OLTC vibration signal noise reduction method based on IREMD-MPE-LKSVD

Abstract

The invention discloses an OLTC vibration signal noise reduction method based on IREMD-MPE-LKSVD, and belongs to the technical field of signal processing. The method comprises the following steps: adaptively decomposing an original converter transformer OLTC vibration signal into IMF components and a residual signal component by using an IREMD method, calculating an MPE value, judging the components with the MPE value greater than a first preset threshold as noisy signal components, and taking the rest as effective signal components; processing the noisy signal component by adopting an improved self-adaptive wavelet threshold noise reduction method, and reconstructing the noisy signal component processed by the improved self-adaptive wavelet threshold noise reduction method and the effective signal component to obtain a primary noise reduction signal; and carrying out secondary noise reduction on the primary noise reduction signal by adopting an LKSVD method to obtain a noise-reduced OLTC vibration signal. According to the invention, effective suppression of OLTC vibration signal noise is realized, and vibration signal quality is improved.

Description

Technical Field

[0001] This invention relates to a noise reduction method for OLTC vibration signals based on IREMD-MPE-LKSVD, belonging to the field of signal processing technology. Background Technology

[0002] Converter transformers, as a crucial bridge connecting AC and DC power grids, are key devices for power conversion. Their core component, the on-load tap changer (OLTC), plays a multifaceted role in power system operation, such as precise reactive power regulation, effective load voltage stabilization, and optimizing the economy and flexibility of the DC system. However, the high-frequency noise signal generated by the electric arc during OLTC switching can affect the effective feature extraction of vibration signals, reducing the accuracy of subsequent OLTC fault diagnosis. Therefore, research on noise reduction of OLTC vibration signals from converter transformers is of great significance.

[0003] OLTC vibration signals are typical non-stationary signals. Current denoising strategies for non-stationary signals typically involve multiple techniques; for example, EMD (Electronic Mode Decomposition) is widely used for fault diagnosis of non-stationary vibration signals. However, the IMF (Integrated Mode Filtering) generated by EMD often suffers from mode aliasing, affecting signal decomposition results. REMD (Reactive Mode Decomposition) is a signal processing technique for time series analysis, essentially an extension of EMD, designed to improve robustness in noisy environments. It addresses the issue of mode aliasing in different datasets that may arise from using a preset threshold method for selecting stopping criteria during EMD decomposition, effectively reducing the impact of noise on the decomposition results and thus improving signal decomposition accuracy. However, traditional REMD methods still have shortcomings in terms of mode aliasing and robustness to noise, and the denoising effect needs further improvement.

[0004] In view of this, the present invention is hereby proposed. Summary of the Invention

[0005] This invention provides a noise reduction method for OLTC vibration signals based on IREMD-MPE-LKSVD, which can be used for deep noise reduction of the original OLTC vibration signals of converter transformers.

[0006] The technical solution of this invention is:

[0007] A noise reduction method for OLTC vibration signals based on IREMD-MPE-LKSVD includes:

[0008] Obtain the original OLTC vibration signal of the converter transformer;

[0009] The original OLTC vibration signal of the converter transformer was adaptively decomposed into the following values ​​using the IREMD method: One IMF component and one residual signal component;

[0010] calculate The MPE values ​​of one IMF component and one residual signal component are used to determine the component with an MPE value greater than a first preset threshold as a noisy signal component, and the rest as valid signal components.

[0011] The noisy signal component is processed by an improved adaptive wavelet threshold denoising method, and the noisy signal component processed by the improved adaptive wavelet threshold denoising method is reconstructed with the effective signal component to obtain a first-stage denoised signal.

[0012] The LKSVD method is used to perform secondary denoising on the primary denoised signal to obtain the denoised OLTC vibration signal.

[0013] Furthermore, the IREMD method is based on the REMD method and is constructed by introducing an adaptive convergence criterion to optimize the screening stopping criterion and adding an orthogonality check.

[0014] Furthermore, the adaptive convergence criterion is as follows: calculate the standardized squared difference of the sum of two consecutive screened components, and stop the screening process of IMF components when the standardized squared difference is less than the adaptive convergence threshold.

[0015] Furthermore, the orthogonality check specifically involves:

[0016] if Or when the modal component list If empty, then accept the current IMF component candidate. As the first One IMF component is added to the modal component list. In; among them, Indicates the current IMF component candidate and Orthogonality index between The maximum value in, List of modal components All extracted IMF components;

[0017] Otherwise, abandon the current IMF weighting candidate. .

[0018] Furthermore, the improved adaptive wavelet thresholding method employs an improved adaptive wavelet threshold function constructed based on the adaptive wavelet threshold.

[0019] Furthermore, the improved adaptive wavelet threshold function is defined as follows:

[0020] ;

[0021] In the formula, After wavelet decomposition, it is located at the th The first layer One original wavelet coefficient, These are the wavelet coefficients after processing with the improved adaptive wavelet threshold function; Represents the step function; It is an adaptive wavelet threshold.

