensionality of the feature matrix output by the previous layer, while retaining the effective features in the matrix and reducing the network parameter quantity; the pooling can be regarded as a convolution operation on the feature map with a special convolution kernel, and the convolution kernel does not act on the adjacent overlapping areas in the image during the pooling operation; given the input feature image and the size of the pooling convolution kernel, the parameter weights of the convolution kernel are all 1 / 4, in the mean pooling operation, the data retains all the features while the image size parameter is reduced to 1 / 4; in the maximum value pooling operation, the value of the convolution kernel at the position of the maximum value in the corresponding original input feature matrix is 1, and the rest of the parameters are 0, and the step length of each movement on the feature map is 2, retaining the most prominent features of the texture in the image and reducing the size to 1 / 4 of the original image; (4) Fully connected layer The image matrix abstracted by the convolution and pooling operations in the convolutional neural network is converted into a column vector of a certain dimension, which is used as the input of the fully connected layer, and the features are extracted again through the layer network, which is used as the input data of the softmax layer and the classification result is calculated by the softmax layer; In the fully connected layer, each neuron in the network is connected to the neurons in the adjacent layer, so this layer contains more parameters; (5) Softmax layer The number of neurons in the input layer is the dimension of the column vector obtained by one-dimensional stretching of the feature map calculated by the convolution; the softmax layer containing a specified number of neurons is used to give the classification result; after output by the last fully connected layer, the feature map is transformed into a one-dimensional feature number column; the neurons in the softmax layer calculate the one-dimensional number column according to formula (2) to obtain the probability that the sample to be recognized belongs to each type, and the classification with the maximum probability is output as the recognition result. where z j is the output result of each neuron in the fully connected layer, and M is the number of neurons contained in the fully connected layer.

3. The automatic recognition method of high-frequency pulse current waveform polarity based on deep learning according to claim 2, characterized in that, The common parameters contained in the convolution calculation include convolution kernel size, number and moving step. The calculation steps of a single convolution kernel are as follows: the parameters of the corresponding positions in each convolution kernel region on the input feature image matrix are multiplied and summed, and then the parameters at the corresponding positions of the output feature map matrix are obtained. The distance of a single movement of the convolution kernel on the feature matrix is called the step. The convolution kernel completes a convolution operation after a round of traversal on the original feature map. The size of the convolution kernel and the step jointly determine the size of the output feature matrix. The more the number of convolution kernels contained in each layer of the network, the richer the extracted feature parameters, but at the same time, the calculation parameters also increase.

4. The method of claim 1, wherein the method is a deep learning-based automatic recognition method of high-frequency pulse current waveform polarity. The specific implementation steps of training the top layer of the neural network using the image data of the waveform subsequence are as follows: 1) Normalizing the pixels of the waveform image input into the network; 2) Removing the top layer parameters of the pre-trained CNN network, and building a new fully connected layer for feature extraction and dimension reduction; The softmax layer outputs the recognition result of the waveform polarity; 3) Importing the model parameters of each network trained by the ImageNet image set, and prohibiting the update of the network parameters before the newly built fully connected layer during the parameter update process; 4) Training the top layer weight parameters of the network using the waveform image training set, calculating the output result of the training set, calculating the error between the output category and the corresponding real category label, using the back propagation algorithm and performing network parameter update for a specified number of iterations.

Images