An intelligent data processing method for a mutual inductor

Abstract

The present application belongs to the technical field of electric data processing, and particularly relates to a data processing method for intelligent mutual inductor, to solve the technical problem of poor accuracy of online evaluation result of the running state of the intelligent mutual inductor in the prior art, comprising the following steps: S1, obtaining time sequence sampling data of current or voltage output by the intelligent mutual inductor; S2, using an unscented Kalman filtering algorithm to denoise the time sequence sampling data to obtain first processed data; S3, windowing and truncating the first processed data, and obtaining a preliminary frequency spectrum through fast Fourier transform; S4, determining a spectral line screening threshold, and generating a corrected harmonic feature vector; S5, performing kernel principal component analysis on the corrected harmonic feature vector, extracting the maximum eigenvalue thereof as a nonlinear principal component energy, and taking the geometric mean of all eigenvalues as a kernel space generalized variance. The highly accurate frequency and amplitude of the harmonic and interharmonic spectral lines screened from the frequency spectrum are ensured.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of electric data processing, and particularly relates to a data processing method for an intelligent transformer. BACKGROUND

[0002] As a key measurement and protection device in the power system, the current or voltage signal output by the intelligent transformer is the direct basis for judging its health condition. In the actual operating environment, especially in the high-voltage substation, there are strong electromagnetic interference, inherent white noise of the system, and pulse noise caused by the operation of starting and stopping the equipment. These complex noise components can seriously pollute the original sampling signal, reduce the signal-to-noise ratio, and make the subsequent signal analysis and feature extraction difficult.

[0003] Traditional filtering methods, such as mean filtering or simple low-pass filtering, are difficult to effectively filter out such non-stationary, non-Gaussian mixed noise, and may cause loss of useful signal information. In addition, in order to perform frequency spectrum analysis, the fast Fourier transform (FFT) is usually used, but this requires windowing and truncation of the signal. The prior art usually uses a fixed type and length of window function, which cannot balance the frequency resolution and local time domain. When the signal contains both steady-state components and transient disturbances, serious spectral leakage and fence effect will occur, resulting in large deviations in the calculation of harmonic and interharmonic amplitudes and frequencies, affecting the accuracy of the analysis.

[0004] After the signal is preliminarily processed, how to accurately extract features from the complex frequency spectrum that can represent the operating state of the intelligent transformer and perform intelligent state recognition is another major challenge faced by current technology. Traditional fault diagnosis methods usually rely on simple statistics or linear features, however, the evolution process of the intelligent transformer from normal to abnormal state is often a complex nonlinear dynamic process, and its fault features may be hidden in the weak nonlinear coupling relationship between multiple harmonic components. Linear principal component analysis (PCA) and other methods cannot effectively mine such deep nonlinear structure information. In addition, in the construction of the state recognition model, support vector machines (SVM) or traditional feedforward neural networks and other methods have been applied, but they usually treat the features at each time point as independent samples, ignoring the continuous evolution rules and dynamic dependencies of the fault features in the time series. Although recurrent neural networks (RNN) can handle time series information, they are prone to gradient vanishing or explosion problems when processing long sequences, limiting their ability to capture long-term evolution trends of the device state. Therefore, the prior art has deficiencies in the depth of feature extraction and the dynamics of the classification model, and it is difficult to achieve high-precision and high-reliability online evaluation of the operating state of the intelligent transformer. SUMMARY

[0005] The application provides a data processing method for intelligent transformers to solve the technical problem of poor accuracy of online evaluation results of the operating state of intelligent transformers in the prior art.