[0022] Furthermore, the adaptive wavelet threshold is defined as follows:

[0023] ;

[0024] In the formula, It is the standard deviation of the noisy signal components. Where c is the signal length and c is the number of wavelet coefficient layers.

[0025] Furthermore, the LKSVD method is based on the K-SVD method and uses a power-law iteration method to solve for the error submatrix. Obtain the maximum singular value and its corresponding left singular vector. and right singular vectors This is how the error submatrix is ​​obtained. The best rank-1 approximation; using left singular vectors Update the learning dictionary matrix The u-th atom in And perform L2 norm normalization; utilize the maximum singular value With right singular vectors The product of transposes with respect to the sparse coefficient matrix The sparse coefficient vector in the u-th row set of position indices of non-zero elements The values ​​are updated, but other positions in the row that were originally zero are forced to remain zero in order to maintain sparsity. Constraints; for the learning dictionary matrix The above process is repeated for each atom until all atoms and their corresponding sparse coefficient vectors are updated.

[0026] The beneficial effects of this invention are:

[0027] This invention significantly suppresses mode aliasing by introducing the IREMD method, thus ensuring higher-precision signal decomposition. Based on this, it distinguishes between noisy and effective signal components by analyzing the differences in MPE values ​​of each component, providing strong support for improving the accuracy of adaptive wavelet threshold denoising. Furthermore, based on the obtained denoised signal, a second denoising is performed using the LKSVD method, effectively suppressing noise in the OLTC vibration signal and improving the vibration signal quality. Verification shows that compared with other denoising methods, the OLTC vibration signal denoising method based on IREMD-MPE-LKSVD proposed in this invention exhibits particularly outstanding denoising performance, effectively preserving the key characteristics of the OLTC vibration signal and providing a reliable solution for OLTC condition monitoring in high-noise environments. Attached Figure Description

[0028] Figure 1 This is a flowchart of the present invention.

[0029] Figure 2 This is a performance comparison chart of the improved adaptive wavelet threshold function compared to hard and soft threshold functions.

[0030] Figure 3 This is a comparison chart of the original OLTC vibration signal and noisy signal components of the converter transformer; among them, Figure 3 (a) Corresponding reference vibration signal Figure 3 (b) Corresponding to the original converter transformer OLTC vibration signal.

[0031] Figure 4 This is a time-domain plot of each component of the IREMD decomposition of the original OLTC vibration signal of the converter transformer, constructed from simulated sub-signals.

[0032] Figure 5 This is a time-domain plot of the OLTC vibration signal after noise reduction; among which, Figure 5 (a) corresponds to a single noise reduction signal. Figure 5 (b) Corresponding to the secondary noise reduction signal.

[0033] Figure 6 These are time-domain plots of simulated vibration signals under two different operating conditions and the original OLTC vibration signal of the converter transformer; among them, Figure 6 (a) Simulated vibration signal corresponding to normal operating conditions. Figure 6 (b) Vibration signal of the original converter transformer under normal operating conditions (OLTC) Figure 6 (c) Simulated vibration signal corresponding to contact wear condition. Figure 6 (d) Vibration signal of the original converter transformer OLTC corresponding to the contact wear condition.

[0034] Figure 7This is a time-domain diagram of each component of the vibration signal of the original converter transformer OLTC under normal operating conditions.

[0035] Figure 8 This is a time-domain diagram of each component of the original OLTC vibration signal of the converter transformer under contact wear conditions.

[0036] Figure 9 These are time-domain plots of OLTC vibration signals after noise reduction under two different operating conditions; among them, Figure 9 (a) Noise reduction signal corresponding to normal operating conditions. Figure 9 (b) Secondary noise reduction signal corresponding to normal operating conditions. Figure 9 (c) Noise reduction signal corresponding to contact wear condition. Figure 9 (d) Secondary noise reduction signal corresponding to contact wear condition.

[0037] Figure 10 These are time-domain plots of OLTC vibration signals after noise reduction using different methods under normal operating conditions; among them, Figure 10 (a) Corresponds to EMD-MPE-WT, Figure 10 (b) Corresponding to EEMD-MPE-WT, Figure 10 (c) Corresponding to CEEMDAN-MPE-WT, Figure 10 (d) Corresponds to LCD-RCMDE-SVD, Figure 10 (e) Corresponds to APSO-SSD-SVD, Figure 10 (f) Corresponding to the method of the present invention.

[0038] Figure 11 These are time-domain plots of OLTC vibration signals after noise reduction using different methods under contact wear conditions; among them, Figure 11 (a) Corresponds to EMD-MPE-WT, Figure 11 (b) Corresponding to EEMD-MPE-WT, Figure 11 (c) Corresponding to CEEMDAN-MPE-WT, Figure 11 (d) Corresponds to LCD-RCMDE-SVD, Figure 11 (e) Corresponds to APSO-SSD-SVD, Figure 11 (f) Corresponding to the method of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0040] Example 1: As Figures 1-11 As shown, an OLTC vibration signal noise reduction method based on IREMD-MPE-LKSVD includes:

[0041] Obtain the original OLTC vibration signal of the converter transformer;

[0042] The original OLTC vibration signal of the converter transformer was adaptively decomposed into the following values ​​using the IREMD method: One IMF component and one residual signal component (i.e., RES component);

[0043] calculate The MPE values ​​of one IMF component and one residual signal component are used to determine the component with an MPE value greater than a first preset threshold as a noisy signal component, and the rest as valid signal components.

