Data processing method for intelligent mutual inductor

Abstract

The invention belongs to the technical field of electric data processing, and particularly relates to a data processing method for an intelligent mutual inductor to solve the technical problem that in the prior art, the accuracy of an online evaluation result of the operation state of the intelligent mutual inductor is poor, and the method comprises the following steps: S1, obtaining time sequence sampling data of current or voltage output by the intelligent mutual inductor; s2, performing noise reduction on the time sequence sampling data by adopting an unscented Kalman filtering algorithm to obtain first processed data; s3, windowing truncation is carried out on the data after the first processing, and a preliminary frequency spectrum is obtained through fast Fourier transform; s4, determining a spectral line screening threshold, and generating a corrected harmonic feature vector; and S5, performing kernel principal component analysis on the corrected harmonic feature vector, extracting the maximum feature value as nonlinear principal component energy, and taking the geometric mean of all the feature values as a kernel space generalized variance. And the high accuracy of the harmonic and inter-harmonic spectral lines screened from the frequency spectrum in the aspects of frequency and amplitude is ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of electrical data processing, and in particular relates to a data processing method for an intelligent mutual inductor. Background Art

[0002] Intelligent instrument transformers are key measurement and protection devices in power systems. Their output current or voltage signals are the direct basis for determining their health. In real-world operating environments, particularly high-voltage substations, strong electromagnetic interference, inherent system white noise, and impulse noise caused by equipment startup and shutdown operations are present. These complex noise components can severely contaminate the original sampled signal, reducing the signal-to-noise ratio and complicating subsequent signal analysis and feature extraction.

[0003] Traditional filtering methods, such as mean filtering or simple low-pass filtering, are unable to effectively filter out this non-stationary, non-Gaussian mixed noise and may result in the loss of useful signal information. Furthermore, for spectral analysis, the Fast Fourier Transform (FFT) is often used, but this requires windowing and truncation of the signal. Existing technologies often use window functions of fixed type and length, which cannot balance frequency resolution and time-domain locality. When the signal contains both steady-state components and transient disturbances, severe spectral leakage and fence effects can occur, leading to large deviations in the calculation of harmonic and interharmonic amplitudes and frequencies, affecting the accuracy of the analysis.

[0004] After initial signal processing, another major challenge facing current technology is how to accurately extract features that characterize the operating status of intelligent transformers from the complex spectrum and perform intelligent state identification. Traditional fault diagnosis methods typically rely on simple statistics or linear features. However, the evolution of an intelligent transformer from normal to abnormal state is often a complex nonlinear dynamic process, and its fault characteristics may be hidden in the weak nonlinear coupling relationships between multiple harmonic components. Methods such as linear principal component analysis (PCA) are difficult to effectively mine this deep nonlinear structural information. Furthermore, while methods such as support vector machines (SVMs) or traditional feedforward neural networks have been applied to construct state identification models, they typically treat the features at each time point as independent samples, ignoring the continuous evolution and dynamic dependencies of fault characteristics over time series. Although recurrent neural networks (RNNs) can process time series information, they are prone to vanishing or exploding gradients when processing long sequences, limiting their ability to capture long-term trends in device state evolution. Therefore, the existing technology has deficiencies in the depth of feature extraction and the dynamics of the classification model, making it difficult to achieve high-precision and high-reliability online evaluation of the operating status of intelligent transformers. Summary of the Invention

[0005] The present invention provides a data processing method for an intelligent transformer, so as to solve the technical problem in the prior art that the accuracy of online evaluation results of the operating status of the intelligent transformer is relatively poor.

[0006] A data processing method for an intelligent mutual inductor comprises the following steps: S1, obtaining the time series sampling data of the current or voltage output by the intelligent transformer; S2, using an unscented Kalman filter algorithm to reduce noise on the time series sampling data to obtain first processed data, in which the noise covariance matrix is ​​constructed online based on wavelet packet energy; S3, adaptively adjusting window function parameters according to signal characteristics of the first processed data, performing windowing and truncation on the first processed data, and obtaining a preliminary spectrum through fast Fourier transform; S4, determining a spectral line screening threshold based on a Teager-Kaiser energy operator calculated from the first processed data, using the spectral line screening threshold to screen effective harmonic and interharmonic spectral lines from the preliminary spectrum, and using a spectral line interpolation correction algorithm to perform frequency and amplitude correction on the screened spectral lines to generate a corrected harmonic feature vector; S5, performing kernel principal component analysis on the corrected harmonic eigenvector, extracting its largest eigenvalue as the nonlinear principal component energy, and taking the geometric mean of all eigenvalues ​​as the kernel space generalized variance.