[0006] A data processing method for intelligent transformers, comprising the following steps:

[0007] S1, obtaining time sequence sampling data of current or voltage output by the intelligent transformer;

[0008] S2, using an unscented Kalman filtering algorithm to denoise the time sequence sampling data to obtain first processed data, wherein in the unscented Kalman filtering algorithm, a noise covariance matrix is constructed online based on wavelet packet energy;

[0009] S3, adaptively adjusting window function parameters according to signal characteristics of the first processed data, windowing and truncating the first processed data, and obtaining a preliminary frequency spectrum through fast Fourier transform;

[0010] S4, determining a spectral line screening threshold based on a Teager-Kaiser energy operator calculated from the first processed data, screening effective harmonic and overtone spectral lines from the preliminary frequency spectrum using the spectral line screening threshold, correcting the frequencies and amplitudes of the screened spectral lines using a spectral line interpolation correction algorithm, and generating a corrected harmonic feature vector;

[0011] S5, performing kernel principal component analysis on the corrected harmonic feature vector, extracting the largest eigenvalue thereof as a nonlinear principal component energy, and taking the geometric mean of all eigenvalues as a kernel space generalized variance.

[0012] Further, in S1, the analog signal output by the secondary side of the intelligent transformer is collected at a set sampling frequency through a data acquisition card and is converted into discrete digital time sequence sampling data sequences through analog-digital conversion.

[0013] Further, in S2, when denoising the time sequence sampling data, the following steps are included:

[0014] The time sequence sampling data is decomposed by a multi-layer wavelet packet to obtain a group of frequency band sub-signals;

[0015] A preset high-frequency sub-signal is selected, the total energy thereof is calculated, and the total energy is taken as a noise variance estimation value at the current time;

[0016] An exponential weighted moving average method is used to update the noise variance estimation value at the current time according to the noise variance estimation value at the current time and the noise variance at the last time, and a process noise covariance diagonal matrix is constructed therefrom .

[0017] Further, in processing the time sequence sampling data of the secondary side output of the intelligent mutual inductor, the signal is subjected to 3-layer wavelet packet decomposition, and the original signal is decomposed into 8 frequency band sub-signals which are evenly divided from low to high.

[0018] Further, in the step S3, when the first processed data is windowed and truncated, the following steps are included:

[0019] The kurtosis value of the first processed data in the current analysis window is calculated.

[0020] If the kurtosis value is greater than the preset transient impulse threshold value, it is determined that the signal contains transient components, and a rectangular window with a window length of 2-4 power frequency periods is used for truncation.

[0021] If the kurtosis value is not greater than the preset transient impulse threshold value, it is determined that the signal is stable, and a Hanning window with a window length of 5-10 power frequency periods is used for truncation.

[0022] Further, in the step S4, when the selected spectral lines are corrected in frequency and amplitude, the following steps are included:

[0023] For the first processed data, the Fourier transform is first performed to obtain a preliminary frequency spectrum, the median of the amplitudes of all spectral lines in the preliminary frequency spectrum is calculated, and the median multiplied by k is set as a reference noise threshold value, wherein k is a preset constant; at the same time, the Teager-Kaiser energy operator is applied to the first processed data for calculation.

[0024] The mean value of the instantaneous energy values calculated by the Teager-Kaiser energy operator is used to determine a dynamic adjustment coefficient, and the dynamic adjustment coefficient is used to correct the reference noise threshold value to generate a final spectral line selection threshold value.

[0025] The three-point parabola or Gaussian interpolation method is used to interpolate each selected spectral line and its left and right adjacent spectral lines to correct their frequency and amplitude.

[0026] Further, the formula of the Teager-Kaiser energy operator is: ;

[0027] wherein, is the instantaneous energy value of the i-th sampling point calculated by the Teager-Kaiser energy operator; is the signal value at the current time, is the signal value at the previous time, is the signal value at the next time.

[0028] Further, in the step S5, when the corrected harmonic feature vector is subjected to kernel principal component analysis, the following steps are included:

[0029] ​The radial basis function is selected as a kernel function of kernel principal component analysis;

[0030] The width parameter of the radial basis function is determined by automatically searching for optimization in a preset parameter range through a method combining grid search and cross validation.

[0031] Further, the gated recurrent unit network comprises an input layer, at least one gated recurrent unit hidden layer, and a fully connected output layer adopting a Softmax activation function, and the number of neurons of the fully connected output layer matches the number of preset intelligentized mutual inductor operating state categories.

