A method and apparatus for noise processing of partial discharge signals

Abstract

The application provides a noise processing method and device for partial discharge signals. The noise processing method comprises the following steps: obtaining real partial discharge samples, wherein the real partial discharge samples comprise real partial discharge signals; performing probability density estimation on the real partial discharge signals in each real partial discharge sample to obtain the probability distribution of the noise of the real partial discharge signals; adding noise signals conforming to the probability distribution to the simulation partial discharge signals without noise to obtain partial discharge signals with noise; generating de-noised partial discharge samples based on the partial discharge signals with noise and the simulation partial discharge signals without noise, training a pre-set neural network model based on the de-noised partial discharge samples, and obtaining a de-noising model; and performing de-noising processing on real-time obtained real partial discharge signals based on the de-noising model obtained through training to obtain partial discharge signals without noise. The application can improve the accuracy of the de-noising process of the partial discharge signals.

Description

Technical Field

[0001] This invention relates to the field of power transmission and transformation technology, and in particular to a method and apparatus for noise processing of partial discharge signals. Background Technology

[0002] In the actual operation of power systems, electrical equipment is affected by internal factors and interference introduced by the external environment. This causes the partial discharge signals collected in real time by sensors and other monitoring equipment to be mixed with varying degrees of noise, resulting in the disturbance characteristics in the partial discharge signal being masked by noise. As a non-stationary, time-varying signal, the partial discharge signal contains the characteristic information required for partial discharge identification. One of the important steps in signal analysis is to extract feature quantities from the signal, but before this, a "clean" signal must be obtained, minimizing interference from various noise sources. This necessitates background noise processing of the noisy signal. Background noise processing of the partial discharge signal is crucial for better extraction of its characteristic information and optimization of the identification algorithm. Therefore, effectively addressing the interference caused by noise is of paramount importance.

[0003] Currently, noise denoising is often achieved by learning the noise characteristics in real partial discharge signals and training models. However, due to the complex on-site environment, the limited number of partial discharge signals collected, and the numerous interference factors, the noise signal data obtained solely through feature learning and simulation cannot closely approximate the noise data of real signals, resulting in low accuracy in the partial discharge signal denoising process. Summary of the Invention

[0004] This invention provides a noise processing method and apparatus for partial discharge signals, which can improve the accuracy of the partial discharge signal denoising process.

[0005] In a first aspect, the present invention provides a noise processing method and apparatus for partial discharge (PD) signals, comprising: acquiring real PD samples, wherein the real PD samples include real PD signals; performing probability density estimation on the real PD signals in each real PD sample to obtain the probability distribution of noise in the real PD signals; adding a noise signal conforming to the probability distribution to a noise-free simulated PD signal to obtain a noisy PD signal; generating denoised PD samples based on the noisy PD signal and the noise-free simulated PD signal, and training a pre-set neural network model based on the denoised PD samples to obtain a denoising model; and performing denoising processing on the real PD signals acquired in real time based on the trained denoising model to obtain a noise-free PD signal.

[0006] This invention provides a noise processing method for partial discharge (PD) signals. By estimating the probability density of each real PD signal, the probability distribution of noise in the real PD signal is obtained. Then, noise signals conforming to the probability distribution are added to a noise-free simulated PD signal to obtain a noisy PD signal. In this way, this invention can obtain all noise signals conforming to the probability distribution of noise in the real PD signal, expanding the number and types of noisy PD signals. Therefore, this invention can obtain richer denoising PD samples and train a more accurate denoising model. Based on the trained denoising model, real PD signals acquired in real time are denoised to obtain a cleaner, noise-free PD signal, improving the accuracy of the PD signal denoising process.

[0007] In one possible implementation, probability density estimation of the real partial discharge signal in each real partial discharge sample is performed to obtain the probability distribution of noise in the real partial discharge signal. This includes: inputting each real partial discharge signal into a sample generation model, which divides each real partial discharge signal into sliding time windows; reading waveform data of each real partial discharge signal in different time windows, the waveform data including waveform, amplitude, and phase angle; calculating kernel density estimates of each real partial discharge signal in different time windows based on the waveform data of each real partial discharge signal in different time windows; and determining the probability distribution of noise in the real partial discharge signal based on the kernel density estimates of each real partial discharge signal in different time windows.

[0008] In one possible implementation, a noise signal conforming to a probability distribution is added to a noise-free simulated partial discharge signal to obtain a noisy partial discharge signal, including: a sample generation model generating a noise-free simulated partial discharge signal; a sample generation model generating a noise signal based on a probability distribution; and superimposing the noise-free simulated partial discharge signal and the noise signal to obtain the noisy partial discharge signal.

[0009] In one possible implementation, denoised samples are generated based on noisy partial discharge (PD) signals and noiseless simulated PD signals. A pre-set neural network model is then trained using these denoised samples to obtain a denoised model. This process includes: determining noisy PD samples, where the noisy PD signal is the input and the noiseless simulated PD signal is the output; iteratively training the pre-set neural network model using a backpropagation algorithm based on the training set from the noisy PD samples; testing the trained neural network model each time based on the test set from the noisy PD samples; and exiting the iterative process if the loss function is less than a set value, or if the training of the noisy PD samples is complete, thus obtaining the denoised model. The loss function characterizes the difference between the PD signal output by the neural network model and the noiseless simulated PD signal.