[0044] The noisy signal component is processed by an improved adaptive wavelet threshold denoising method, and the noisy signal component processed by the improved adaptive wavelet threshold denoising method is reconstructed with the effective signal component to obtain a first-stage denoised signal.

[0045] The LKSVD method is used to perform secondary denoising on the primary denoised signal to obtain the denoised OLTC vibration signal.

[0046] Furthermore, the IREMD method is based on the REMD method and is constructed by introducing an adaptive convergence criterion to optimize the screening stopping criterion and adding an orthogonality check.

[0047] Furthermore, the adaptive convergence criterion is as follows: calculate the standardized squared difference of the sum of two consecutive screened components, and stop the screening process of IMF components when the standardized squared difference is less than the adaptive convergence threshold.

[0048] This invention optimizes the screening stopping criterion by introducing an adaptive convergence criterion, thereby forming the screening stopping criterion in the IREMD method, specifically including:

[0049] (1) Criteria for the number of extreme points and zero points: The number of extreme points and zero points should be consistent, or the difference between the two should not exceed 1.

[0050] (2) Local Minimum Criterion for Objective Function Value: During the screening process, when and When established, it is assumed that the objective function value has been found during the screening process. The local minimum of ; where, , , The first , , The objective function value of the next iteration.

[0051] (3) Adaptive convergence criterion: Calculate the components selected in two consecutive screenings. and Standardized squared difference When standardized squared difference Less than the adaptive convergence threshold They believe that the current IMF weighting is stable enough.

[0052] ;

[0053] In the formula, The original converter transformer OLTC vibration signal through The filtered components after the next iteration; This represents the total length of the signal.

[0054] The screening process stops when any of the above screening stopping criteria is met, and the updated and retained objective function value during the iteration process is output. Optimal selection component that achieves local minimum As a candidate for IMF component.

[0055] Furthermore, the expression for the objective function value in the screening stopping criterion is:

[0056] ;

[0057] ;

[0058] ;

[0059] In the formula, For the first The objective function value of the next iteration; The original converter transformer OLTC vibration signal through The selected components after the next iteration. , Indicates the maximum number of iterations. For sampling points; for The root mean square of; for Excess kurtosis; This is the total length of the signal; for The average value; weight parameters .

[0060] It should be noted that in the IREMD method, the screening stopping criterion is determined by... Determine, and require The value should be as small as possible, but this invention aims to balance... and The influence of different signal characteristics was investigated by constructing an expression for the objective function using weighted summation, which achieved the effect of adjusting the relative importance of the two.

[0061] Furthermore, the decomposition process of the IREMD method is as follows:

[0062] S2.1 Initialization:

[0063] Import the original OLTC vibration signal of the converter transformer Set the current IMF component index. Let the signal to be decomposed Initialize the modal component list. (Used to store the extracted IMF components). Set weight parameters. Adaptive convergence threshold and orthogonality threshold .

[0064] S2.2 Initialization of the screening process:

[0065] Initialize the number of filtering iterations before each IMF component extraction. Let the current filtered component .initialization The objective function value of the next iteration , The objective function value of the next iteration ,current The objective function value of the next iteration At the same time, initialize the minimum objective function value. and optimal filtering components Used to record the objective function value during the screening process. The local minimum and its corresponding components.

[0066] S2.3, Main Filtering Loop (Extracting IMF Component Candidates):

[0067] Repeat the following steps until any of the above "screening stopping criteria" are met:

[0068] S2.3.1 Find the mean envelope Identify and filter components All local maxima and minima are identified; the upper envelope is obtained by connecting the maxima and minima using natural cubic spline interpolation. and lower envelope Calculate the mean envelope :

[0069] ;

[0070] S2.3.2. Based on the mean envelope, obtain the new screening components. :

[0071] ;

[0072] S2.3.3, Calculation And update the best filter component:

[0073] based on calculate and Then calculate .if Then update and .

[0074] S2.3.4, Inspection Does it meet any of the criteria in the "Screening Stop Criteria"?

[0075] If satisfied, stop the filtering loop and... As a current IMF weight candidate .

[0076] If the screening stopping criteria are not met, then ,Will As a new and update , Continue executing S2.3.1.

[0077] S2.4 When IMF component candidates are obtained Then, perform the following orthogonality check:

[0078] if Or when If empty, then accept the current IMF component candidate. As the first One IMF component is added to the modal component list. In; among them, Indicates the current IMF component candidate and Orthogonality index between The maximum value in, List of modal components All extracted IMF components;

[0079] Otherwise, consider the current IMF weight candidates Severe mode aliasing exists. Therefore, the current IMF component candidate should be abandoned. To avoid falling into an infinite loop, in the current signal to be decomposed Add Gaussian white noise, then return to step S2.2 to restart the filtering process for this perturbed signal until a candidate IMF component that satisfies the orthogonality condition is successfully obtained. Or the number of retries has been reached.