[0007] Furthermore, in S1, the analog signal output from the secondary side of the intelligent transformer is collected by a data acquisition card at a set sampling frequency, and analog-to-digital conversion is performed to form a discrete digital time-series sampling data sequence.

[0008] Furthermore, in S2, when performing noise reduction on the time series sampling data, the following steps are included: Perform multi-layer wavelet packet decomposition on the time series sampling data to obtain a set of frequency band sub-signals; Select a preset high-frequency sub-signal, calculate its total energy and use the total energy as the noise variance estimate at the current moment; The exponentially weighted moving average method is used to update the current noise variance estimate based on the current noise variance estimate and the previous noise variance, and the process noise covariance diagonal matrix is ​​constructed based on this. .

[0009] Furthermore, when processing the time series sampling data outputted from the secondary side of the intelligent transformer, the signal is subjected to three-layer wavelet packet decomposition, and the original signal is decomposed into eight frequency band sub-signals evenly divided from low to high.

[0010] Furthermore, in S3, when windowing and truncating the first processed data, the following steps are included: Calculating the kurtosis value of the first processed data within the current analysis window; If the kurtosis value is greater than the preset transient shock threshold, the signal is determined to contain a transient component and is truncated using a rectangular window with a window length of 2-4 power frequency cycles; If the kurtosis value is not greater than the preset transient shock threshold, the signal is determined to be steady state and truncated using a Hanning window with a window length of 5-10 power frequency cycles.

[0011] Furthermore, in S4, when performing frequency and amplitude correction on the screened spectral lines, the following steps are included: For the first processed data, a preliminary spectrum is first obtained by Fourier transform. The median of the amplitudes of all spectral lines in the preliminary spectrum is calculated and k times the median is set as the baseline 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 according to 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 line screening threshold; The three-point parabola or Gaussian interpolation method is used to interpolate each filtered spectral line and its left and right adjacent spectral lines to correct their frequency and amplitude.

[0012] Furthermore, the formula of the Teager-Kaiser energy operator is: ; in, is the first energy calculated by the Teager-Kaiser energy operator The instantaneous energy value of each sampling point; is the signal value at the current moment, is the signal value at the previous moment, is the signal value at the next moment.

[0013] Furthermore, in S5, when performing kernel principal component analysis on the corrected harmonic eigenvector, the following steps are included: The radial basis function is selected as the kernel function of kernel principal component analysis; By combining grid search with cross validation, the width parameter of the radial basis function is determined by automatically optimizing within the preset parameter range.

[0014] Furthermore, the gated recurrent unit network includes an input layer, at least one gated recurrent unit hidden layer, and a fully connected output layer using a Softmax activation function. The number of neurons in the fully connected output layer matches the number of preset intelligent transformer operating state types.

[0015] Furthermore, when judging the current operating state of the intelligent transformer through 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 its corresponding index is mapped to the current operating state of the intelligent transformer; wherein, the types of the current operating state of the intelligent transformer include at least: normal operating state, magnetic circuit saturation state and transient fault state.

[0016] The beneficial effects are: compared with the existing technology, the present invention can effectively filter out complex mixed noises such as strong electromagnetic interference, white noise and impulse noise by adopting an unscented Kalman filter that constructs a noise covariance matrix based on wavelet packet energy. In the spectrum analysis stage, the window function parameters are optimized according to the signal content, the spectrum leakage and the fence effect are suppressed, and the Teager-Kaiser energy operator is combined with the spectrum line interpolation correction to ensure the high accuracy of the harmonic and interharmonic spectrum lines screened out from the spectrum in terms of frequency and amplitude. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The figure is a flow chart of a data processing method for an intelligent transformer. DETAILED DESCRIPTION

[0018] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 of the present invention, all other embodiments obtained by those skilled in the art without making creative efforts shall fall within the scope of protection of the present invention.