[0010] In one possible implementation, the pre-set neural network model includes multiple convolutional layers; the pre-set neural network model is trained based on denoised partial discharge samples to obtain a denoised model, and the training process further includes: during the training process, for each convolutional layer, calculating the attention information of each feature in the output of that convolutional layer; wherein the attention information of the feature is used to characterize the importance of the feature; determining the weight of each feature based on the attention information of each feature, wherein the attention information of each feature is different, and different attention information corresponds to different weights; and determining the input of the next convolutional layer based on each feature and the weight of each feature.

[0011] In one possible implementation, a pre-set neural network model is trained based on denoised partial discharge samples to obtain a denoised model, and the method further includes: calculating the loss function after each iteration based on the following formula;

[0012]

[0013] Where loss represents the value of the loss function, C j H represents the number of channels in the feature map of a neural network model. j W represents the length of the feature map in the neural network model. j This represents the width of the feature map in the neural network model. Represents the loss function. The loss network function representing a sharp image, This represents the loss network function of the denoised image.

[0014] In one possible implementation, the real partial discharge sample also includes the partial discharge type corresponding to the real partial discharge signal; after denoising the real partial discharge signal acquired in real time based on the trained denoising model to obtain a noise-free partial discharge signal, the implementation further includes: determining a noise-free partial discharge sample based on the noise-free partial discharge signal and the partial discharge type corresponding to the noise-free partial discharge signal; training a pre-set partial discharge recognition model based on the noise-free partial discharge sample to obtain a trained partial discharge recognition model; and inputting the real partial discharge signal acquired in real time into the trained partial discharge recognition model to obtain the partial discharge type corresponding to the real partial discharge signal acquired in real time.

[0015] Secondly, embodiments of the present invention provide a noise processing apparatus for partial discharge signals, comprising: a communication module for acquiring real partial discharge samples, the real partial discharge samples including real partial discharge signals; a processing module for estimating the probability density of the real partial discharge signals in each real partial discharge sample to obtain the probability distribution of noise in the real partial discharge signals; adding a noise signal conforming to the probability distribution to a noise-free simulated partial discharge signal to obtain a noisy partial discharge signal; generating denoised partial discharge samples based on the noisy partial discharge signal and the noise-free simulated partial discharge signal, and training a pre-set neural network model based on the denoised partial discharge samples to obtain a denoising model; and performing denoising processing on the real partial discharge signals acquired in real time based on the trained denoising model to obtain a noise-free partial discharge signal.

[0016] In one possible implementation, the processing module is specifically used to input each real partial discharge signal into the sample generation model. The sample generation model divides each real partial discharge signal into sliding time windows, reads the waveform data of each real partial discharge signal in different time windows, and the waveform data includes waveform, amplitude, and phase angle. Based on the waveform data of each real partial discharge signal in different time windows, the kernel density estimate of each real partial discharge signal in different time windows is calculated. Based on the kernel density estimate of each real partial discharge signal in different time windows, the probability distribution of noise in the real partial discharge signal is determined.

[0017] In one possible implementation, the processing module is specifically used to generate a noise-free simulated partial discharge signal by a sample generation model; the sample generation model generates a noise signal based on a probability distribution; and the noise-free simulated partial discharge signal and the noise signal are superimposed to obtain a noisy partial discharge signal.

[0018] In one possible implementation, the processing module is specifically used to determine noisy partial discharge (PD) samples, which take noisy PD signals as input and noise-free simulated PD signals as output. Based on the training set in the noisy PD samples, a pre-set neural network model is iteratively trained using the backpropagation algorithm. During the training process, the neural network model is tested for each completed training session based on the test set in the noisy PD samples. If the loss function is less than a set value, or the training of the noisy PD samples is completed, the iterative process is exited, and a denoised model is obtained. The loss function is used to characterize the difference between the PD signal output by the neural network model and the noise-free simulated PD signal.

[0019] In one possible implementation, the pre-configured neural network model includes multiple convolutional layers; a processing module, specifically used during training, to calculate the attention information of each feature output by each convolutional layer; wherein the attention information of a feature is used to characterize the importance of that feature; based on the attention information of each feature, the weight of each feature is determined, wherein the attention information of each feature is different, and different attention information corresponds to different weights; based on each feature and the weight of each feature, the input of the next convolutional layer is determined.

[0020] In one possible implementation, the processing module is specifically used to calculate the loss function after each iteration based on the following formula;

[0021]

[0022] Where loss represents the value of the loss function, C j H represents the number of channels in the feature map of a neural network model. j W represents the length of the feature map in the neural network model. j This represents the width of the feature map in the neural network model. Represents the loss function. The loss network function representing a sharp image, This represents the loss network function of the denoised image.

[0023] In one possible implementation, the processing module is further configured to determine noiseless partial discharge samples based on noiseless partial discharge signals and the partial discharge types corresponding to the noiseless partial discharge signals; train a pre-set partial discharge identification model based on the noiseless partial discharge samples to obtain a trained partial discharge identification model; and input the real partial discharge signals acquired in real time into the trained partial discharge identification model to obtain the partial discharge type corresponding to the real partial discharge signals acquired in real time.