[0032] Further, when judging the current operating state of the intelligentized mutual inductor through the network output, the maximum value in the output probability values of each neuron of the output layer of the gated recurrent unit network is taken, and the index corresponding to the maximum value is mapped as the current operating state of the intelligentized mutual inductor; wherein the categories of the current operating state of the intelligentized mutual inductor at least include: a normal operating state, a magnetic circuit saturation state and a transient fault state.

[0033] The beneficial effects are: compared with the prior art, the present application can effectively filter out strong electromagnetic interference, white noise, impulse noise and other complex mixed noise by adopting the unscented Kalman filter based on the wavelet packet energy to construct the noise covariance matrix, in the frequency spectrum analysis stage, the window function parameters are optimized according to the signal content, the frequency spectrum leakage and the fence effect are suppressed, and the Teager-Kaiser energy operator and the spectral line interpolation correction are combined, so that the high accuracy of the harmonic and the interharmonic spectral line in the frequency and the amplitude is ensured. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 The flowchart of the data processing method for the intelligentized mutual inductor. DETAILED DESCRIPTION

[0035] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application, and those skilled in the art should know that the embodiments described below are part of the embodiments of the present disclosure, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0036] The embodiments of the data processing method for the intelligentized mutual inductor provided by the present application are as follows:

[0037] As shown in Figure 1 A data processing method for an intelligentized mutual inductor, comprising the following steps:

[0038] S1, acquiring time sequence sampling data of current or voltage output by the intelligentized mutual inductor.

[0039] In an optional embodiment, the analog signal outputted by the intelligent mutual inductor on the secondary side is collected by a data acquisition card at a set sampling frequency (e.g. 10 kHz) and is subjected to analog-digital conversion to form a discrete digital time-series sampling data sequence wherein is the sampling point number.

[0040] S2, the time-series sampling data is subjected to noise reduction by a unscented Kalman filter algorithm to obtain first processed data, in which a noise covariance matrix is constructed on-line based on wavelet packet energy.

[0041] In an optional embodiment, when the time-series sampling data is subjected to noise reduction, the following steps are included:

[0042] The time-series sampling data is subjected to multi-layer wavelet packet decomposition to obtain a group of frequency band sub-signals;

[0043] A preset high frequency band sub-signal is selected, the total energy thereof is calculated and the total energy is taken as the noise variance estimation value at the current time;

[0044] An exponential weighted moving average method is adopted to update the noise variance estimation value at the current time according to the noise variance at the current time and the noise variance at the last time, and the noise variance estimation value at the current time is used to construct the process noise covariance diagonal matrix .

[0045] Specifically, when the time-series sampling data outputted by the intelligent mutual inductor on the secondary side is processed, the signal is subjected to 3-layer wavelet packet decomposition to decompose the original signal into 8 frequency band sub-signals which are evenly divided from low to high. For example, for a system with a sampling rate of 10 kHz, the frequency range of the 8th frequency band sub-signal is 4.375-5 kHz. Generally, the signal in the high frequency band is mainly composed of random interference such as Gaussian white noise, therefore, one or several sub-signals in the highest frequency band, such as the 7th and 8th frequency band sub-signals, can be selected, the square sum of the signal coefficients thereof is calculated to obtain the energy value at the current time, for example, 0.02 is calculated, which is regarded as the noise variance estimation value at the current time.

[0046] In order to smooth the noise variance estimation value and avoid sharp fluctuations caused by instantaneous disturbance, an exponential weighted moving average method is introduced. A weighting factor (e.g. 0.2) is set, the noise variance obtained by iteration at the last time is 0.025; the noise variance estimation value at the current time is 0.02, and the updated noise variance estimation value at the current time is 0.2 x 0.02 + (1-0.2) x 0.025 = 0.024. The updated noise variance estimation value (0.024) is used to construct the process noise covariance diagonal matrix required by the unscented Kalman filter algorithm where all the diagonal elements are 0.024 and the non-diagonal elements are 0. This online updating mechanism enables the UKF algorithm to adaptively track the change of the system background noise, thus achieving effective noise reduction under different working conditions.

[0047] S3, adaptively adjusting the window function parameters according to the signal characteristics of the first processed data, performing windowing and truncation on the first processed data, and obtaining a preliminary frequency spectrum via fast Fourier transform.

