Partial discharge signal feature extraction method and system

Abstract

The invention discloses a partial discharge signal feature extraction method and system, and the method comprises the following steps: collecting a current signal on a grounding loop cable of detected electrical equipment, and carrying out the preprocessing of the current signal; converting the preprocessed signal into a two-dimensional matrix, and performing singular value decomposition on the two-dimensional matrix to obtain a principal component; matrix reconstruction is carried out based on the principal component to generate a de-noising matrix, and the de-noising matrix is restored to a time domain signal sequence; performing discrete wavelet transform on the time domain signal sequence to obtain a multi-level approximation coefficient and a detail coefficient; the discrete wavelet transform selects a wavelet basis function from the candidate wavelet basis set based on a preset adaptive wavelet basis selection mechanism, and adaptively determines the number of wavelet decomposition layers based on the signal dominant frequency characteristics; and inputting the approximation coefficient and the detail coefficient into an inverse wavelet transform module to obtain a final partial discharge signal. According to the method, efficient, accurate and universal partial discharge signal feature extraction can be realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power equipment state monitoring and signal processing, and particularly relates to a partial discharge signal feature extraction method and system. BACKGROUND

[0002] Partial discharge (PD) detection, as an important means of monitoring the insulation state of high-voltage electrical equipment, plays a key role in preventing equipment failures and ensuring the safe operation of power systems. In large-scale power equipment monitoring, how to efficiently process and store a large amount of PD data has become a research hotspot. PD signals exhibit high frequency, short duration, and low amplitude characteristics, which result in a large amount of data and susceptibility to various noise interference, including white noise, impulse interference, and periodic narrowband interference. Among them, white noise, which is widely distributed in various frequency bands and difficult to effectively suppress, is the main challenge in PD signal detection. These noises not only reduce the clarity of the signal, but also increase the complexity and computational burden of data processing. Therefore, efficient data compression and denoising processing of PD signals can not only reduce the storage and transmission requirements, but also improve the real-time performance and accuracy of signal processing.

[0003] Chinese patent application with publication number CN116383609A discloses a partial discharge signal denoising method combining singular value decomposition and wavelet transform. The method includes: selecting a suitable Hankel matrix to construct an adaptive singular value decomposition model, which is responsible for singular value decomposition of the original noisy partial discharge signal; calculating the singular entropy increment, and finding the position with the maximum bending degree of the singular entropy curve according to its asymptotic property; using the position with the maximum curvature as a threshold, and using the singular values greater than the threshold to reconstruct the original signal; selecting a suitable wavelet basis function to construct a one-dimensional two-level dual-tree complex wavelet transform, which is responsible for signal decomposition and reconstruction; and using a q-shift scheme to jointly construct the filter bank inside the DT-CWT.

[0004] The above technical solution uses the maximum point of singular entropy curvature as the basis for judging the threshold. This approach has the problems of strong rigidity and poor robustness. In the case of relatively flat or multi-section energy distribution of the signal, a single threshold may not accurately distinguish between signal and noise, which can easily cause problems such as misremoval of effective signal components, insufficient suppression of residual noise, and poor adaptability of the denoising result to different signal samples. Moreover, the wavelet transform method is complex but lacks sufficient adaptability. The fixed two-level DT-CWT decomposition does not consider the variation range of the main frequency of the partial discharge signal, which can result in a mismatch between the decomposition scale and the signal spectrum. The use of a fixed wavelet basis function limits the method's ability to generalize to different scenarios. SUMMARY

[0005] The application provides a partial discharge signal feature extraction method and system, aiming at solving the problems of poor robustness and lack of adaptability in the prior art.

[0006] To solve the above technical problems, the partial discharge signal feature extraction method provided by the application comprises the following steps: Collecting a current signal on a grounding loop cable of an electrical equipment to be measured, and preprocessing the current signal; Converting the preprocessed signal into a two-dimensional matrix, and performing singular value decomposition on the two-dimensional matrix to obtain principal components; Based on the principal components, a denoising matrix is generated by matrix reconstruction, and the denoising matrix is restored to a time domain signal sequence; Performing discrete wavelet transform on the time domain signal sequence to obtain multi-level approximation coefficients and detail coefficients; the discrete wavelet transform selects a wavelet base function from a candidate wavelet base set based on a preset adaptive wavelet base selection mechanism, and adaptively determines a wavelet decomposition level based on a signal main frequency feature; The approximation coefficients and the detail coefficients are input into an inverse wavelet transform module to obtain a final partial discharge signal.