[0024] Thirdly, embodiments of the present invention provide an electronic device, characterized in that the electronic device includes a memory and a processor, the memory storing a computer program, and the processor being configured to call and run the computer program stored in the memory to perform the steps of the method as described in the first aspect and any possible implementation thereof.

[0025] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of the method as described in the first aspect and any possible implementation thereof.

[0026] The technical effects of any of the implementation methods in the second to fourth aspects mentioned above can be found in the technical effects of the corresponding implementation method in the first aspect, and will not be repeated here. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a flowchart illustrating a noise processing method for partial discharge signals provided in an embodiment of the present invention.

[0029] Figure 2 This is a flowchart illustrating another noise processing method for partial discharge signals provided in an embodiment of the present invention.

[0030] Figure 3 This is a schematic diagram of a noisy partial discharge signal provided in an embodiment of the present invention;

[0031] Figure 4 This is a schematic diagram of a noise-free partial discharge signal provided in an embodiment of the present invention.

[0032] Figure 5 This is a schematic diagram of the structure of a noise processing device for partial discharge signals provided in an embodiment of the present invention;

[0033] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0034] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0035] In the description of this invention, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The term "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" and "more than one" refer to two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0036] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner to facilitate understanding.

[0037] Furthermore, the terms "comprising" and "having," and any variations thereof, used in the description of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or modules is not limited to the steps or modules listed, but may optionally include other steps or modules not listed, or may optionally include other steps or modules inherent to such process, method, product, or device.

[0038] To make the objectives, technical solutions, and advantages of the present invention clearer, the following description will be provided in conjunction with the accompanying drawings and specific embodiments.

[0039] Currently, denoising methods are mainly divided into two categories: traditional denoising methods based on artificial features and denoising methods based on deep learning. Deep learning-based denoising methods possess powerful learning capabilities, not only fitting complex noise distributions but also saving computation time. However, deep learning denoising requires a large amount of data to train the model; otherwise, the denoising effect may be significantly reduced. Due to limitations, sufficient field data is often unavailable for training the denoising model. Therefore, simulation or programming is needed to simulate real-world noise, typically adding Gaussian white noise and narrowband impulse noise. However, due to the complex field environment, the limited number of partial discharge signals collected, and the numerous interference factors, noise signal data obtained solely through feature learning and simulation cannot closely approximate the noise data of the real signal, resulting in low accuracy in the partial discharge signal denoising process.

[0040] To address the aforementioned technical problems, this invention provides a noise processing method for partial discharge signals. Based on kernel density estimation and convolutional neural networks, the noise is analyzed to obtain its probability distribution, thereby yielding a noise signal that more closely resembles reality and expanding the quantity and types of noise signals. This provides sufficient data for denoising models based on convolutional neural networks. Consequently, during the training process of the convolutional neural network, data mining capabilities can be better utilized, improving the performance of the denoising model.

[0041] It should be noted that kernel density estimation (KDE), also known as Parzen window estimation, is a density estimation method proposed and developed in the 1950s and 60s. KDE belongs to the nonparametric estimation method. Probability distribution is a core issue in statistical inference; once the joint probability density is given, all questions regarding subsets of variables can be answered. In other words, the core of parametric statistics is the estimation of density. Often in practical problems, the distribution of much data is unknown and can only be determined through assumptions; kernel density estimation arose precisely to address this problem.

[0042] like Figure 1 As shown, this embodiment of the invention provides a noise processing method for partial discharge signals, wherein the execution subject is a noise processing device, and the noise processing method includes steps S101-S105.

[0043] S101. Obtain real partial discharge samples.

[0044] In this embodiment of the application, the real partial discharge sample includes the real partial discharge signal.

[0045] In some embodiments, the real partial discharge sample also includes the partial discharge type corresponding to the real partial discharge signal.

[0046] S102. Estimate the probability density of the real partial discharge signal in each real partial discharge sample to obtain the probability distribution of noise in the real partial discharge signal.

[0047] In some embodiments, the probability distribution of noise is used to characterize the probability of various types of noise occurring.

[0048] As one possible implementation, the noise processing device can determine the probability distribution of noise in the real partial discharge signal based on steps S1021-S1023.

[0049] S1021. Input each real partial discharge signal into the sample generation model. The sample generation model divides each real partial discharge signal into sliding time windows and reads the waveform data of each real partial discharge signal in different time windows.

[0050] In some embodiments, waveform data includes waveform, amplitude, and phase angle;

[0051] S1022. Based on the waveform data of each real partial discharge signal in different time windows, calculate the kernel density estimate of each real partial discharge signal in different time windows.

[0052] In some embodiments, the kernel density estimate may be determined based on the kernel density function.

[0053] For example, the kernel density function can be a uniform kernel function. A uniform kernel function can be expressed as follows.

[0054]

[0055] Where K(u) is the kernel density function, and u is the independent variable of the kernel density function.

[0056] For example, the kernel density function can be a quartic kernel function. A quartic kernel function can be expressed as follows.

[0057]

[0058] Where K(u) is the kernel density function, and u is the independent variable of the kernel density function.