[0019] The embodiment of the data processing method for intelligent mutual inductor provided by the present invention is as follows: like Figure 1 As shown, a data processing method for an intelligent mutual inductor includes the following steps: S1, obtains the time series sampling data of the current or voltage output by the intelligent transformer.

[0020] In an optional embodiment, the analog signal output from the secondary side of the intelligent transformer is collected by a data acquisition card at a set sampling frequency (such as 10kHz), and analog-to-digital conversion is performed to form a discrete digital time-series sampling data sequence. ,in is the sampling point number.

[0021] S2, using an unscented Kalman filter algorithm to reduce noise on the time series sampling data to obtain first processed data. In the unscented Kalman filter algorithm, a noise covariance matrix is ​​constructed online based on wavelet packet energy.

[0022] In an optional embodiment, when performing noise reduction on time series sampling data, the following steps are included: Perform multi-layer wavelet packet decomposition on the time series sampling data to obtain a set of frequency band sub-signals; Select a preset high-frequency sub-signal, calculate its total energy and use the total energy as the noise variance estimate at the current moment; The exponentially weighted moving average method is used to update the current noise variance estimate based on the current noise variance estimate and the previous noise variance, and the process noise covariance diagonal matrix is ​​constructed based on this. .

[0023] Specifically, when processing the time-series sampling data output from the secondary side of the intelligent transformer, the signal is subjected to a three-layer wavelet packet decomposition, decomposing the original signal into eight frequency band sub-signals evenly divided from low to high. For example, for a system with a sampling rate of 10kHz, the frequency range of the eighth frequency band sub-signal is 4.375-5kHz. Typically, high-frequency signals are mainly composed of random interference such as Gaussian white noise. Therefore, one or several sub-signals with the highest frequency, such as the seventh and eighth frequency band sub-signals, can be selected, and the sum of the squares of their signal coefficients can be calculated to obtain the energy value at that moment. For example, 0.02 is calculated, which is regarded as the noise variance estimate at the current moment.

[0024] In order to smooth the noise variance estimate and avoid the violent fluctuation caused by instantaneous disturbance, the exponentially weighted moving average method is introduced. Set a weighting factor (For example, take 0.2). The noise variance obtained by iterative update at the previous moment is 0.025; the noise variance estimate at the current moment is 0.02, and the updated noise variance estimate at the current moment is 0.2×0.02+(1-0.2)×0.025=0.024. The updated noise variance estimate (0.024) is used to construct the process noise covariance diagonal matrix required by the unscented Kalman filter algorithm. , where all diagonal elements are 0.024 and all off-diagonal elements are 0. This online update mechanism enables the unscented Kalman filter algorithm to adaptively track changes in the system background noise, thereby achieving effective noise reduction under different working conditions.

[0025] S3, adaptively adjusting window function parameters according to signal characteristics of the first processed data, performing windowing and truncation on the first processed data, and obtaining a preliminary spectrum through fast Fourier transform.

[0026] In an optional embodiment, when windowing and truncating the first processed data, the following steps are included: Calculating the kurtosis value of the first processed data within the current analysis window; If the kurtosis value is greater than the preset transient shock threshold, the signal is determined to contain a transient component and is truncated using a rectangular window with a window length of 2-4 power frequency cycles; If the kurtosis value is not greater than the preset transient shock threshold, the signal is determined to be steady state and truncated using a Hanning window with a window length of 5-10 power frequency cycles.

[0027] Specifically, to balance the conflict between time resolution and frequency resolution in spectrum analysis. Kurtosis is a statistic that measures the steepness of a signal waveform. For a normally distributed stationary signal, the kurtosis value is close to 3. Therefore, a value of 5 can be set as the transient impact threshold. The reason for setting 5 as the transient impact threshold is that the peak value of a transient impact is necessarily greater than 3. However, the specific threshold value must be chosen to balance "avoiding missed detections" and "reducing false detections." If the threshold is set too low, a slight increase in kurtosis caused by slight fluctuations in a normal signal may be mistakenly identified as a transient impact, resulting in poor anti-interference performance. If the threshold is set too high, some weak but real transient impacts may be missed, resulting in insufficient sensitivity. The choice of 5 as the threshold is based on empirical judgment.

