Extra-high voltage transformer partial discharge signal denoising method and system

Abstract

The invention relates to the technical field of power equipment state monitoring, and discloses an extra-high voltage transformer partial discharge signal denoising method and system.The method comprises the steps that firstly, the high-frequency interference energy proportion is detected in real time, the wavelet packet decomposition layer number is dynamically adjusted, and the high-frequency interference fine isolation capacity is improved; a clustering combination threshold value is dynamically adjusted based on a noise cluster proportion, and an interference frequency band is accurately identified; and finally, reconstructing the signal through an inverse wavelet packet, and carrying out optimized output. According to the method, the signal-to-noise ratio can be remarkably improved in a complex noise environment, the pulse detection rate is improved, meanwhile, real-time optimization is achieved through sliding window processing and FPGA acceleration, and efficient and reliable technical support is provided for online monitoring of the insulation state of the extra-high voltage transformer.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment condition monitoring, and particularly to a method and system for denoising partial discharge signals of ultra-high voltage transformers. Background Art

[0002] Monitoring partial discharge of ultra-high voltage transformers is the core means to evaluate their insulation status. However, on-site high-frequency detection faces severe challenges in complex electromagnetic environments such as power frequency harmonics, random pulse noise, and radio frequency band interference, resulting in serious signal distortion. Among existing methods, the fixed-frequency band filtering method relies on a preset frequency band range, such as 30 - 100 MHz, and cannot dynamically adapt to sudden interferences, such as 28 MHz radio noise. Field tests show that its effective signal loss rate exceeds 30%, and its high-frequency noise suppression ability is insufficient. Wavelet threshold denoising requires manual setting of the decomposition layer number and threshold parameters, and the misdeletion rate of weak discharge pulses with amplitudes lower than 0.5 mV is as high as 40%, and high-frequency noise (>20 MHz) residue is significant. Although empirical mode decomposition (EMD) has self-adaptability, it has mode mixing and end effects, especially in the presence of strong power frequency interference, the signal separation accuracy drops sharply. Traditional methods generally have two major defects: First, parameter fixation leads to the inability to respond to the time-frequency characteristic changes of non-stationary interferences. For example, the fixed wavelet decomposition layer number results in insufficient frequency band resolution or computational redundancy. Second, the dimensional singularity leads to the separation of time-domain and frequency-domain information. For example, the time-domain threshold method ignores the energy focusing law of frequency bands, and the frequency-domain filtering method is difficult to capture the time-domain mutation characteristics of transient pulses. It can be seen that it is difficult to achieve high-fidelity extraction of weak discharge signals in complex electromagnetic environments in the prior art. Summary of the Invention

[0003] The present invention provides a method and system for denoising partial discharge signals of ultra-high voltage transformers to solve the problem that it is difficult to achieve high-fidelity extraction of weak discharge signals in complex electromagnetic environments in the prior art.

[0004] To achieve the above object, the present invention is realized through the following technical solutions:

[0005] In a first aspect, the present invention provides a method for denoising partial discharge signals of ultra-high voltage transformers, including:

[0006] S1: Collect partial discharge signals of the transformer and preprocess the electrical signals;

[0007] S2: Perform dynamic wavelet packet decomposition on the preprocessed electrical signals, and extract sub-band energy characteristics from the results of the dynamic wavelet packet decomposition;

[0008] S3: Perform adaptive hierarchical clustering based on the extraction results; according to the hierarchical clustering results, divide the sub-bands into high-energy discharge clusters and low-energy noise clusters, and filter out the frequency bands corresponding to the low-energy noise clusters;

[0009] S4: Reconstruct and optimize based on the wavelet packet nodes corresponding to the high-energy discharge clusters to obtain the finally output denoised signal.

[0010] In a second aspect, the present application provides a partial discharge signal denoising system for an UHV transformer, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the method described in the first aspect above are implemented.

[0011] Beneficial effects:

[0012] The partial discharge signal denoising method for an UHV transformer provided by the present invention first dynamically adjusts the wavelet packet decomposition layer number by real-time detecting the high-frequency interference energy ratio to enhance the refined isolation ability of high-frequency interference; and dynamically adjusts the clustering and merging threshold based on the noise cluster ratio to accurately identify the interference frequency band; finally, reconstructs the signal through inverse wavelet packet and optimizes the output. It can significantly improve the signal-to-noise ratio in a complex noise environment, improve the pulse detection rate, and at the same time optimize the real-time performance through sliding window processing and FPGA acceleration, providing an efficient and reliable technical support for the on-line monitoring of the insulation state of UHV transformers.

[0013] In a further technical solution, a hybrid distance metric strategy combining weighted Euclidean distance and cosine similarity can accurately distinguish the energy distribution characteristics of low-frequency and high-frequency subbands. Description of the drawings

[0014] Figure 1 One of the flowcharts of a partial discharge signal denoising method for an UHV transformer according to a preferred embodiment of the present invention;

[0015] Figure 2 Another flowchart of a partial discharge signal denoising method for an UHV transformer according to a preferred embodiment of the present invention. Detailed implementation manners

[0016] The technical solutions of the present invention will be described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments of the present invention belong to the scope of protection of the present invention.

[0017] Unless otherwise defined, technical terms or scientific terms used in the present invention shall have the ordinary meanings understood by those of ordinary skill in the field to which the present invention pertains. The "first", "second" and similar terms used in the present invention do not denote any order, quantity or importance, but are only used to distinguish different components. Similarly, similar terms such as "a" or "one" do not denote a quantity limitation, but mean that there is at least one. The terms such as "connected" or "linked" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms such as "upper", "lower", "left" and "right" are only used to indicate relative position relationships. When the absolute position of the object being described changes, the relative position relationships also change accordingly.