[0059] For example, the kernel density function can be a triangular kernel function. The triangular kernel function can be expressed as follows.

[0060] K(u) = 1 - |u| (|u| ≤ 1);

[0061] Where K(u) is the kernel density function, and u is the independent variable of the kernel density function.

[0062] For example, the kernel density function can be a Gaussian kernel function. The Gaussian kernel function can be expressed as follows.

[0063]

[0064] Where K(u) is the kernel density function, and u is the independent variable of the kernel density function.

[0065] It should be noted that extensive simulations show that when the sample size is large enough, the type of kernel density function is not the most important factor in estimating the probability density function of the unknown variable; any function can be used as the kernel function.

[0066] As one possible implementation, for any real partial discharge signal at any time window, the noise processing device can determine the kernel density estimate of the real partial discharge signal at that time window based on the following formula.

[0067]

[0068]

[0069] Where f(x) represents the kernel density estimate, n represents the sample size, h represents the window width of the time window, K(u) represents the kernel density function, exp[u] represents the exponential function with base e, u represents the independent variable of the kernel density function, and x i Let x represent the monitoring point of the i-th sample, and let x represent the noise estimation point.

[0070] In some embodiments, the noise processing apparatus may determine the window width of the time window based on the following formula.

[0071]

[0072] Where h is the window width of the time window, K(t) represents the kernel density function, n represents the number of samples, t represents the noise level, i.e. the size of the noise pixels in the image, f′(x) represents the derivative of the probability density function, f(x) represents the probability density function of the sample noise, and x represents the noise estimation point.

[0073] It should be noted that the window width h of the time window, also known as the smoothing parameter, affects the fitting effect and has a significant impact on the accuracy of probability density function estimation. It is the key to the quality of the estimation.

[0074] As can be seen from the formula for calculating the kernel density estimate, when h is too large, the influence of averaging becomes prominent, obscuring the details of the density function and failing to clearly reflect its characteristics. When h is too small, the influence of randomness increases, making the density function less smooth and failing to reflect its trend well. Therefore, h should be appropriately chosen to balance these two effects.

[0075] The quality of a window width can be judged using the integral mean square error (IMSE). Currently, the main principles followed include the Minimum Integral Mean Square Error (MISE) method and the Minimum Asymptotic Integral Mean Square Error (AMISE) method.

[0076]

[0077] Where AMISE represents the minimum value of the mean squared error. Let f(x) represent the probability density estimation function of the noise, f′(x) represent the probability density function of the sample noise, K(t) represent the kernel function, x represent the noise monitoring point, n represent the number of samples, h represent the window width of the time window, i.e. the smoothing parameter, and t represent the noise magnitude, i.e. the size of the noise pixels in the image.

[0078] To minimize the mean square error of the asymptotic integral, an appropriate value of h must be chosen, i.e., the optimal window width h must be obtained by differentiation.

[0079] make but

[0080]

[0081] The optimal window width can be calculated based on the above formula, as shown below.

[0082]

[0083] When choosing K(u) as the Gaussian kernel function, a simplified calculation method using a practically applicable window width estimation method can be employed: the Silverman's thumb rule. This rule assumes that the random variable X follows a normal distribution with a mean of 0 and a variance equal to the sample variance. The optimal window width can then be expressed as the following formula.

[0084]

[0085] in, Let f(x) represent the probability density estimation function of the noise, f′(x) represent the probability density function of the sample noise, K(t) represent the kernel function, x represent the noise monitoring points, n represent the number of samples, h represent the window width of the time window (i.e., the smoothing parameter), and t represent the noise magnitude (i.e., the size of the noisy pixels in the image). This represents the sample variance.

[0086] It is understandable that the optimal window width determined by the above formula, i.e., the adaptive smoothing parameter method, makes the smoothing parameter smaller in areas of high density and larger in areas of low density, thereby obtaining a smoother estimate and making the data fit better.

[0087] S1023. Based on the kernel density estimates of each real partial discharge signal in different time windows, determine the probability distribution of noise in the real partial discharge signal.

[0088] It should be noted that, after kernel density estimation, the true noise distribution is fitted. Based on the noise distribution, noise synthesis is performed on the simulated partial discharge pulse signal. The synthesized noisy signal is then set as different types of noisy partial discharge samples Y. u U = 1, 2, 3, 4, corresponding to the tip discharge dataset, bubble discharge dataset, suspension discharge dataset, and surface discharge dataset, respectively. A CNN neural network model is established and trained using the training set. During training, the network continuously learns and eventually outputs a noise map that is infinitely close to the real noise level. Finally, the noise map is subtracted from the noisy image to obtain a clean image, thus realizing the denoising process of the partial discharge signal.

[0089] S103. Add a noise signal that conforms to the probability distribution to the noise-free simulated partial discharge signal to obtain a noisy partial discharge signal.

[0090] As one possible implementation, the noise processing device can determine the noisy partial discharge signal based on steps S1031-S1033.

[0091] S1031, The sample generation model generates noise-free simulated partial discharge signals.

[0092] S1032, The sample generation model generates noise signals based on probability distribution.

[0093] S1033. Superimpose the noise-free simulated partial discharge signal and the noise signal to obtain the noisy partial discharge signal.