[0080] The above are the current IMF weighted candidates. and Orthogonality index between The expression is:

[0081] .

[0082] S2.5, Iteratively extract the next IMF component:

[0083] If the current IMF component candidate If accepted, update the residual signal: .

[0084] make Replace with and set .

[0085] Repeat steps S2.2 to S2.5 until a new signal to be decomposed is obtained. Until it exhibits monotonic or constant value characteristics (at this point) That is, the final residual.

[0086] When the decomposition process ends, the following is obtained: There are 1 IMF components, and the final residual signal components are expressed as follows: Therefore, the original converter transformer OLTC vibration signal It can be represented as:

[0087] ;

[0088] In the formula, The original OLTC vibration signal of the converter transformer. The total number of IMF components. The first one obtained after decomposition One IMF component that meets the conditions ( ).

[0089] Through the above-mentioned improved and optimized process for the REMD method, the IREMD method is made more suitable for the OLTC vibration signal decomposition task of UHV converter transformers, ensuring that the IMF components decomposed by the algorithm are optimized and conducive to the subsequent calculation and screening of multi-scale permutation entropy values ​​under complex field conditions.

[0090] Furthermore, the aforementioned The MPE values ​​of each IMF component and one residual signal component are calculated using multi-scale permutation entropy. Under the measure of multi-scale permutation entropy, noisy signal components and effective signal components exhibit significant differences. Specifically, the introduction of noise enhances the randomness of the signal, resulting in noisy signal components having higher permutation entropy values. In contrast, the permutation entropy of the effective signal component, which contains true information, evolves more regularly across different scales, typically exhibiting lower entropy values. Therefore, the MPE value introduced in this invention can effectively distinguish between noisy and effective signal components. This method helps to obtain the characteristic information of the OLTC vibration signal of the converter transformer at different scales. By calculating the MPE values ​​of each IMF component and residual signal component after decomposition using the IREMD method, the complexity of the time series is quantified into a value between 0 and 1. A value close to 1 indicates high randomness; while a value close to 0 indicates strong regularity. For noisy signal components with an MPE value close to 1, denoising algorithms should be used for subsequent noise reduction. For valid signal components with an MPE value close to 0, they should be retained for subsequent signal reconstruction. In this embodiment, a first preset threshold of 0.6 is set. IMF components with an MPE value greater than or equal to 0.6 are determined to be noisy signal components, while components with an MPE value less than 0.6 are determined to be valid signal components.

[0091] Furthermore, the noisy signal components selected by the MPE value are mainly noise, but also contain weak vibration signal characteristics, requiring refined denoising processing. Therefore, this invention employs an improved adaptive wavelet thresholding denoising method, which uses an improved adaptive wavelet threshold function constructed based on an adaptive wavelet threshold.

[0092] (1)Wavelet decomposition

[0093] Perform a discrete wavelet transform (DWT) on each noisy signal component individually. Decompose it into wavelet coefficients at different levels.

[0094] (2) Threshold processing

[0095] Thresholding operations are performed on the wavelet coefficients obtained from the decomposition, including adaptive wavelet thresholding and improved adaptive wavelet thresholding functions.

[0096] ① Adaptive wavelet threshold

[0097] Traditional wavelet thresholding denoising methods often use a fixed global threshold. However, the noise figure during wavelet decomposition decreases as the number of decomposition levels increases, so using a fixed threshold can negatively impact the denoising performance. The adaptive wavelet threshold is defined as:

[0098] ;

[0099] In the formula, For adaptive wavelet thresholding, It is the standard deviation of the noisy signal components. Let be the signal length, and c be the number of wavelet coefficient layers. By using an adaptive wavelet threshold, noise-related wavelet coefficients can be reduced without affecting coefficients containing valuable information.

[0100] ② Improved adaptive wavelet threshold function

[0101] To reduce the impact of soft and hard thresholding functions on signal denoising in traditional wavelet denoising methods, this invention constructs an improved adaptive wavelet thresholding function based on an adaptive wavelet threshold. The improved adaptive wavelet thresholding function is defined as follows:

[0102] ;

[0103] In the formula, After wavelet decomposition, it is located at the th The first layer One original wavelet coefficient (input wavelet coefficient), The wavelet coefficients (output wavelet coefficients) are processed by the improved adaptive wavelet threshold function. This represents the step function.

[0104] like Figure 2 The figure shows a comparison between the improved adaptive wavelet thresholding function used in this invention and hard and soft thresholding functions (the horizontal axis represents the input wavelet coefficients, and the vertical axis represents the output wavelet coefficients). Figure 2 It can be seen that, compared with hard threshold and soft threshold functions, the present invention has better smoothness and continuity.

[0105] Furthermore, the noisy signal component processed by the improved adaptive wavelet threshold denoising method is reconstructed with the effective signal component to obtain the first-stage denoised signal. Specifically, the wavelet coefficients obtained by processing with the improved adaptive wavelet threshold function are used to reconstruct the first-stage denoised noisy signal component through inverse wavelet transform (IDWT). The first-stage denoised noisy signal component and the unprocessed effective signal component are then superimposed to obtain the final first-stage denoised signal. .