[0028] When a short circuit occurs or a large switching overvoltage occurs, the current signal experiences a sharp transient surge, resulting in a sharp waveform and a calculated kurtosis value of 8 or higher. When the kurtosis value of the data within the current analysis window exceeds a preset threshold of 5, the signal is identified as a transient. Once a transient signal is identified, a short window is selected to accurately capture the timing and shape of the surge. For example, for a 50Hz power frequency system, a window length of three power frequency cycles, or 60ms, is selected, using a rectangular window. A rectangular window has no amplitude attenuation in the time domain and best preserves the amplitude information of the transient surge.

[0029] Conversely, if the calculated kurtosis value is 2.8, which is lower than the threshold of 5, the signal is considered steady-state. To obtain high-precision harmonic analysis results and suppress spectrum leakage, the system automatically switches to a long window, for example, a window length of 8 power frequency cycles (160ms), and uses a Hanning window for better sidelobe attenuation.

[0030] The adaptive windowing strategy ensures the best analysis effect under different signal conditions.

[0031] S4, based on the Teager-Kaiser energy operator calculated from the first processed data, determines the spectral line screening threshold, uses the spectral line screening threshold to screen out effective harmonic and interharmonic spectral lines from the preliminary spectrum, uses the spectral line interpolation correction algorithm to correct the frequency and amplitude of the screened spectral lines, and generates a corrected harmonic feature vector.

[0032] In an optional embodiment, when performing frequency and amplitude correction on the screened spectral lines, the following steps are included: For the first processed data, a preliminary spectrum is obtained by Fourier transform. The median of all spectral line amplitudes in the preliminary spectrum is calculated and k times the median is set as the baseline 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, and the formula is: ; in, is the first energy calculated by the Teager-Kaiser energy operator The instantaneous energy value of each sampling point; is the signal value at the current moment, is the signal value at the previous moment, is the signal value at the next moment.

[0033] The dynamic adjustment coefficient is determined according to 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 line screening threshold; The three-point parabola or Gaussian interpolation method is used to interpolate each filtered spectral line and its left and right adjacent spectral lines to correct their frequency and amplitude.

[0034] Specifically, to effectively distinguish true harmonic components from background noise, a dynamic threshold that can adapt to signal fluctuations is required. Specifically, the Teager-Kaiser energy operator is applied to the first processed data after noise reduction to obtain an energy sequence. Simultaneously, a Fourier transform is performed on the first processed data to obtain a preliminary spectrum, and the median of all spectral line amplitudes is calculated. Assume that the median value is 0.002A. An empirical constant k is set to 4 (4 is chosen as an empirical constant to balance missed detections and false positives), and the baseline noise threshold is set to 0.008A. Specifically, the mean of the energy sequence is then calculated. When the signal is stable, this mean is small, for example, 0.5. Based on this, the system generates a dynamic adjustment coefficient, such as 1.0. The final threshold is then the baseline threshold of 0.008A. When the signal experiences transient shocks or strong harmonic distortion, the mean of the energy sequence increases significantly, for example, to 2.0, and the dynamic adjustment coefficient is correspondingly increased to 1.5. The final threshold is then adjusted to 0.012A. A higher threshold can effectively prevent high-frequency components in transient processes from being misidentified as harmonics. After using the final threshold to filter out valid spectral lines, for example, a spectral line with an amplitude of 0.5A can be corrected from the discrete spectrum's 249.5Hz to a more accurate 250.1Hz, and its amplitude from 0.49A to 0.51A, using three-point parabolic interpolation using the line itself and three points on its left and right adjacent spectral lines, thereby eliminating errors caused by the picket fence effect.

[0035] S5, performing kernel principal component analysis on the corrected harmonic eigenvector, extracting its largest eigenvalue as the nonlinear principal component energy, and taking the geometric mean of all eigenvalues ​​as the kernel space generalized variance.

[0036] In an optional embodiment, performing kernel principal component analysis on the corrected harmonic eigenvector includes the following steps: The radial basis function (RBF) is selected as the kernel function of kernel principal component analysis; By combining grid search with cross validation, the width parameter of the radial basis function is determined by automatically optimizing within the preset parameter range.