[0094] S104. Based on the noisy partial discharge signal and the noiseless simulated partial discharge signal, generate denoised partial discharge samples, and train the pre-set neural network model based on the denoised partial discharge samples to obtain the denoised model.

[0095] As one possible implementation, the noise processing device can determine the denoising model based on steps S1041-S1044.

[0096] S1041. Identify noisy partial discharge samples.

[0097] In some embodiments, the noisy partial discharge sample takes a noisy partial discharge signal as input and a noiseless simulated partial discharge signal as output.

[0098] For example, noisy partial discharge samples can be configured to be of different types. Among them, noisy partial discharge samples can include tip discharge noisy partial discharge samples Y. 1 Bubble discharge with noisy partial discharge sample Y 2 Suspended discharge with noisy partial discharge sample Y 3 Sample Y with noisy partial discharge along the surface 4 .

[0099] As one possible implementation, the noise processing device can normalize the noisy partial discharge samples and train a neural network model based on the normalized noisy partial discharge samples.

[0100] S1042. Based on the training set in the noisy partial discharge samples, the pre-set neural network model is iteratively trained using the backpropagation algorithm.

[0101] In some embodiments, the noise processing apparatus may divide the noisy partial discharge samples into a training set and a test set. For example, the noise processing apparatus may divide the noisy partial discharge samples into the training set and the test set in a 7:3 ratio.

[0102] Optionally, the pre-configured neural network model may include multiple convolutional layers.

[0103] Correspondingly, during the iterative training process, the noise processing device can update the input of each convolutional layer by analyzing the attention information of the features in each convolutional layer.

[0104] For example, the noise processing device may determine the input of the next convolutional layer based on steps A1-A3.

[0105] A1. During training, for each convolutional layer, calculate the attention information of each feature in the output of that convolutional layer.

[0106] The attention information of a feature is used to characterize the importance of that feature.

[0107] A2. Determine the weight of each feature based on the attention information of each feature.

[0108] Among them, the attention information of each feature is different, and different attention information corresponds to different weights;

[0109] A3. Based on each feature and its weight, determine the input to the next convolutional layer.

[0110] It should be noted that the noise processing device first inputs the noisy partial discharge signal into the convolutional layer. The convolutional layer adopts a 3×3 structure. The convolutional layer saves the spatial relationship between images by learning image features from local regions of the noisy partial discharge signal. The operation formula of the convolution kernel is shown below.

[0111]

[0112] Where the size of the convolution kernel is S×S, a i,j The parameter value in the i-th row and j-th column represents the convolution kernel, f represents the activation function, D represents the number of image channels, and w represents the parameter value in the i-th row and j-th column. m,n,d x represents the convolution kernel weights. i+m,j+n,d is the value of the pixel (i+m, j+d) in the d-th channel of the feature map, and b is the kernel bias.

[0113] It should be noted that convolution operations primarily extract local features from the input image. Shallow convolutional layers extract low-level image features, while deeper convolutional layers extract more global and complex features, which are combinations of simpler features. The ReLU activation function is used for convolutional layers: y = max(0, x).

[0114] Optionally, the pre-configured neural network model may also include pooling layers.

[0115] Pooling layers further process the features extracted by convolutional layers. Pooling is essentially downsampling, aimed at reducing the size of the feature map while retaining important information. Like convolution, pooling is a local connection operation. A pooling layer defines a spatial neighborhood and performs pooling operations within that neighborhood.

[0116] Pooling layers not only reduce size but also decrease the number of parameters and computations in the neural network, thus mitigating overfitting to some extent. Furthermore, pooling makes the network more resistant to small deformations, translations, and distortions, improving the model's fault tolerance.

[0117] Optionally, the noise processing device can also modify the gradient parameters of the pre-set neural network model in real time during training to improve the model's generalization ability.

[0118] It should be noted that since the training process involves gradient descent, there is a possibility of gradient vanishing during backpropagation, which leads to slow training convergence. Batch standardization is used to reduce the dependence of gradients on parameters, speed up network training, and improve the model's generalization ability.

[0119] The formulas for the mean, variance, standardization, scaling, and translation used in batch standardization are as follows:

[0120]

[0121] Where, x i (i = 1, 2, ..., m) represents the hidden layer output of the neural network, ε is the first preset value, representing an infinitesimal quantity used to avoid dividing the variance by zero, γ is the second preset value, β is the second preset value, and γ and β are learnable parameters. μ represents the sample mean, m represents the number of samples after batch processing, σ represents the sample variance, and x... i Let x represent the value of the i-th sample. i,norm This represents the mean value of the sample.

[0122] As one possible implementation, the noise processing device can use the backpropagation algorithm to train the convolution kernel parameters and update the parameters in the network by implementing negative gradient reduction.

[0123] For example, the noise processing device can update the parameters based on the following formula.

[0124]

[0125] Where w2 represents the updated convolutional kernel parameters, w1 represents the original convolutional kernel parameters, C represents the loss function, and α represents the learning rate.

[0126] S1043. During the training process, the neural network model is tested for each completed training session based on the test set in the noisy partial discharge samples.