[0048] In an optional embodiment, when performing windowing and truncation on the first processed data, the following steps are included:

[0049] calculating the kurtosis value of the first processed data within the current analysis window;

[0050] If the kurtosis value is greater than the preset transient impulse threshold, it is determined that the signal contains a transient component, and a rectangular window with a window length of 2-4 power frequency periods is used for truncation.

[0051] If the kurtosis value is not greater than the preset transient impulse threshold, it is determined that the signal is stable, and a Hanning window with a window length of 5-10 power frequency periods is used for truncation.

[0052] Specifically, in order to balance the contradiction between time resolution and frequency resolution in spectral analysis, the kurtosis value is a statistical quantity that measures the steepness of the signal waveform. For a normal distribution of stationary signals, the kurtosis value is close to 3, so 5 can be set as the transient impulse threshold. The reason for setting 5 as the transient impulse threshold is that the peak value of the transient impulse must be greater than 3, but the specific threshold needs to consider both "avoiding missed judgment" and "reducing false positives". If the threshold is set too low, normal signals with slight fluctuations may be misjudged as transient impulses, resulting in poor anti-interference performance. If the threshold is set too high, some weak but real transient impulses may be missed, resulting in insufficient sensitivity. Selecting 5 as the threshold is based on empirical judgment.

[0053] When a short-circuit fault occurs or a large operating overvoltage appears, the current signal will have a sharp transient impulse, and the calculated kurtosis value may reach 8 or higher. When the data kurtosis value within the current analysis window is 8, which exceeds the preset threshold of 5, it is determined that the signal is a transient signal. Once it is determined to be a transient signal, in order to accurately capture the time point and shape of the impulse, a short window is selected, for example, for a 50Hz power frequency system, a window length of 3 power frequency periods, i.e. 60ms, is selected, and a rectangular window is used. The rectangular window has no amplitude decay in the time domain, and can best preserve the amplitude information of the transient impulse.

[0054] Conversely, if the calculated kurtosis value is 2.8, which is below the threshold of 5, the signal is determined to be a steady-state signal. In this case, in order to obtain high-precision harmonic analysis results and suppress spectral leakage, a long window will be automatically switched, for example, selecting a window length of 8 power frequency cycles, i.e., 160ms, and using the Hanning window, which has better sidelobe attenuation performance.

[0055] The adaptive windowing strategy ensures optimal analysis results under different signal conditions.

[0056] S4. Based on the Teager-Kaiser energy operator calculated from the first processed data, a spectral line screening threshold is determined. This spectral line screening threshold is used to screen effective harmonic and interharmonic spectral lines from the preliminary spectrum. The selected spectral lines are then corrected for frequency and amplitude using a spectral line interpolation correction algorithm to generate corrected harmonic feature vectors.

[0057] In an optional embodiment, frequency and amplitude correction of the selected spectral lines includes the following steps:

[0058] For the first processed data, a preliminary spectrum is obtained by Fourier transform. The median of the amplitude of all spectral lines in the preliminary spectrum is calculated, and k times this median is set as the reference noise threshold, where k is a preset constant. Simultaneously, the Teager-Kaiser energy operator is applied to the first processed data for calculation, with the following formula: ;

[0059] in, The first energy obtained by the Teager-Kaiser energy operator The instantaneous energy value of each sampling point; The signal value at the current moment. The signal value at the previous moment. This represents the signal value at the next moment.

[0060] The dynamic adjustment coefficient is determined based on the mean of the instantaneous energy values ​​calculated by the Teager-Kaiser energy operator, and the reference noise threshold is corrected using the dynamic adjustment coefficient to generate the final spectral screening threshold.

[0061] The frequency and amplitude of each selected spectral line and its left and right adjacent spectral lines are interpolated using the three-point parabolic or Gaussian interpolation method.