[0007] Preferably, the method for obtaining the principal components is as follows: first, the cumulative energy proportion of all singular values is calculated, when the cumulative energy reaches a preset energy threshold, the number of singular values contained in the current cumulative energy is determined as the number of selected principal components, which is used for subsequent matrix reconstruction.

[0008] Preferably, the method for obtaining the principal components is as follows: the information entropy of all singular values after normalization is calculated, and the information entropy trend under different numbers of singular values is analyzed, when the information entropy trend tends to be stable or the change rate is less than a preset threshold, the current number of singular values is determined as the number of selected principal components.

[0009] Preferably, the method for obtaining the principal components is as follows: all singular values are arranged in descending order, and the difference sequence between adjacent singular values is calculated, when the maximum gap appears, the number of singular values before the maximum gap is selected as the number of selected principal components.

[0010] Preferably, the method for selecting the wavelet base function is as follows: each wavelet base function in a preset candidate wavelet base function set is used to perform wavelet decomposition and reconstruction on the time domain signal sequence, the reconstruction error and signal entropy index corresponding to each wavelet base function are calculated, and the optimal wavelet base function is selected for signal decomposition and reconstruction according to the principle of minimum reconstruction error and minimum signal entropy.

[0011] Preferably, the method for determining the wavelet decomposition level is as follows: the frequency spectrum of the time domain signal sequence is analyzed to determine the signal main frequency, the appropriate decomposition level is calculated according to the relationship between the signal main frequency and the sampling frequency and the frequency band division characteristics of wavelet decomposition, and the calculation formula is:

[0012] wherein, is the number of decomposition layers calculated, is the sampling frequency, is the signal fundamental frequency.

[0013] Preferably, the detail coefficients are further processed by a soft threshold function before inputting into the inverse wavelet transform module, the soft threshold function estimates the threshold value of each layer of detail coefficients based on the median absolute deviation, and then realizes the processing of zeroing the coefficients less than the threshold value and reducing the coefficients greater than the threshold value; the threshold value is adaptively calculated, and a scaling factor can be introduced in different decomposition layers for weighted adjustment to enhance the denoising sensitivity under multi-scale; the detail coefficients after threshold processing and the last layer of approximation coefficients are jointly input into the inverse wavelet transform module.

[0014] Preferably, the current signal is collected by a high-frequency current transformer, and the original signal collected is digitized by a data acquisition card to form an original current signal.

[0015] Preferably, the preprocessing includes: removing the direct current component, and the removal method includes subtracting the global mean value of the signal or using a high-pass filter to filter out the low-frequency components; normalizing the signal amplitude to normalize the signal to the interval [-1, 1]; and / or amplitude limiting filtering or median filtering is performed on the signal.

[0016] Another aspect of the present application also provides a partial discharge signal feature extraction system, which is used to implement the above-mentioned feature extraction method, and includes: a signal acquisition module, configured to acquire a current signal on a grounding loop cable of an electrical equipment under test, and to preprocess the current signal; a signal matrix conversion module, configured to convert the preprocessed current signal into a two-dimensional matrix; a principal component extraction module, configured to perform singular value decomposition on the two-dimensional matrix and extract principal components; a denoising reconstruction module, configured to reconstruct the matrix based on the principal components, generate a denoised matrix, and restore the denoised matrix to a time-domain signal sequence; a wavelet decomposition module, configured to perform discrete wavelet transform on the time-domain signal sequence to obtain multi-level approximation coefficients and detail coefficients, and the wavelet decomposition module has the following capabilities: adaptively determining the number of decomposition layers based on the signal fundamental frequency characteristics; selecting a wavelet base function from a set of candidate wavelet base functions based on a preset wavelet base selection mechanism; The wavelet reconstruction module is used to input the approximation coefficients and detail coefficients into the inverse wavelet transform module to generate the final partial discharge signal.

[0017] Compared with the prior art, the present invention has the following technical effects: 1. The partial discharge signal feature extraction method proposed in this invention adopts a composite signal processing algorithm that integrates Adaptive Singular Value Decomposition (SVD) and Discrete Wavelet Transform (DWT). By fusing the low-rank features extracted by SVD with the multi-scale analysis capability of DWT, it effectively realizes the joint modeling of structural and frequency information, and significantly improves the recognition rate and integrity retention rate of non-Gaussian pulses (such as partial discharge pulses).