[0037] Specifically, traditional linear principal component analysis (PCA) struggles to process the harmonic characteristics generated by nonlinear processes such as intelligent transformer saturation. Therefore, kernel PCA is employed. This uses a kernel function to map the original harmonic feature vectors (e.g., vectors containing the amplitude and phase of the fundamental, third, fifth, and seventh harmonics) into a high-dimensional feature space. This allows data from different operating states to be linearly separable within this high-dimensional feature space.

[0038] 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 is crucial to the performance of kernel principal component analysis. It will cause the model to overfit, and too large This makes the model underfitting and unable to distinguish samples. In order to find the best sigma value, a combination of grid search and cross validation is used.

[0039] First set The search range is (for example, 0.01 to 1000), and a series of candidate values ​​are selected at power intervals of 10, such as 0.01, 0.1, 1, 10, 100, 1000. For each candidate The performance of the model is evaluated on the training data using 5-fold cross validation. The evaluation index can be the inter-class separation of the features after dimensionality reduction or the accuracy of the subsequent classifier. Finally, the model with the highest average performance score in the cross validation is selected. A value (e.g. 10) is chosen as the final width parameter.

[0040] S6, combines the nonlinear principal component energy, the kernel space generalized variance and their respective first-order difference values ​​over time into a four-dimensional feature vector sequence, and inputs the four-dimensional feature vector sequence into a pre-trained gated recurrent unit network, and judges the current operating status of the intelligent transformer through the network output.

[0041] Specifically, in each analysis time window , the current nonlinear principal component energy is recorded as , the generalized variance in kernel space is recorded as , calculate their difference with the previous time window The first-order difference values ​​of the corresponding values ​​are 、 ,in, The previous time window The corresponding nonlinear principal component energy, The previous time window The corresponding generalized variance of the kernel space. 、 、 、 The four numerical values ​​are combined into a four-dimensional feature vector; as time goes by, a four-dimensional feature vector sequence is formed. This four-dimensional feature vector sequence is input into a gated recurrent unit network that has been trained using historical normal and fault data. The final output layer of the gated recurrent unit network (for example, a fully connected layer plus a Softmax activation function) will output probability values ​​corresponding to preset states such as normal, minor fault, and major fault. The state with the highest probability value is taken as the current operating state of the intelligent transformer.

[0042] 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 using a Softmax activation function, and the number of neurons in the fully connected output layer matches the number of preset intelligent transformer operating state types.

[0043] The operating state of intelligent transformers, particularly the evolution from normal to saturated, exhibits significant time dependence. Harmonic characteristics at a single point in time are insufficient for accurate judgment. Therefore, a recurrent neural network, known as a gated recurrent unit network, is employed to process the time series consisting of feature vectors output by kernel principal component analysis. For example, a feature vector is generated every 20 milliseconds, and a series of 10 consecutive feature vectors is input into the network, enabling it to learn the dynamic patterns of state changes.

[0044] In an optional embodiment, in the gated recurrent unit network, the number of nodes in the input layer is the same as the dimension of the input feature vector (for example, 15 features after dimensionality reduction).

[0045] The input layer is followed by a hidden layer consisting of 64 neurons, each containing a gated recurrent unit (GRU), to capture both short-term and long-term dependencies in the time series. The output of the hidden layer is fed into a fully connected output layer. If the states to be diagnosed are normal operation, magnetic circuit saturation, and transient fault, the output layer would have three neurons. Applying the Softmax activation function to these three neurons converts the raw values ​​in the output layer into a probability distribution. For example, the output [0.1, 0.8, 0.1] represents the probability of the current sequence being judged as normal, saturated, and transient fault, respectively.

[0046] In an optional embodiment, when judging the current operating state of the intelligent transformer through 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 its corresponding index is mapped to the current operating state of the intelligent transformer; wherein the types of the current operating state of the intelligent transformer include at least: normal operating state, magnetic circuit saturation state and transient fault state.

[0047] Specifically, the output layer of the gated recurrent unit network provides a probability estimate for each preset operating state. For example, assume that the three output neurons of the gated recurrent unit network correspond, in order, to normal operation, magnetic circuit saturation, and transient fault conditions. At a certain moment, after processing a sequence of feature vectors, the gated recurrent unit network outputs a probability vector of [0.05, 0.92, 0.03]. This means that the model predicts that the intelligent transformer has a 5% probability of being in a normal state, a 92% probability of being in a magnetic circuit saturation state, and a 3% probability of being in a transient fault state. To provide a clear diagnostic conclusion, the maximum probability principle is adopted. In the above example, 92% is the maximum of the three probabilities, and the neuron index corresponding to this maximum value is 1, that is, the second neuron. According to the pre-set mapping relationship, the second neuron represents the magnetic circuit saturation state. Therefore, the final determination is that the current operating state of the intelligent transformer is magnetic circuit saturation.