[0062] Specifically, in order to effectively distinguish the real harmonic components and background noise, a dynamic threshold that can adapt to signal changes is needed. Specifically, for the first processed data after noise reduction, an energy sequence is obtained by applying the Teager-Kaiser energy operator. At the same time, the preliminary frequency spectrum is obtained by Fourier transform on the first processed data, and the median of the amplitudes of all spectral lines is calculated, assuming that the obtained median value is 0.002A. Set an empirical constant k as 4 (select 4 as the empirical constant, which is the experience selection of balancing "missed detection" and "false judgment"), then the reference noise threshold is set to 0.008A. Specifically, then the average value of the energy sequence is calculated. When the signal is stable, the average value is small, for example, 0.5, and the system generates a dynamic adjustment coefficient accordingly, such as 1.0. At this time, the final threshold is the reference threshold 0.008A. When the signal appears transient impact or strong harmonic distortion, the average value of the energy sequence will increase significantly, for example, to 2.0, and the dynamic adjustment coefficient will also increase to 1.5 accordingly. The final threshold is corrected to 0.012A. A higher threshold can effectively avoid the high-frequency components in the transient process being misjudged as harmonics. After screening out the effective spectral lines with the final threshold, for example, a spectral line with an amplitude of 0.5A, the frequency of which is corrected from the discrete frequency spectrum of 249.5Hz to the more accurate 250.1Hz, and the amplitude is corrected from 0.49A to 0.51A, thereby eliminating the error caused by the fence effect.

[0063] S5, performing kernel principal component analysis on the corrected harmonic feature vector, extracting the largest eigenvalue as the nonlinear principal component energy, and taking the geometric mean of all eigenvalues as the kernel space generalized variance.

[0064] In an optional embodiment, when performing kernel principal component analysis on the corrected harmonic feature vector, the following steps are included:

[0065] Radial basis function (RBF) is selected as the kernel function of kernel principal component analysis;

[0066] Through the method of grid search combined with cross-validation, the width parameter of the radial basis function is automatically optimized in the preset parameter range.

[0067] Specifically, the traditional linear principal component analysis is difficult to process the harmonics generated by the nonlinear process of intelligent mutual inductor saturation. Therefore, kernel principal component analysis is adopted, and the original harmonic feature vector (such as the vector containing the fundamental, 3rd, 5th, and 7th harmonic amplitudes and phases) is mapped to a high-dimensional feature space through a kernel function. The purpose is to make the data of different operating states linearly separable in this high-dimensional feature space.

[0068] The form of the kernel function determines how the similarity between sample points is measured, thus affecting the mapping effect. The radial basis function is selected because it can handle complex nonlinear relationships and has only one key parameter . The width parameter of the radial basis function is crucial to the performance of kernel principal component analysis. A too small will cause the model to overfit, while a too large will make the model underfit and unable to distinguish samples. In order to find the best sigma value, a grid search combined with cross-validation is used.

[0069] First, set the search range of (e.g. 0.01 to 1000) and select a series of candidate values at 10 power intervals, such as 0.01, 0.1, 1, 10, 100, 1000. For each candidate value, use 5-fold cross-validation to evaluate its performance on the training data. The evaluation index can be the class separation degree of the features after dimensionality reduction or the accuracy of the subsequent classifier. Finally, the value (e.g. 10) with the highest average performance score in cross-validation is selected as the final width parameter.

[0070] S6, the nonlinear principal component energy, the kernel space generalized variance and their first order difference values with respect to time are combined into a four-dimensional feature vector sequence, and the four-dimensional feature vector sequence is input into a pre-trained gated recurrent unit network to output the current running state of the intelligent current transformer through the network.

[0071] Specifically, at each analysis time window , the current nonlinear principal component energy is denoted as , and the kernel space generalized variance is denoted as . Their first order difference values with respect to the corresponding values of the previous time window are calculated as , , where is the nonlinear principal component energy of the previous time window , is the kernel space generalized variance of the previous time window . The , , , The four values are combined into a four-dimensional feature vector; a sequence of four-dimensional feature vectors is formed over time, and the sequence of four-dimensional feature vectors is input into a gated recurrent unit network that has been trained through historical normal and fault data. The final output layer (for example, a fully connected layer plus a Softmax activation function) of the gated recurrent unit network outputs probability values corresponding to preset states such as normal, slight fault, and serious fault. The state with the highest probability value is taken as the current operating state of the intelligent mutual inductor.