[0018] 2. The partial discharge signal feature extraction method proposed in this invention introduces an adaptive objective function composed of kurtosis, MSE, and SNR to guide the denoising process. Maximizing kurtosis promotes the enhancement of abrupt pulse features, minimizing MSE suppresses structural distortion, and maximizing SNR improves the overall signal quality. This objective function is used to optimize the parameter adjustment of each step in the denoising process, including the singular value threshold, the number of wavelet decomposition layers, and the threshold function form. The three-index collaborative constraint mechanism can preserve abrupt information to the greatest extent without sacrificing the signal structure, and has good generalization ability, significantly improving the algorithm's adaptability to different electrical equipment and different operating conditions.

[0019] 3. The partial discharge signal feature extraction method proposed in this invention introduces a variety of singular value selection strategies, such as the energy ratio method, entropy change method, and maximum gap method, which can automatically determine the optimal number of principal components for different signal types, effectively distinguish between signals and noise, avoid misjudgments caused by subjective threshold settings, and thus improve the stability and accuracy of the singular value decomposition denoising stage.

[0020] 4. In the wavelet transform detail coefficient processing stage, the partial discharge signal feature extraction method proposed in this invention adopts a soft threshold function based on median absolute deviation (MAD) estimation, combined with a hierarchical weighting mechanism to adjust the threshold size, thereby effectively suppressing noise in different frequency bands, preserving key partial discharge feature information, avoiding signal distortion or abrupt changes caused by traditional hard thresholding methods, and improving the smoothness and fidelity of the reconstructed signal. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the partial discharge signal feature extraction method of the present invention; Figure 2 This is a step-by-step schematic diagram of the feature extraction method described in the embodiments of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present application and with reference to the accompanying drawings.

[0023] Example 1 This embodiment describes a method for extracting features from partial discharge signals, such as... Figure 1 As shown, it includes the following steps one through five: Step 1: Collect the current signal on the grounding circuit cable of the electrical equipment under test, and preprocess the current signal.

[0024] In this embodiment, the current signal is acquired using a high-frequency current transformer, and the acquired raw signal is digitized by a data acquisition card to form the raw current signal.

[0025] Specifically, the partial discharge signal is acquired using a high-frequency current transformer (HFCT) as the front-end sensing device. HFCT is a non-invasive, high-bandwidth, low-noise current sensing device widely used in the field of partial discharge detection in high-voltage equipment. It possesses excellent high-frequency response characteristics and can effectively capture high-frequency current pulse signals generated during the discharge process.

[0026] In practical implementation, the HFCT is installed on the grounding loop cable of the electrical equipment under test (such as the transformer casing or the grounding end of the cable shield) to sense the high-frequency pulse current caused by partial discharge in a non-contact manner. The bandwidth of this device typically covers tens of kilohertz to tens of megahertz, which can effectively avoid power frequency signal interference and ensure the integrity of the partial discharge pulse signal.

[0027] The acquired raw signal is digitized using a high-bandwidth, high-sampling-rate data acquisition card (an oscilloscope with a sampling rate of 65 MS / s) to form the raw current signal, which is a one-dimensional time-domain signal sequence. The raw current signal is usually mixed with power frequency interference (such as the 50Hz fundamental wave and its harmonic components), random impulse noise, environmental electromagnetic interference, and other interference components. In addition, the partial discharge signal itself has low amplitude and short duration, exhibiting non-stationary and sparse abrupt change characteristics.

[0028] To ensure the stability and effectiveness of the subsequent denoising process, the raw current signal acquired by the HFCT needs to be standardized and preprocessed.

[0029] The preprocessing includes: To remove the DC component and eliminate baseline offset in the measurement circuit, one method could be to subtract the global mean of the signal or use a high-pass filter to remove low-frequency components. The signal amplitude is normalized to the range of [-1,1] to unify the dimensional differences between different sampling batches or devices; In addition, in some other embodiments of the present invention, the signal can be subjected to amplitude limiting filtering or median filtering to suppress extreme abnormal pulses generated by lightning strikes, switching operations, etc.

[0030] After the above preprocessing steps, the HFCT acquired signal is converted into a standardized one-dimensional partial discharge signal sequence, providing a good input foundation for subsequent matrix construction and denoising processing. This signal retains the high-frequency characteristics of the original partial discharge pulse while significantly reducing the interference of power frequency interference and background noise, thus providing a high-quality and structurally clear initial signal input for subsequent denoising algorithms.