[0048] In addition, in the description of this specification, “a plurality of” means at least two, for example, two, three or more, etc., unless otherwise clearly and specifically defined.

Claims

1. A data processing method for an intelligent mutual inductor, characterized in that: The following steps are involved: S1, obtaining the time series sampling data of the current or voltage output by the intelligent transformer; S2, using an unscented Kalman filter algorithm to reduce noise on the time series sampling data to obtain first processed data, in which the noise covariance matrix is ​​constructed online based on wavelet packet energy; S3, adaptively adjusting window function parameters according to signal characteristics of the first processed data, performing windowing and truncating on the first processed data, and obtaining a preliminary spectrum through fast Fourier transform; S4, determining a spectral line screening threshold based on a Teager-Kaiser energy operator calculated from the first processed data, using the spectral line screening threshold to screen effective harmonic and interharmonic spectral lines from the preliminary spectrum, and using a spectral line interpolation correction algorithm to perform frequency and amplitude correction on the screened spectral lines to generate a corrected harmonic feature vector; S5, performing kernel principal component analysis on the corrected harmonic eigenvector, extracting its largest eigenvalue as the nonlinear principal component energy, and taking the geometric mean of all eigenvalues ​​as the kernel space generalized variance.

2. The data processing method for intelligent mutual inductor according to claim 1, characterized in that: In S1, the analog signal output from the secondary side of the intelligent transformer is collected by a data acquisition card at a set sampling frequency, and analog-to-digital conversion is performed to form a discrete digital time-series sampling data sequence.

3. The data processing method for intelligent mutual inductor according to claim 1, characterized in that: In S2, when denoising the time series sampling data, the following steps are included: Perform multi-layer wavelet packet decomposition on the time series sampling data to obtain a set of frequency band sub-signals; Select a preset high-frequency sub-signal, calculate its total energy and use the total energy as the noise variance estimate at the current moment; The exponentially weighted moving average method is used to update the current noise variance estimate based on the current noise variance estimate and the previous noise variance, and the process noise covariance diagonal matrix is ​​constructed based on this. .

4. The data processing method for intelligent mutual inductor according to claim 3, characterized in that: When processing the time series sampling data output from the secondary side of the intelligent transformer, the signal is decomposed by three-layer wavelet packet decomposition, and the original signal is decomposed into eight frequency band sub-signals that are evenly divided from low to high.

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

6. The data processing method for intelligent mutual inductor according to claim 1, characterized in that: In S4, when performing frequency and amplitude correction on the screened spectral lines, the following steps are included: For the first processed data, a preliminary spectrum is first obtained by Fourier transform. The median of the amplitudes of all spectral lines in the preliminary spectrum is calculated and k times the median is set as the baseline 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 according to 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 line screening threshold; The three-point parabola or Gaussian interpolation method is used to interpolate each filtered spectral line and its left and right adjacent spectral lines to correct their frequency and amplitude.

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

8. The data processing method for intelligent mutual inductor according to claim 1, characterized in that: In S5, when performing kernel principal component analysis on the corrected harmonic eigenvector, the following steps are included: The radial basis function is selected as the kernel function of kernel principal component analysis; By combining grid search with cross validation, the width parameter of the radial basis function is determined by automatically optimizing within the preset parameter range.

9. The data processing method for an intelligent mutual inductor according to any one of claims 1 to 8, characterized in that: The gated recurrent unit network includes an input layer, at least one gated recurrent unit hidden layer, and a fully connected output layer using a Softmax activation function. The number of neurons in the fully connected output layer matches the number of preset intelligent transformer operating state types.

10. The intelligent transformer data processing method according to claim 9, characterized in that: When judging the current operating state of the intelligent transformer through 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 its corresponding index is mapped to the current operating state of the intelligent transformer; wherein the types of the current operating state of the intelligent transformer include at least: normal operating state, magnetic circuit saturation state and transient fault state.

Images