[0072] In an optional embodiment, the gated recurrent unit network includes an input layer, at least one gated recurrent unit hidden layer, and a fully connected output layer with a Softmax activation function. The number of neurons in the fully connected output layer matches the number of preset intelligent mutual inductor operating states.

[0073] The operating state of the intelligent mutual inductor, especially the evolution process from normal to saturation, has obvious time dependence. The harmonic features at a single time point are not enough to make an accurate judgment. Therefore, a gated recurrent unit network, which is a type of recurrent neural network, is used to process the time sequence composed of the feature vectors output by the kernel principal component analysis. For example, a feature vector is generated every 20 ms, and the next 10 feature vectors form a time sequence that is input into the network, enabling it to learn the dynamic pattern of state changes.

[0074] In an optional embodiment, the number of nodes in the input layer of the gated recurrent unit network is the same as the dimension of the input feature vector (for example, 15 features after dimension reduction).

[0075] A gated recurrent unit hidden layer containing 64 neurons is connected to the input layer to capture short-term and long-term dependencies in the time sequence. The output of the hidden layer is sent to a fully connected output layer. If the states to be diagnosed are divided into normal operation, magnetic circuit saturation, and transient fault, the output layer is set to have 3 neurons. A Softmax activation function is applied to the 3 neurons, which can convert the original values of the output layer into a probability distribution. For example, the output [0.1, 0.8, 0.1] represents the probabilities of the current sequence being judged as normal, saturated, and transient fault, respectively.

[0076] In an optional embodiment, when the current operating state of the intelligent mutual inductor is determined through the network output, the maximum value of the output probability values of each neuron in the output layer of the gated recurrent unit network is taken, and the index corresponding to the maximum value is mapped to the current operating state of the intelligent mutual inductor. The types of the current operating state of the intelligent mutual inductor include at least normal operation, magnetic circuit saturation, and transient fault.

[0077] Specifically, the output layer of the Gated Recurrent Unit network provides a probability estimation for each preset operating state. For example, assume that the three output neurons of the Gated Recurrent Unit network correspond to the normal operation, magnetic path saturation and transient fault states in sequence. At a certain time, after the Gated Recurrent Unit network processes a certain sequence of feature vectors, the output probability vector is [0.05, 0.92, 0.03]. This means that the model predicts that the current smart mutual inductor has a 5% probability of being in a normal state, a 92% probability of being in a magnetic path saturation state, and a 3% probability of being in a transient fault state. In order to give a clear diagnosis conclusion, the maximum probability principle is adopted. In the above example, 92% is the maximum value among the three probabilities, and the neuron index corresponding to the maximum value is 1, that is, the second neuron. According to the pre-set mapping relationship, the second neuron represents the magnetic path saturation state. Therefore, the final determination of the current operating state of the smart mutual inductor is the magnetic path saturation.

[0078] In addition, in the description of the present specification, the meaning of "a plurality of" is at least two, such as two, three or more, etc., unless otherwise explicitly specifically limited.

Claims

1. A data processing method for an intelligent instrument transformer, characterized in that, Includes the following steps: S1, acquire timing sampling data of the current or voltage output by the intelligent current transformer; S2, the unscented Kalman filter algorithm is used to denoise the time-series sampled data to obtain the first processed data. In the unscented Kalman filter algorithm, the noise covariance matrix is ​​constructed online based on the wavelet packet energy. S3, adaptively adjust the window function parameters according to the signal characteristics of the first processed data, window and truncate the first processed data, and obtain the preliminary spectrum through fast Fourier transform; S4. Based on the Teager-Kaiser energy operator calculated from the first processed data, a spectral line screening threshold is determined. The spectral line screening threshold is used to screen effective harmonic and interharmonic spectral lines from the preliminary spectrum. The spectral line interpolation correction algorithm is used to correct the frequency and amplitude of the screened spectral lines to generate the corrected harmonic feature vector. S5. Perform kernel principal component analysis on the corrected harmonic eigenvectors, extract the largest eigenvalue as the nonlinear principal component energy, and take the geometric mean of all eigenvalues ​​as the kernel space generalized variance. S6 combines the nonlinear principal component energy, the kernel space generalized variance, and their respective first-order differences over time into a four-dimensional feature vector sequence, and inputs this four-dimensional feature vector sequence into a pre-trained gated recurrent unit network. The network output is used to determine the current operating status of the intelligent current transformer. The gated recurrent unit network includes an input layer, at least one gated recurrent unit hidden layer, and a fully connected output layer using the Softmax activation function. The number of neurons in the fully connected output layer matches the number of preset intelligent transformer operating state types. When determining the current operating state of the intelligent current transformer through network output, the maximum value of the output probability value of each neuron in the output layer of the gated recurrent unit network is taken, and its corresponding index is mapped to the current operating state of the intelligent current transformer; among them, the types of the current operating state of the intelligent current transformer include at least: normal operating state, magnetic circuit saturation state, and transient fault state.