[0127] S1044. If the loss function is less than the set value, or the training of noisy partial discharge samples is completed, then exit the iteration process and obtain the denoising model.

[0128] In some embodiments, the loss function is used to characterize the difference between the partial discharge signal output by the neural network model and the noise-free simulated partial discharge signal.

[0129] As one possible implementation, the noise processing device can calculate the loss function after each iteration based on the following formula;

[0130]

[0131] Where loss represents the value of the loss function, C j H represents the number of channels in the feature map of a neural network model. j W represents the length of the feature map in the neural network model. j This represents the width of the feature map in the neural network model. Represents the loss function. The loss network function representing a sharp image, This represents the loss network function of the denoised image.

[0132] It should be noted that the formula for calculating gradient descent is shown below.

[0133] z l =w l a l-1 +b l ,a l =σ(z) l );

[0134] Where x is the input and a is the activation value.

[0135] After forward propagation to the output layer, an error will be generated. The formula for calculating the error is shown below.

[0136]

[0137] The error of the current layer is used to represent the error of the previous layer using the chain rule, and the calculation formula is shown below.

[0138] δ l =((w l+1 ) T )δ l+1 *σ l (z l )

[0139] Finally, gradient descent is used to train the parameters and update the weights w. l and bias parameter b l .

[0140]

[0141]

[0142] Among them, z l w represents the output value of the forward propagation at layer l. l w represents the weight of the l-th layer. l+1 a represents the weight of the (l+1)th layer. l-1 b represents the activation value of the input in the (l-1)th layer. l σ(z) represents the threshold of the l-th layer. l )express δ represents the error generated during forward propagation. l δ represents the error of the l-th layer. l+1 Let C represent the error of the (l+1)th layer, and let C represent the loss function. σ represents the output value of the output layer. l This represents the derivative of the activation function. η represents the output value of the forward pass, m represents the number of hidden layer outputs, and δ represents the number of hidden layer outputs. x,l Let a represent the x-th error of the l-th layer. x,l-1 S105 represents the x-th activation value of the (l-1)-th layer. Based on the denoising model obtained through training, the real partial discharge signal acquired in real time is denoised to obtain a noise-free partial discharge signal.

[0143] This invention provides a noise processing method for partial discharge (PD) signals. By estimating the probability density of each real PD signal, the probability distribution of noise in the real PD signal is obtained. Then, noise signals conforming to the probability distribution are added to a noise-free simulated PD signal to obtain a noisy PD signal. In this way, this invention can obtain all noise signals conforming to the probability distribution of noise in the real PD signal, expanding the number and types of noisy PD signals. Therefore, this invention can obtain richer denoising PD samples and train a more accurate denoising model. Based on the trained denoising model, real PD signals acquired in real time are denoised to obtain a cleaner, noise-free PD signal, improving the accuracy of the PD signal denoising process.

[0144] Optionally, the real partial discharge samples also include the partial discharge type corresponding to the real partial discharge signal.

[0145] Correspondingly, such as Figure 2As shown, this embodiment of the invention provides a noise processing method for partial discharge signals, and further includes steps S201-S203 after step S105.

[0146] S201. Based on the noiseless partial discharge signal and the partial discharge type corresponding to the noiseless partial discharge signal, determine the noiseless partial discharge sample.

[0147] S202. Based on noiseless partial discharge samples, train the pre-set partial discharge identification model to obtain the trained partial discharge identification model.

[0148] S203. Input the real partial discharge signal acquired in real time into the trained partial discharge identification model to obtain the partial discharge type corresponding to the real partial discharge signal acquired in real time.

[0149] In this way, embodiments of the present invention can train a partial discharge (PD) identification model based on denoised, noise-free PD samples. Compared with noisy PD signals, the present invention enables the PD identification model to more accurately acquire the characteristics of PD signals, achieve more accurate prediction of PD types, and improve the accuracy of PD type identification.

[0150] For example, such as Figure 3 As shown in the figure, the noisy partial discharge signals corresponding to each partial discharge type provided in the embodiments of the present invention are as follows: the upper left figure is the noisy partial discharge signal of tip discharge, the upper right figure is the noisy partial discharge signal of bubble discharge, the lower left figure is the noisy partial discharge signal of suspension discharge, and the lower right figure is the noisy partial discharge signal of surface discharge.

[0151] For example, such as Figure 4 As shown in the figure, the noiseless partial discharge signals corresponding to each partial discharge type provided in the embodiments of the present invention are as follows: the upper left figure is the noiseless partial discharge signal for tip discharge, the upper right figure is the noiseless partial discharge signal for bubble discharge, the lower left figure is the noiseless partial discharge signal for suspension discharge, and the lower right figure is the noiseless partial discharge signal for surface discharge.

[0152] Depend on Figure 3 and Figure 4 As can be seen, the embodiments of the present invention generate various types of noisy partial discharge signals, expand the number and types of noisy partial discharge signals, improve the performance of the denoising model, obtain a purer noise-free partial discharge signal, and improve the accuracy of the partial discharge signal denoising process.

[0153] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0154] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0155] Figure 5A schematic diagram of a noise processing device for partial discharge signals provided in an embodiment of the present invention is shown. The noise processing device 300 includes a communication module 301 and a processing module 302.