[0106] Furthermore, for complex engineering vibration signals, the IREMD method combined with MPE value and improved adaptive wavelet thresholding for noise reduction cannot completely separate background noise (such as electromagnetic interference, mechanical noise, etc.) that overlaps with the characteristic spectrum of the vibration signal. Therefore, this invention employs the LKSVD method for secondary noise reduction, the specific process of which is as follows:

[0107] (1) Matrix construction

[0108] First, a sliding window processing method is used to process the noise reduction signal. Construct as a sample matrix Initialize the learning dictionary matrix (Include (atom column vectors) and sparsity ;in, Atomic length, The total number of samples; The dimension is , The dimension is .

[0109] (2) Sparse coding

[0110] Based on the constructed sample matrix In sparsity Under the constraints, the sparse coefficient matrix is ​​solved using the Orthogonal Matching Pursuit (OMP) algorithm. To minimize reconstruction error.

[0111] (3) Learning to update atoms in a dictionary matrix

[0112] The core idea of ​​the K-SVD method is to update the atoms in the learned dictionary matrix one by one. First, based on the current sparse coefficient matrix... Find the set of indices of the non-zero elements in the u-th row. Next, the overall error matrix is ​​calculated. That is, from the sample matrix Subtracting the current atom The sum of the products of all atoms except the one with the corresponding coefficient. Then, based on the position set... From the overall error matrix Extract the currently updated atoms The column vectors of the overall error matrix with non-zero coefficients are used to obtain the error submatrix. .

[0113] The traditional K-SVD method performs this step on the error submatrix. Performing a full singular value decomposition (SVD) to obtain its best rank-1 approximation is computationally expensive. To improve efficiency, the LKSVD method of this invention is based on the K-SVD method and uses power iteration to quickly solve for the error submatrix. The maximum singular value and its corresponding left and right singular vectors are used to replace the complete SVD, forming the LKSVD method. Specifically, this includes:

[0114] S5.1 Optimal Rank-1 Approximation Based on Power Iteration Method

[0115] For the error submatrix Using the power iteration method, only the largest singular value is extracted. and its corresponding left singular vector and right singular vectors The power iteration method is an efficient iterative algorithm that converges directly to the principal components of a matrix through multiple matrix multiplication iterations. This yields the error submatrix. The best rank-1 approximation:

[0116] .

[0117] S5.2, Learning Atomic Updates in a Dictionary Matrix: Using Left Singular Vectors Update the learning dictionary matrix The u-th atom in And perform L2 norm normalization to ensure that it is a unit vector: After obtaining the optimal rank-1 approximation, the maximum singular value is used. With right singular vectors The product of transposes with respect to the sparse coefficient matrix The sparse coefficient vector in the u-th row set of position indices of non-zero elements The values ​​are updated, but other positions in the row that were originally zero are forced to remain zero in order to maintain sparsity. constraint: ; Represents the sparse coefficient matrix The sparse coefficient vector in the u-th row A set of position indices for non-zero elements.

[0118] S5.3, Learning the dictionary matrix Repeat steps S5.1 and S5.2 for each atom (i.e. each column) until all atoms and their corresponding sparse coefficient vectors have been updated.

[0119] Once the preset number of iterations is reached or the error converges, the finally learned learning dictionary matrix D and sparse coefficient matrix are used. Calculate the reconstruction matrix The reconstructed matrix Each column is restored to a one-dimensional time series to obtain the final denoised OLTC vibration signal. The LKSVD method proposed in this invention can solve the problem of residual noise in vibration signals after local noise reduction processing, aiming to further improve the signal-to-noise ratio and waveform fidelity of the output signal.

[0120] Furthermore, the following explanation is based on the data:

[0121] I. Simulation Signal Verification

[0122] Due to the harsh operating environment and complex structure of the on-load tap changer of an ultra-high voltage converter transformer, its vibration signal exhibits strong nonlinearity and multi-time-scale characteristics. This invention employs a simulation analysis strategy, generating four simulated sub-signals with different frequencies and amplitudes and superimposing them to construct a reference vibration signal. This verifies the effectiveness of the proposed method in noise reduction of the OLTC vibration signal of the converter transformer. The construction of the reference vibration signal is as follows:

[0123] ;

[0124] In the formula, signal It is a low-frequency signal with an amplitude of 1 and a frequency of 5Hz; signal It is a low-to-medium frequency signal with an amplitude of 0.2 and a frequency of 40Hz; It is an intermediate frequency signal with an amplitude of 0.4 and a frequency of 50Hz; the signal It is a high-frequency signal with an initial amplitude of 0.3 and a frequency of 100Hz, whose amplitude has an attenuation effect; by superimposing the above four signals, a reference vibration signal is constructed. .