[0031] Step 2: Convert the preprocessed signal into a two-dimensional matrix, and perform singular value decomposition on the two-dimensional matrix to obtain principal components.

[0032] To facilitate subsequent singular value decomposition (SVD) processing, the preprocessed signal is first converted into a two-dimensional data matrix. The main purpose of this step is to transform the time-domain signal into a form suitable for matrix decomposition using the sliding window method, thus facilitating SVD.

[0033] In this embodiment, a sliding window method is used to convert the preprocessed signal into a two-dimensional matrix: First, a fixed-length window L is set, and the sliding step of the window is s. This method divides the original signal sequence into multiple windows, with the signal within each window forming a subsequence. Assuming the original signal length is N, the sliding window function will form K = (NL) / s + 1 windows. Each signal segment is then stored column-wise in a two-dimensional matrix. In, that is:

[0034] in The constructed two-dimensional matrix has each column representing a signal window, containing local features of the time series. In practical applications, the window length is typically set to tens to hundreds, and the specific value needs to be adjusted according to the time-frequency characteristics of the actual signal.

[0035] Once the two-dimensional matrix has been constructed Next, adaptive singular value decomposition (ASVD) is performed. The core idea of ​​ASVD is to decompose the signal matrix into multiple independent components, and to achieve noise reduction by selecting important components and removing noise components.

[0036] After signal preprocessing, the next step is to prepare the input data by constructing a two-dimensional matrix and then perform noise reduction on the original signal using the Singular Value Decomposition (SVD) method. This step effectively extracts the principal components of the partial discharge signal while removing noise components, providing a high-quality signal input for subsequent noise reduction operations.

[0037] Singular Value Decomposition (SVD) is a classic matrix factorization method widely used in signal processing, noise reduction, and feature extraction. SVD decomposes a matrix (in this example, the two-dimensional matrix described above) into its constituent parts. The product of three matrices is expressed as follows:

[0038] in, It is a two-dimensional matrix. It is a left singular vector matrix. It is a right singular vector matrix. It is a diagonal matrix containing singular values. That is, its diagonal elements are singular values, arranged in descending order. The larger the singular value, the higher the energy of the component in the signal, usually representing the principal component of the signal; while smaller singular values ​​often correspond to noise components.

[0039] In this embodiment, the method for obtaining the principal components is as follows: first, the cumulative energy percentage of all singular values ​​is calculated; when the cumulative energy reaches a preset energy threshold, the number of singular values ​​included in the current accumulation is determined as the number of selected principal components for subsequent matrix reconstruction.

[0040] Specifically, when obtaining principal components, the energy percentage of each singular value is first calculated, and an energy threshold η is set. The first r singular values ​​are retained, where r satisfies:

[0041] In another embodiment of the present invention, the method for obtaining the principal components is as follows: calculate the information entropy after normalization of all singular values, and analyze the trend of information entropy change under different numbers of singular values. When the information entropy change tends to stabilize or the rate of change is less than a preset threshold, determine the current number of singular values ​​as the number of selected principal components. That is, the location of noise components is determined by monitoring the inflection point of entropy change. The location with a large entropy change usually represents the transition region of the signal, while the part with a small entropy change corresponds to noise components.

[0042] In another embodiment of the present invention, the method for obtaining the principal components is as follows: all singular values ​​are arranged in descending order, the difference sequence between adjacent singular values ​​is calculated, and when the largest difference gap occurs, the number of singular values ​​before the largest gap is selected as the number of principal components.

[0043] Step 3: Based on the principal components, perform matrix reconstruction to generate a denoising matrix, and restore the denoising matrix to a time-domain signal sequence.

[0044] The denoised matrix is ​​reconstructed using the selected principal components. , means as follows:

[0045] in, This is the filtered diagonal singular value matrix, containing the first r retained singular values. It is obtained through matrix reconstruction. It contains the main components of the signal, and the noise has been effectively removed.

[0046] Then, the matrix is ​​averaged using the sliding window method. The signal is restored to a one-dimensional time-domain signal sequence, resulting in the denoised signal x(t), which prepares for the subsequent Discrete Wavelet Transform (DWT).