[0156] The communication module 301 is used to acquire real partial discharge samples, which include real partial discharge signals.

[0157] The processing module 302 is used to estimate the probability density of the real partial discharge signal in each real partial discharge sample to obtain the probability distribution of noise in the real partial discharge signal; add a noise signal that conforms to the probability distribution to the noiseless simulated partial discharge signal to obtain a noisy partial discharge signal; generate denoised partial discharge samples based on the noisy partial discharge signal and the noiseless simulated partial discharge signal, and train a pre-set neural network model based on the denoised partial discharge samples to obtain a denoising model; and perform denoising processing on the real partial discharge signal acquired in real time based on the trained denoising model to obtain a noiseless partial discharge signal.

[0158] In one possible implementation, the processing module 302 is specifically used to input each real partial discharge signal into the sample generation model. The sample generation model divides each real partial discharge signal into sliding time windows, reads the waveform data of each real partial discharge signal in different time windows, and the waveform data includes waveform, amplitude and phase angle. Based on the waveform data of each real partial discharge signal in different time windows, the kernel density estimate of each real partial discharge signal in different time windows is calculated. Based on the kernel density estimate of each real partial discharge signal in different time windows, the probability distribution of noise in the real partial discharge signal is determined.

[0159] In one possible implementation, the processing module 302 is specifically used to generate a noise-free simulated partial discharge signal by a sample generation model; the sample generation model generates a noise signal based on a probability distribution; and the noise-free simulated partial discharge signal and the noise signal are superimposed to obtain a noisy partial discharge signal.

[0160] In one possible implementation, the processing module 302 is specifically used to determine noisy partial discharge samples, which take noisy partial discharge signals as input and noiseless simulated partial discharge signals as output; based on the training set in the noisy partial discharge samples, the pre-set neural network model is iteratively trained using the backpropagation algorithm; during the training process, the neural network model is tested for each completed training based on the test set in the noisy partial discharge samples; if the loss function is less than a set value, or the training of the noisy partial discharge samples is completed, the iteration process is exited, and a denoised model is obtained. The loss function is used to characterize the difference between the partial discharge signal output by the neural network model and the noiseless simulated partial discharge signal.

[0161] In one possible implementation, the pre-set neural network model includes multiple convolutional layers; the processing module 302 is specifically used to, during training, calculate the attention information of each feature output by each convolutional layer; wherein the attention information of the feature is used to characterize the importance of the feature; determine the weight of each feature based on the attention information of each feature, wherein the attention information of each feature is different, and different attention information corresponds to different weights; and determine the input of the next convolutional layer based on each feature and the weight of each feature.

[0162] In one possible implementation, the processing module 302 is specifically used to calculate the loss function after each iteration based on the following formula;

[0163]

[0164] Where loss represents the value of the loss function, C j H represents the number of channels in the feature map of a neural network model. j W represents the length of the feature map in the neural network model. j This represents the width of the feature map in the neural network model. Represents the loss function. The loss network function representing a sharp image, This represents the loss network function of the denoised image.

[0165] In one possible implementation, the processing module 302 is further configured to determine a noiseless partial discharge sample based on a noiseless partial discharge signal and a partial discharge type corresponding to the noiseless partial discharge signal; train a pre-set partial discharge identification model based on the noiseless partial discharge sample to obtain a trained partial discharge identification model; and input the real partial discharge signal acquired in real time into the trained partial discharge identification model to obtain a partial discharge type corresponding to the real partial discharge signal acquired in real time.

[0166] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. For example... Figure 6 As shown, the electronic device 400 of this embodiment includes: a processor 401, a memory 402, and a computer program 403 stored in the memory 402 and executable on the processor 401. When the processor 401 executes the computer program 403, it implements the steps in the above-described method embodiments, for example... Figure 1 Steps 101 to 105 are shown. Alternatively, when the processor 401 executes the computer program 403, it implements the functions of each module / unit in the above-described device embodiments, for example... Figure 5 The functions of the communication module 301 and the processing module 302 shown are illustrated.

[0167] For example, the computer program 403 can be divided into one or more modules / units, which are stored in the memory 402 and executed by the processor 401 to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program 403 in the electronic device 400. For example, the computer program 403 can be divided into... Figure 5 The communication module 301 and the processing module 302 are shown.

[0168] The processor 401 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0169] The memory 402 can be an internal storage unit of the electronic device 400, such as a hard disk or memory of the electronic device 400. The memory 402 can also be an external storage device of the electronic device 400, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device 400. Furthermore, the memory 402 can include both internal and external storage units of the electronic device 400. The memory 402 is used to store the computer program and other programs and data required by the terminal. The memory 402 can also be used to temporarily store data that has been output or will be output.