[0125] The simulation experiment sampling frequency was set to 1000Hz, corresponding to 4000 sampling points, with a total sampling duration of 4 seconds. To simulate complex noise interference in the actual signal, Gaussian white noise with a signal-to-noise ratio of 10dB was introduced into the reference vibration signal to construct the required original converter transformer OLTC vibration signal. The reference vibration signal and the original converter transformer OLTC vibration signal are shown below. Figure 3 As shown; where, Figure 3 (a) Corresponding reference vibration signal Figure 3 (b) Corresponding to the original converter transformer OLTC vibration signal.

[0126] The original OLTC vibration signal of the converter transformer was decomposed using the IREMD method, and the maximum number of iterations of the IREMD method was determined. Setting the value to 30, we obtained seven intrinsic IMF components (IMF1~IMF7) and one residual signal component (RES). The specific decomposition results are as follows: Figure 4 As shown. By Figure 4 It can be seen that the original converter transformer OLTC vibration signal did not exhibit mode aliasing after being decomposed by the IREMD method, and the number of decomposed IMF components was relatively small.

[0127] Furthermore, the present invention calculated the MPE values ​​of each IMF component, and the results are summarized in Table 1. From the results in Table 1, it can be seen that the components with earlier serial numbers have MPE values ​​closer to 1. In addition, components with higher MPE values ​​generally exhibit stronger random fluctuations. Specifically, the MPE values ​​of IMF1 and IMF2 components exceed the first preset threshold of 0.6, and are therefore identified as noisy signal components; the rest are valid signal components.

[0128] Table 1. MPE values ​​of each component

[0129]

[0130] For the two noisy signal components IMF1 and IMF2, this invention employs an improved adaptive wavelet threshold denoising method. Subsequently, the processed components are reconstructed with the effective signal components to generate a first-order denoised signal, as shown below. Figure 5 As shown in (a). The noise-reduced signal is then decomposed using the LKSVD method to complete the secondary noise reduction process, finally obtaining the OLTC vibration signal after secondary noise reduction, as shown in (a). Figure 5 As shown in (b); via Figure 5 It can be seen that the noise-reduced OLTC vibration signal obtained by the method of the present invention is close to the reference vibration signal before noise addition, indicating that the present invention effectively suppresses noise interference.

[0131] To further illustrate the effectiveness of the noise reduction method proposed in this invention, a comparative analysis was conducted with EMD-MPE-WT, EEMD-MPE-WT, CEEMDAN-MPE-WT, LCD-RCMDE-SVD, and APSO-SSD-SVD noise reduction methods. By introducing root mean square error (RMSE), signal-to-noise ratio (SNR), and waveform similarity parameter (NCC) as evaluation indicators, the comparison results are shown in Table 2. As can be seen from Table 2, the method proposed in this invention has a smaller RMSE, a larger SNR and NCC compared to the other methods, and the NCC value is closest to 1, indicating a very high similarity to the reference vibration signal, demonstrating excellent noise reduction performance.

[0132] Table 2 Comparison of noise reduction performance indicators of various methods

[0133]

[0134] To better simulate the actual operation of the OLTC converter transformer, this invention selected Gaussian white noise, arc noise, and a mixture of the two noises for simulation analysis, and obtained the noise reduction effects of different methods under different noise environments as shown in Table 3. As can be seen from Table 3, the noise reduction effect of the method proposed in this invention is superior to other methods under different noise interferences, and it can preserve the characteristic information of the OLTC vibration signal to the greatest extent.

[0135] Table 3. Comparison of noise reduction effects of different methods under different noise environments

[0136]

[0137] II. Experimental Analysis

[0138] This invention utilizes an ABB-manufactured UCLRE on-load tap changer to conduct prototype simulation tests on a standby converter transformer. Normal and fault conditions were simulated for the OLTC prototype, with OLTC contact wear selected as the fault condition. An AD100T vibration sensor was chosen, with a sensitivity of 100mV / g, a frequency response range of 0.3~15000Hz, and a range of ±50g. Four vibration sensors were installed on the top cover protrusions and edges of the OLTC prototype. Three sensors were installed on the top cover edge, and one on the top cover edge protrusion. This multi-point synchronous measurement ensured that the collected vibration signals comprehensively and accurately reflected the OLTC's operating characteristics. The analog signals collected by the vibration sensors were input into an 8-channel data acquisition card (100kHz frequency) and stored in a computer via USB.

[0139] In practical applications, it is difficult to collect the actual vibration signal of the operating converter transformer OLTC. However, the OLTC vibration signal collected from the standby converter transformer has relatively low noise and can be used as a simulated vibration signal. To simulate the actual operating conditions of the OLTC, this invention adds 10dB of mixed noise, consisting of Gaussian white noise and arc noise, to the simulated vibration signal collected from the standby converter transformer OLTC, thus constructing the required original converter transformer OLTC vibration signal. The simulated vibration signals and the original converter transformer OLTC vibration signals under the two operating conditions are shown below. Figure 6 As shown; where, Figure 6 (a) Simulated vibration signal corresponding to normal operating conditions. Figure 6 (b) Vibration signal of the original converter transformer under normal operating conditions (OLTC) Figure 6 (c) Simulated vibration signal corresponding to contact wear condition. Figure 6 (d) Vibration signal of the original converter transformer OLTC corresponding to the contact wear condition.