2. The data processing method for intelligent instrument transformers according to claim 1, characterized in that, In S1, the analog signal output from the secondary side of the intelligent current transformer is acquired by the data acquisition card at a set sampling frequency, and then converted from analog to digital to form a discrete digital time-series sampling data sequence.

3. The data processing method for intelligent instrument transformers according to claim 1, characterized in that, In S2, the noise reduction of time-series sampled data includes the following steps: Multi-level wavelet packet decomposition is performed on the time-series sampled data to obtain a set of frequency band sub-signals; Select a preset high-frequency segment signal, calculate its total energy, and use the total energy as the noise variance estimate at the current moment; An exponentially weighted moving average method is used to update the noise variance estimate at the current time step based on the noise variance estimate at the previous time step, and this update is used to construct the process noise covariance diagonal matrix. .

4. The data processing method for intelligent instrument transformers according to claim 3, characterized in that, When processing the timing sampling data output from the secondary side of the intelligent instrument transformer, the signal is decomposed into 8 frequency band sub-signals that are evenly divided from low to high by performing 3-level wavelet packet decomposition.

5. The data processing method for intelligent instrument transformers according to claim 1, characterized in that, In S3, when windowing and truncating the first processed data, the following steps are included: Calculate the kurtosis value of the first-processed data within the current analysis window; If the kurtosis value is greater than the preset transient impact threshold, the signal is determined to contain transient components, and a rectangular window with a window length of 2-4 power frequency cycles is used for truncation. If the kurtosis value is not greater than the preset transient impact threshold, the signal is determined to be in a steady state, and a Hanning window with a window length of 5-10 power frequency cycles is used for truncation.

6. The data processing method for intelligent instrument transformers according to claim 1, characterized in that, In S4, when performing frequency and amplitude correction on the selected spectral lines, the following steps are included: For the first processed data, a preliminary spectrum is obtained by Fourier transform. The median of the amplitude of all spectral lines in the preliminary spectrum is calculated and k times it is set as the reference noise threshold, where k is a preset constant. At the same time, the Teager-Kaiser energy operator is applied to the first processed data for calculation. The dynamic adjustment coefficient is determined based on the mean of the instantaneous energy values ​​calculated by the Teager-Kaiser energy operator, and the reference noise threshold is corrected using the dynamic adjustment coefficient to generate the final spectral screening threshold. The frequency and amplitude of each selected spectral line and its left and right adjacent spectral lines are interpolated using the three-point parabolic or Gaussian interpolation method.

7. The data processing method for intelligent instrument transformers according to claim 6, characterized in that, The formula for the Teager-Kaiser energy operator is: ; in, The first energy obtained by the Teager-Kaiser energy operator The instantaneous energy value of each sampling point; The signal value at the current moment. The signal value at the previous moment. This represents the signal value at the next moment.

8. The data processing method for intelligent instrument transformers according to claim 1, characterized in that, In S5, when performing kernel principal component analysis on the corrected harmonic eigenvectors, the following steps are included: Radial basis functions were selected as the kernel functions for kernel principal component analysis. By combining grid search and cross-validation, the method automatically optimizes within a preset parameter range to determine the width parameter of the radial basis function.

Images