[0047] Step four: Perform discrete wavelet transform on the time-domain signal sequence to obtain multi-level approximation coefficients and detail coefficients; the discrete wavelet transform selects wavelet basis functions from the candidate wavelet basis set based on a preset adaptive wavelet basis selection mechanism, and adaptively determines the number of wavelet decomposition levels based on the signal dominant frequency characteristics.

[0048] While partial discharge signals denoised by SVD initially suppress most background interference and system noise, they may still retain non-ideal components such as high-frequency sharp interference and power frequency harmonic fluctuations. This is especially true when the signal amplitude is small or the partial discharge amplitude is close to the noise lower limit, making effective separation difficult with traditional filters. To further improve the denoising accuracy and feature preservation capability, this technical solution introduces a wavelet threshold denoising mechanism based on discrete wavelet transform. Furthermore, it proposes an adaptive parameter selection strategy for wavelet basis functions and decomposition levels to achieve optimal denoising performance with signal structure awareness.

[0049] DWT is a multi-resolution analysis tool with good localization capabilities in both the time and frequency domains. Its basic idea is to decompose the original signal layer by layer using a set of orthogonal or bioorthogonal wavelet filters (high-pass and low-pass filter pairs), thereby extracting the low-frequency (approximate) and high-frequency (detail) components of the signal at different scales. By analyzing the energy distribution of the signal at different scales, DWT effectively captures transient anomalous signals (such as partial discharge pulses) while distinguishing between noise and useful components.

[0050] For the input signal x(t), the wavelet decomposition process can be described as follows:

[0051] in, The approximation coefficients of the J-th layer represent low-frequency information. This represents the detail coefficients of the J-th layer, which contain high-frequency information. By preserving approximation coefficients and performing threshold denoising on the detail coefficients, a synergistic effect of signal denoising and abrupt feature enhancement can be achieved.

[0052] The selection of wavelet basis functions has a decisive impact on the DWT decomposition and reconstruction results. Different wavelet basis functions differ in symmetry, orthogonality, tight support, and vanishing moments, which determine their signal matching degree and reconstruction accuracy. Commonly used wavelet basis functions include: Daubechies wavelets (dbN): such as db4 and db6, have good orthogonality and tight support, and are suitable for capturing sharp changes; Symlets wavelet (symN): Approximately symmetrical, suitable for reconstructing signals with high accuracy requirements; Coiflets wavelet (coifN): possesses the orthogonality of both function values ​​and derivatives, making it suitable for complex signal analysis; Biorthogonal wavelet (biorN.M): Supports perfect reconstruction and is suitable for signal compression and recovery.

[0053] In this embodiment, in order to achieve optimal fit and dynamic selectivity of wavelet basis functions, an adaptive wavelet basis function selection method and a wavelet decomposition layer number determination method are designed.

[0054] The method for selecting wavelet basis functions is as follows: from a preset set of candidate wavelet basis functions, each wavelet basis function is used to perform wavelet decomposition and reconstruction on the time-domain signal sequence, and the corresponding reconstruction error and signal entropy index are calculated. Based on the principle of minimizing reconstruction error and minimizing signal entropy, the optimal wavelet basis function is selected for signal decomposition and reconstruction.

[0055] The specific steps are as follows: Define the set of candidate wavelet bases

[0056] For each DWT decomposition and denoising reconstruction were performed separately, and the following performance metrics were evaluated: mean squared error (MSE), kurtosis, and signal-to-noise ratio (SNR).

[0057] Mean square error is used to measure the mean squared error between the denoised signal and the original signal; the smaller the value, the better the reconstruction effect.

[0058] In the formula, Indicates the original signal at the 1st... The amplitude of each sampling point This represents the signal value after denoising (or reconstruction); N represents the total number of sampling points.

[0059] Kurtosis is a statistical measure describing the non-Gaussianity of a signal. For Gaussian distributed signals, the theoretical value is 3. The kurtosis of pulse signals such as partial discharge is usually significantly greater than 3. Therefore, the larger the kurtosis, the more obvious the instantaneous spike or pulse characteristics are in the signal.

[0060] in, , where represents the mean of the signal, the denominator is the square of the variance, and N represents the total number of sampling points.

[0061] Signal-to-noise ratio (SNR) represents the ratio of signal energy to noise energy, and its unit is decibel (dB). The higher the value, the higher the noise reduction quality.

[0062] The numerator is the energy of the original signal (signal power), and the denominator is the energy of the signal reconstruction error (noise power).