[0140] The original OLTC vibration signal of the converter transformer was decomposed using the IREMD method to obtain a series of IMF components. For both operating conditions, 10 components were decomposed, including 9 IMF components (IMF1 to IMF9) and a residual signal component (RES). The waveforms of all these components in the time domain are shown below. Figure 7 , Figure 8 As shown. By Figure 7 , Figure 8It can be seen that the number of decomposition components is the same under both operating conditions of OLTC using the IREMD method.

[0141] Next, the MPE values ​​of each IMF component and residual signal component are calculated to distinguish between noisy and effective signal components. In the MPE value calculation process, three key parameters need to be determined beforehand: embedding dimension, time delay, and scale factor. The rationality of parameter settings, especially the scale factor and embedding dimension, plays a decisive role in the coarse-grained processing effect of time series data. If the scale factor is too small, it will restrict the full description of the inherent characteristics of the time series; however, if its value is too large, it may obscure the essential differences in complexity between different sequences. Regarding the embedding dimension, too small a dimension will cause feature loss; too large a dimension may increase the computational load of the model. Through experiments, the optimal parameter combination was determined to be: embedding dimension set to 5, time delay set to 1, and scale factor set to 3. Under these conditions, the calculated MPE value is the most stable and the disturbance to the original vibration signal is minimal. The calculated MPE values ​​are shown in Table 4.

[0142] Table 4. MPE values ​​of each component after IREMD method decomposition.

[0143]

[0144] Analysis of Table 4 shows that the MPE values ​​of IMF1 and IMF2 components are higher than 0.6 under both operating conditions, indicating that the noise components are mainly concentrated in these two components. Therefore, this invention uses an improved adaptive wavelet threshold denoising method to process the IMF1 and IMF2 components and reconstructs them with the effective signal components to obtain a first-order denoised signal. Subsequently, the first-order denoised signal is decomposed using LKSVD, and the waveforms of the denoised signals under the two operating conditions in the time domain are shown below. Figure 9 As shown; Figure 9 (a) OLTC vibration signal after noise reduction under normal operating conditions Figure 9 (b) OLTC vibration signal after secondary noise reduction under normal operating conditions Figure 9 (c) OLTC vibration signal after noise reduction under corresponding contact wear conditions. Figure 9 (d) OLTC vibration signal after secondary noise reduction corresponding to contact wear conditions. Figure 9 It can be seen that, compared with the original OLTC vibration signal and the primary denoising signal, the secondary denoising of the LKSVD method further filters out residual noise that overlaps with the signal characteristic spectrum, making the transient impact characteristics during OLTC switching more prominent, and the signal amplitude does not show significant attenuation. This result verifies that the LKSVD method can effectively solve the problem of residual noise after local denoising, significantly improve the quality of the output signal, and meet the requirements of high-precision fault diagnosis.

[0145] To evaluate the performance of the noise reduction method of this invention, it was compared and analyzed with EMD-MPE-WT, EEMD-MPE-WT, CEEMDAN-MPE-WT, LCD-RCMDE-SVD, and APSO-SSD-SVD methods. All methods used the same experimental procedures to process the original OLTC vibration signals of the converter transformer under two operating conditions. The results are as follows: Figure 10 , Figure 11 As shown; where, Figure 10 (a) Figure 11 (a) Corresponds to EMD-MPE-WT, Figure 10 (b) Figure 11 (b) Corresponding to EEMD-MPE-WT, Figure 10 (c) Figure 11 (c) Corresponding to CEEMDAN-MPE-WT, Figure 10 (d) Figure 11 (d) Corresponds to LCD-RCMDE-SVD, Figure 10 (e) Figure 11 (e) Corresponds to APSO-SSD-SVD, Figure 10 (f) Figure 11 (f) Corresponding to the method of this invention. By Figure 10 , Figure 11 It can be seen that the original OLTC vibration signals of the converter transformer under both OLTC operating conditions have the highest similarity to the simulated vibration signals after being denoised by the method of this invention, and the denoising effect is better than other methods.

[0146] To comprehensively evaluate the performance differences of different methods in noise reduction, this invention introduces RMSE, SNR, and NCC as key evaluation indicators. The comparison results of each indicator are shown in Table 5. Table 5 shows that the root mean square error (RMSE) of the original OLTC vibration signals of the converter transformer under both OLTC operating conditions after noise reduction using the method of this invention is lower than that of other methods, with 1.0537 for normal operating conditions and 0.5871 for contact wear conditions. The signal-to-noise ratio (SNR) is higher than that of other methods, with 17.04 for normal operating conditions and 16.80 for contact wear conditions. The similarity to simulated vibration signals is highest, with the waveform similarity parameter NCC being 0.9901 for normal operating conditions and 0.9895 for contact wear conditions. As demonstrated above, when using the method of this invention to denoise the original OLTC vibration signals of the converter transformer under both OLTC operating conditions, useless noise signals can be removed, while retaining the useful information of the vibration signal to the greatest extent. Experiments have shown that the noise reduction method for OLTC vibration signals based on IREMD-MPE-LKSVD proposed in this invention has better noise reduction waveform and noise reduction performance than other methods, achieving the best balance between noise suppression and feature preservation, and providing reliable technical support for accurate fault diagnosis of UHV converter transformer OLTC.