[0170] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0171] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0172] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0173] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0174] If the integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0175] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A noise processing method for partial discharge signals, characterized in that, include: Acquire real partial discharge samples, wherein the real partial discharge samples include real partial discharge signals; The probability density estimation of the real partial discharge signal in each real partial discharge sample is performed to obtain the probability distribution of noise in the real partial discharge signal. This includes: inputting each real partial discharge signal into a sample generation model, which divides each real partial discharge signal into sliding time windows; reading waveform data of each real partial discharge signal in different time windows, the waveform data including waveform, amplitude, and phase angle; calculating kernel density estimates of each real partial discharge signal in different time windows based on the waveform data of each real partial discharge signal in different time windows; and determining the probability distribution of noise in the real partial discharge signal based on the kernel density estimates of each real partial discharge signal in different time windows. A noise signal conforming to the aforementioned probability distribution is added to a noise-free simulated partial discharge signal to obtain a noisy partial discharge signal. Based on noisy partial discharge signals and noiseless simulated partial discharge signals, denoised partial discharge samples are generated, and a pre-set neural network model is trained based on the denoised partial discharge samples to obtain a denoised model. Based on the trained denoising model, the real partial discharge signal acquired in real time is denoised to obtain a noise-free partial discharge signal.

2. The noise processing method for partial discharge signals according to claim 1, characterized in that, The step of adding a noise signal conforming to the probability distribution to a noise-free simulated partial discharge signal to obtain a noisy partial discharge signal includes: The sample generation model generates noise-free simulated partial discharge signals; The sample generation model generates a noise signal based on the probability distribution. By superimposing the noise-free simulated partial discharge signal and the noise signal, a noisy partial discharge signal is obtained.

3. The noise processing method for partial discharge signals according to claim 1, characterized in that, The process of generating denoised samples based on the noisy partial discharge signal and the noise-free simulated partial discharge signal, and training a pre-set neural network model based on the denoised samples to obtain a denoised model includes: A noisy partial discharge sample is determined, wherein the noisy partial discharge sample takes the noisy partial discharge signal as input and the noiseless simulated partial discharge signal as output; Based on the training set in the noisy partial discharge samples, the pre-set neural network model is iteratively trained using the backpropagation algorithm; During the training process, the neural network model is tested for each completed training session based on the test set in the noisy partial discharge samples. If the loss function is less than the set value, or the training of the noisy partial discharge sample is completed, the iteration process is exited, and the denoising model is obtained. The loss function is used to characterize the difference between the partial discharge signal output by the neural network model and the noise-free simulated partial discharge signal.

4. The noise processing method for partial discharge signals according to claim 3, characterized in that, The pre-configured neural network model includes multiple convolutional layers; The step of training a pre-set neural network model based on denoised partial discharge samples to obtain a denoised model further includes: During training, for each convolutional layer, the attention information of each feature in the output of that convolutional layer is calculated; whereby the attention information of a feature is used to characterize the importance of that feature. Based on the attention information of each feature, the weight of each feature is determined, wherein the attention information of each feature is different, and different attention information corresponds to different weights; Based on the features and their weights, the input to the next convolutional layer is determined.

5. The noise processing method for partial discharge signals according to claim 3, characterized in that, The step of training a pre-set neural network model based on denoised partial discharge samples to obtain a denoised model further includes: The loss function is calculated after each iteration based on the following formula; ； in, This represents the value of the loss function. This represents the number of channels in the feature map of a neural network model. This represents the length of the feature map in the neural network model. This represents the width of the feature map in the neural network model. Represents the loss function. The loss network function representing a sharp image, This represents the loss network function of the denoised image.

6. The noise processing method for partial discharge signals according to claim 1, characterized in that, The real partial discharge samples also include the partial discharge types corresponding to the real partial discharge signals; The denoising model, based on training, denoises the real-time acquired partial discharge signal to obtain a noise-free partial discharge signal, and then further includes: Based on the noise-free partial discharge signal and the partial discharge type corresponding to the noise-free partial discharge signal, a noise-free partial discharge sample is determined. Based on the noiseless partial discharge samples, the pre-set partial discharge identification model is trained to obtain the trained partial discharge identification model. The real partial discharge signal acquired in real time is input into the trained partial discharge recognition model to obtain the partial discharge type corresponding to the real partial discharge signal acquired in real time.

7. A noise processing device for partial discharge signals, characterized in that, include: The communication module is used to acquire real partial discharge samples, wherein the real partial discharge samples include real partial discharge signals; The processing module is used to estimate the probability density of the real partial discharge signal in each real partial discharge sample to obtain the probability distribution of noise in the real partial discharge signal; add a noise signal that conforms to the probability distribution to the noise-free simulated partial discharge signal to obtain a noisy partial discharge signal; generate denoised partial discharge samples based on the noisy partial discharge signal and the noise-free simulated partial discharge signal, and train a pre-set neural network model based on the denoised partial discharge samples to obtain a denoising model; and perform denoising processing on the real partial discharge signal acquired in real time based on the trained denoising model to obtain a noise-free partial discharge signal. The processing module is specifically used to input each real partial discharge signal into the sample generation model. The sample generation model divides each real partial discharge signal into sliding time windows, reads the waveform data of each real partial discharge signal in different time windows, and the waveform data includes waveform, amplitude and phase angle. Based on the waveform data of each real partial discharge signal in different time windows, the kernel density estimate of each real partial discharge signal in different time windows is calculated. Based on the kernel density estimate of each real partial discharge signal in different time windows, the probability distribution of noise of the real partial discharge signal is determined.

8. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor being configured to invoke and run the computer program stored in the memory to perform the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 6 above.

Images