[0063] Optimal wavelet basis function selection: ,in , , This is a weighting factor that can be dynamically set according to the application scenario. The wavelet basis function that ultimately yields the best overall performance is selected. For use in formal DWT processing.

[0064] The method for determining the number of wavelet decomposition levels is as follows: Perform spectral analysis on the time-domain signal sequence to determine the signal's dominant frequency. Based on the relationship between the signal's dominant frequency and the sampling frequency, and combined with the wavelet decomposition frequency band division characteristics, calculate the appropriate number of decomposition levels. The calculation formula is:

[0065] In the formula, The number of decomposition layers is calculated. Sampling frequency, This is the signal's main frequency. If... Between 100kHz and 10MHz, the typical number of decomposition layers is 3 to 6.

[0066] In other embodiments of the present invention, if the energy distribution of the main frequency band is wide, the entropy method or energy ratio method is used to compensate and correct the number of decomposition layers.

[0067] This mechanism ensures that the scale coverage of wavelet decomposition closely matches the dominant frequency distribution of the signal, avoiding information loss or over-decomposition.

[0068] Step 5: Input the approximation coefficients and detail coefficients into the inverse wavelet transform module to obtain the final partial discharge signal.

[0069] In other embodiments of the present invention, the detail coefficients are further subjected to a soft thresholding function to suppress noise before being input into the inverse wavelet transform module. The soft thresholding function estimates the corresponding threshold based on the median absolute deviation of the detail coefficients at each level, thereby setting coefficients smaller than the threshold to zero and reducing coefficients larger than the threshold. The threshold is adaptively calculated, and a scaling factor can be introduced in different decomposition levels for weighted adjustment to enhance the denoising sensitivity at multiple scales. The detail coefficients after thresholding are input into the inverse wavelet transform module together with the last level of approximation coefficients.

[0070] The soft threshold function is as follows:

[0071] In the formula, These are the detail coefficients after soft thresholding. The detail coefficients before processing. The adaptive threshold for the detail coefficients of the j-th layer is obtained using the following method:

[0072]

[0073] In the formula, M represents the number of coefficients in the current layer.

[0074] For different numbers of layers, a scaling factor can be introduced to adjust the threshold to improve sensitivity at multiple scales (multi-layer weighted strategy). After processing, all thresholded detail coefficients are then processed. Approximation coefficients The signal is then fed into an inverse discrete wavelet transform (IDWT) to obtain a smoothed and denoised reconstructed signal. .

[0075] To evaluate the effectiveness of the method described in this embodiment for partial discharge signal denoising, the signal-to-noise ratio (SNR), normalized cross-correlation coefficient (NCC), and root mean square error (RMSE) are used to evaluate the effectiveness of the method.

[0076] Signal-to-noise ratio (SNR): SNR is an important indicator for evaluating signal quality. It is defined as the ratio of signal power to noise power, usually expressed in decibels (dB). The higher the SNR, the better the signal quality and the less noise interference.

[0077] Normalized cross-correlation coefficient (NCC): NCC is used to quantify the similarity between two signals, and its calculation range is between [-1, 1]. The closer the value is to 1, the higher the similarity between the signals, indicating a good denoising effect.

[0078] Root Mean Square Error (RMSE): RMSE is an important metric for measuring the difference between the processed signal and the original signal. It is calculated as the square root of the average of the sum of squares of the errors between the predicted and true values. The formula is: The Noise Rejection Ratio (NRR) reflects how prominent the signal is after noise reduction. A higher NRR value indicates a better noise reduction effect.

[0079] Experiments A and B were conducted to compare and analyze the performance of four denoising methods in discharge signal processing: ASVD-DWT (in this embodiment), ASVD-IWT (Adaptive Singular Value Decomposition-Inverse Wavelet Transform), EMD-DWT (Empirical Mode Decomposition-Discrete Wavelet Transform), and DWT (Discrete Wavelet Transform), in order to verify the effectiveness of the ASVD-DWT method.

[0080] Experiment A: Comparison of noise reduction for simulated PD signals Experiment A used simulated PD signals. The simulation was run using PyCharm software on an Intel(R) Core(TM) i7-8750H processor with a clock frequency of 2.20GHz and 16GB of memory. Different denoising methods were applied to this data set, and the results are shown in the figure below. Table 1 presents the evaluation results of different data based on SNR, NCC, and RMSE.