[0147] Table 5 Comparison of noise reduction performance indicators of different methods

[0148]

[0149] The following conclusions were drawn through experimental verification:

[0150] (1) In view of the common mode aliasing problem in the traditional EMD method when processing OLTC fault vibration signals, the present invention introduces the IREMD method, which can significantly suppress the mode aliasing phenomenon, thereby ensuring higher accuracy signal decomposition.

[0151] (2) The MPE value can effectively capture the nonlinear characteristics of the OLTC vibration signal of the UHV converter transformer at different time scales. By analyzing the differences in the MPE values ​​of each component, the noisy signal component and the effective signal component can be distinguished, thus providing strong support for improving the accurate noise reduction of the adaptive wavelet threshold noise reduction method.

[0152] (3) The present invention performs secondary noise reduction using the LKSVD method, which effectively suppresses the noise of the OLTC vibration signal and improves the quality of the vibration signal.

[0153] (4) Compared with other noise reduction methods, the noise reduction method of OLTC vibration signal based on IREMD-MPE-LKSVD proposed in this invention has shown particularly good noise reduction effect, effectively preserving the key features of OLTC vibration signal, and providing a reliable solution for OLTC condition monitoring of converter transformer under strong noise background.

[0154] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for denoising OLTC vibration signals based on IREMD-MPE-LKSVD, characterized in that, include: Obtain the original OLTC vibration signal of the converter transformer; The original OLTC vibration signal of the converter transformer was adaptively decomposed into the following values ​​using the IREMD method: One IMF component and one residual signal component; calculate The MPE values ​​of one IMF component and one residual signal component are used to determine the component with an MPE value greater than a first preset threshold as a noisy signal component, and the rest as valid signal components. The noisy signal component is processed by an improved adaptive wavelet threshold denoising method, and the noisy signal component processed by the improved adaptive wavelet threshold denoising method is reconstructed with the effective signal component to obtain a first-stage denoised signal. The LKSVD method is used to perform secondary denoising on the primary denoised signal to obtain the denoised OLTC vibration signal.

2. The OLTC vibration signal noise reduction method based on IREMD-MPE-LKSVD according to claim 1, characterized in that, The IREMD method is based on the REMD method and is constructed by introducing an adaptive convergence criterion to optimize the screening stopping criterion and adding an orthogonality check.

3. The OLTC vibration signal noise reduction method based on IREMD-MPE-LKSVD according to claim 2, characterized in that, The adaptive convergence criterion is as follows: calculate the standardized squared difference of the sum of two consecutive screened components, and stop the screening process of IMF components when the standardized squared difference is less than the adaptive convergence threshold.

4. The OLTC vibration signal noise reduction method based on IREMD-MPE-LKSVD according to claim 2, characterized in that, The orthogonality check specifically involves: if Or when the modal component list If empty, then accept the current IMF component candidate. As the first One IMF component is added to the modal component list. In; among them, Indicates the current IMF component candidate and Orthogonality index between The maximum value in, List of modal components All extracted IMF components; Otherwise, abandon the current IMF weighting candidate. .

5. The OLTC vibration signal noise reduction method based on IREMD-MPE-LKSVD according to claim 1, characterized in that, The improved adaptive wavelet thresholding method employs an improved adaptive wavelet threshold function constructed based on the adaptive wavelet threshold.

6. The OLTC vibration signal noise reduction method based on IREMD-MPE-LKSVD according to claim 5, characterized in that, The improved adaptive wavelet threshold function is defined as follows: ； In the formula, After wavelet decomposition, it is located at the th The first layer The original wavelet coefficients, These are the wavelet coefficients after processing with the improved adaptive wavelet threshold function; Represents the step function; It is an adaptive wavelet threshold.

7. The OLTC vibration signal noise reduction method based on IREMD-MPE-LKSVD according to claim 6, characterized in that, The adaptive wavelet threshold is defined as follows: ； In the formula, It is the standard deviation of the noisy signal components. Where c is the signal length and c is the number of wavelet coefficient layers.

8. The OLTC vibration signal noise reduction method based on IREMD-MPE-LKSVD according to claim 1, characterized in that, The LKSVD method is based on the K-SVD method and uses the power iteration method to solve for the error submatrix. Obtain the maximum singular value and its corresponding left singular vector. and right singular vector This is how the error submatrix is ​​obtained. The best rank-1 approximation; using left singular vectors Update the learning dictionary matrix The u-th atom in And perform L2 norm normalization; utilize the maximum singular value With right singular vectors The product of transposes with respect to the sparse coefficient matrix The sparse coefficient vector in the u-th row set of position indices of non-zero elements The values ​​are updated, but other positions in the row that were originally zero are forced to remain zero in order to maintain sparsity. Constraints; for the learning dictionary matrix The above process is repeated for each atom until all atoms and their corresponding sparse coefficient vectors are updated.

Images