[0081]

[0082] Table 1 Comparison of noise reduction effects of simulated signals As can be seen from the data in the table, the ASVD-DWT method proposed in this embodiment demonstrates significantly better denoising performance than other methods across all evaluation metrics. Specifically, ASVD-DWT excels in all denoising metrics, significantly outperforming other methods. In contrast, other methods such as ASVD-IWT, EMD-DWT, and DWT exhibit poor denoising performance, showing low values ​​in NCC, RMSE, and SNR metrics, with ASVD-IWT achieving an SNR of only 2.20 dB, while EMD-DWT and DWT also fall short of the performance of ASVD-DWT. Therefore, ASVD-DWT demonstrates superior denoising performance across all metrics, proving its effectiveness in noise suppression.

[0083] Experiment B: Comparison of PD signal denoising measurements Experiment B selected 5000 points before and after the position of the highest amplitude within a single cycle of the PD signal measured in the laboratory as experimental data for denoising. The Non-Rating Detection Ratio (NRR) was used as the evaluation metric for denoising. The denoising results for each method on the real PD signal are shown in Table 2 below.

[0084]

[0085] Table 2 Comparison of Signal Denoising Effects in the Laboratory Based on the analysis in Table 2 above, the ASVD-DWT algorithm exhibits the best denoising performance across all four sets of different PD signals (A, B, C, and D). In each signal, the NRR of ASVD-DWT is significantly higher than the other three algorithms. Although EMD-DWT also performs well, with a consistently high NRR value, it still falls short of ASVD-DWT. In contrast, the denoising performance of DWT and ASVD-IWT is mediocre, with ASVD-IWT showing the worst performance, consistently remaining at a low level. Overall, the data analysis indicates that ASVD-DWT demonstrates strong noise suppression capabilities across all four sets of real PD signals.

[0086] This embodiment proposes a partial discharge signal processing method based on the combination of Adaptive Singular Value Decomposition (ASVD) and Discrete Wavelet Transform (DWT). Through processing simulation and measured data, the results show that ASVD can effectively extract the main features from the signal and significantly reduce the impact of noise. Especially when processing partial discharge signals in high-noise environments, the ASVD-DWT method exhibits a higher signal-to-noise ratio (SNR) and a lower error rate compared to traditional denoising techniques. Experimental results further verify the superiority of the ASVD-DWT method in partial discharge signals through comparison with other denoising methods. The ASVD method adaptively adjusts the singular values ​​of the signal, resulting in more accurate denoising, while the multi-scale decomposition characteristics of DWT effectively suppress high-frequency noise. This combination not only improves the clarity and accuracy of the signal but also provides more reliable support for feature extraction of partial discharge signals. The denoising results show that ASVD-DWT has a better denoising effect compared to ASVD-IWT, EMD-DWT, and wavelet transform denoising methods.

[0087] Example 2 A partial discharge signal feature extraction system, the system being used to implement the feature extraction method as described in Embodiment 1, comprising: The signal acquisition module is used to acquire the current signal on the grounding circuit cable of the electrical equipment under test, and to preprocess the current signal. The signal matrix conversion module is used to convert the preprocessed current signal into a two-dimensional matrix; The principal component extraction module is used to perform singular value decomposition on the two-dimensional matrix and extract principal components; The denoising and reconstruction module is used to reconstruct the matrix based on the principal components, generate a denoised matrix, and restore the denoised matrix to a time-domain signal sequence. The wavelet decomposition module is used to perform discrete wavelet transform on the time-domain signal sequence to obtain multi-level approximation coefficients and detail coefficients. The wavelet decomposition module has the following capabilities: the ability to adaptively determine the number of decomposition levels based on the signal's dominant frequency characteristics; and the ability to select wavelet basis functions from a set of candidate wavelet basis functions based on a preset wavelet basis selection mechanism. The wavelet reconstruction module is used to input the approximation coefficients and detail coefficients into the inverse wavelet transform module to generate the final partial discharge signal.

[0088] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A partial discharge signal feature extraction method, characterized by, The method comprises the following steps: Collecting a current signal on a grounding loop cable of an electrical equipment to be tested, and preprocessing the current signal; Converting the preprocessed signal into a two-dimensional matrix, and performing singular value decomposition on the two-dimensional matrix to obtain principal components; Based on the principal components, a denoising matrix is generated by matrix reconstruction, and the denoising matrix is restored to a time-domain signal sequence; Performing discrete wavelet transform on the time-domain signal sequence to obtain multi-level approximation coefficients and detail coefficients; the discrete wavelet transform selects a wavelet base function from a set of candidate wavelet base functions based on a preset adaptive wavelet base selection mechanism, and adaptively determines a wavelet decomposition level based on a signal main frequency feature; The approximation coefficients and the detail coefficients are input into an inverse wavelet transform module to obtain a final partial discharge signal.

2. The method of claim 1, wherein, The method for obtaining the principal components is as follows: first, the cumulative energy proportion of all singular values is calculated; when the cumulative energy reaches a preset energy threshold, the number of singular values contained in the current cumulative energy is determined as the number of selected principal components, which is used for subsequent matrix reconstruction.

3. The method of claim 1, wherein, The method for obtaining the principal components is as follows: the information entropy of all singular values after normalization is calculated, and the information entropy trend under different numbers of singular values is analyzed; when the information entropy trend tends to be stable or the change rate is less than a preset threshold, the current number of singular values is determined as the number of selected principal components.

4. The method of claim 1, wherein, The method for obtaining the principal components is as follows: all singular values are arranged in descending order, and the difference sequence between adjacent singular values is calculated; when the maximum difference gap occurs, the number of singular values before the maximum gap is selected as the number of selected principal components.

5. The method of claim 1, wherein, The method for selecting the wavelet base function is as follows: each wavelet base function in a set of preset candidate wavelet base functions is used to perform wavelet decomposition and reconstruction on the time-domain signal sequence, the reconstruction error and signal entropy index corresponding to each wavelet base function are calculated, and the optimal wavelet base function is selected for signal decomposition and reconstruction according to the principle of minimum reconstruction error and minimum signal entropy.

6. The method of claim 1, wherein, The method for determining the wavelet decomposition level is as follows: the frequency spectrum of the time-domain signal sequence is analyzed to determine the signal main frequency; the appropriate decomposition level is calculated according to the relationship between the signal main frequency and the sampling frequency and the frequency band division characteristics of wavelet decomposition; and the calculation formula is as follows: In the formula, is the number of decomposition layers calculated, is the sampling frequency, is the signal main frequency.

7. The method of claim 1, wherein, Before being input into the inverse wavelet transform module, the detail coefficients are further processed by a soft threshold function to suppress noise; the soft threshold function estimates the threshold value of each layer of detail coefficients based on the median absolute deviation of the detail coefficients, and then realizes the processing of zeroing the coefficients less than the threshold value and reducing the coefficients greater than the threshold value; the threshold value is adaptively calculated, and a proportional factor can be introduced for weighted adjustment in different decomposition layers to enhance the denoising sensitivity in multiple scales; the detail coefficients after threshold processing and the last layer of approximation coefficients are input into the inverse wavelet transform module.

8. The method of claim 1, wherein, The current signal is collected by a high-frequency current transformer, and the original signal collected is digitized by a data acquisition card to form an original current signal.

9. The method of claim 1, wherein, The preprocessing includes: Removing the direct current component, and the removal method includes subtracting the global mean value of the signal or filtering out the low-frequency component by using a high-pass filter; Normalizing the signal amplitude to standardize the signal to the interval [-1, 1]; and / or Limiting the amplitude of the signal or performing median filtering.

10. A partial discharge signal feature extraction system characterized by, The system is used for implementing the feature extraction method as claimed in any one of claims 1-9, comprising: a signal acquisition module, configured to acquire a current signal on a grounding loop cable of an electrical equipment under test, and to pre-process the current signal; a signal matrix conversion module, configured to convert the pre-processed current signal into a two-dimensional matrix; a principal component extraction module, configured to perform singular value decomposition on the two-dimensional matrix, and to extract principal components; a denoising reconstruction module, configured to reconstruct the matrix based on the principal components, to generate a denoised matrix, and to restore the denoised matrix into a time-domain signal sequence; a wavelet decomposition module, configured to perform discrete wavelet transform on the time-domain signal sequence, to obtain multi-level approximation coefficients and detail coefficients, the wavelet decomposition module having the ability to adaptively determine a decomposition level based on a signal main frequency feature, and the ability to select a wavelet base function from a candidate wavelet base set based on a preset wavelet base selection mechanism; a wavelet reconstruction module, configured to input the approximation coefficients and the detail coefficients into an inverse wavelet transform module, to generate a final partial discharge signal.

Images