[0018] It should be understood that traditional partial discharge monitoring methods (such as fixed-frequency band filtering, wavelet threshold denoising, etc.) have significant defects: relying on artificially preset parameters (such as decomposition levels, clustering thresholds), insufficient frequency band resolution, and being unable to dynamically adapt to the time-frequency characteristic changes of complex electromagnetic interference, resulting in weak discharge signals being easily submerged by noise. Based on this, the present application provides a method for denoising partial discharge signals of ultra-high voltage transformers. By dynamically analyzing the energy of frequency bands to extract signal features, combining a hybrid metric of weighted Euclidean distance and cosine similarity, and a threshold adjustment strategy driven by the noise proportion, accurate identification and filtering of interference frequency bands are achieved. This method overcomes the limitations of traditional methods such as fixed parameters and single dimension, significantly improves the denoising accuracy and real-time performance of partial discharge signals in the complex noise environment on-site of ultra-high voltage transformers, and provides reliable technical support for the online monitoring of the insulation status of power equipment.

[0019] Please refer to Figure 1 - Figure 2 , a method for denoising partial discharge signals of ultra-high voltage transformers provided by the present application, includes:

[0020] S1: Collect partial discharge signals of the transformer and preprocess the electrical signals;

[0021] S2: Perform dynamic wavelet packet decomposition on the preprocessed electrical signals, and extract sub-band energy features from the results of the dynamic wavelet packet decomposition;

[0022] S3: Perform adaptive hierarchical clustering based on the extraction results; according to the hierarchical clustering results, divide the sub-bands into high-energy discharge clusters and low-energy noise clusters, and filter out the frequency bands corresponding to the low-energy noise clusters;

[0023] S4: Reconstruct and optimize based on the wavelet packet nodes corresponding to the high-energy discharge clusters to obtain the finally output denoised signal.

[0024] The above-mentioned method for denoising partial discharge signals of UHV transformers first dynamically adjusts the wavelet packet decomposition level by detecting the proportion of high-frequency interference energy in real time, improving the refined isolation ability of high-frequency interference; and dynamically adjusts the clustering and merging threshold based on the proportion of noise clusters to accurately identify the interference frequency band; finally, reconstructs the signal through inverse wavelet packet and optimizes the output. It can significantly improve the signal-to-noise ratio in a complex noise environment, improve the pulse detection rate, and at the same time optimize the real-time performance through sliding window processing and FPGA acceleration, providing an efficient and reliable technical support for the on-line monitoring of the insulation state of UHV transformers.

[0025] Next, with two complete examples, the steps of the above-mentioned method for denoising partial discharge signals of UHV transformers will be described in detail:

[0026] Example 1

[0027] 1. Signal preprocessing and dynamic wavelet packet decomposition

[0028] Signal acquisition and preprocessing: Collect the partial discharge signals of the transformer through UHF sensors, and perform baseline correction and normalization processing on the original signals to eliminate DC offset and dimensional differences.

[0029] Dynamic wavelet packet decomposition:

[0030] Wavelet basis selection: Select the Db4 wavelet basis function, taking into account both time-frequency resolution and computational efficiency;

[0031] Dynamic adjustment of decomposition level:

[0032] Trigger condition and adjustment strategy

[0033] The dynamic adjustment of the decomposition level is based on the real-time detection and state machine control of the proportion of high-frequency interference energy. The specific process is as follows: First, perform a fast Fourier transform on the preprocessed signal to calculate the energy proportion in the 20–30 MHz frequency band:

[0034]

[0035] where X(f) represents the complex component of the signal at frequency f, and |X(f)| 2 represents the energy at frequency f

[0036] If R HF ≥ 15%, it is determined that there is significant high-frequency interference (such as radio noise), and the decomposition level is triggered to increase from the default 3 layers (generating 8 sub-bands) to 4 layers (generating 16 sub-bands) to improve the high-frequency band resolution from 6.25 MHz to 3.125 MHz and accurately isolate the interference; if R is detected in 5 consecutive analysis windows (each window is 1 ms) HFIf it is ≤ 10%, it will return to 3 layers to reduce the computational cost. By introducing a hysteresis interval (layer-up threshold 15%, layer-down threshold 10%), frequent switching is avoided to ensure the stability of the algorithm.

[0037] Algorithm implementation and mathematical verification

[0038] Dynamic adjustment is achieved through state machine control, and its mathematical core is the formula for the bandwidth of the frequency band:

[0039]

[0040] f s represents the sampling frequency of the signal.

[0041] When the number of layers n = 3, the sub-bandwidth is 6.25 MHz, and the computational complexity is O(Nlog N ); when the number of layers increases to n = 4, the sub-bandwidth shrinks to 3.125 MHz, and the computational complexity increases by 30%, but the high-frequency interference suppression rate increases from 68% to 92%. To ensure real-time performance, sliding window processing (segmented every 1 ms) and pre-generated filter bank optimization are adopted. When deployed in hardware, parallel acceleration can be achieved by using FPGA, reducing the single processing time from 3.8 ms for fixed 4 layers to 3.1 ms for dynamic adjustment, and the memory occupancy remains < 50 KB. This design solves the problem of frequency band confusion or computational redundancy caused by fixed number of layers through dynamic balance of resolution and efficiency.

[0042] 2. Adaptive hierarchical clustering for identifying interference frequency bands

[0043] (1) Feature vector construction: Use the energy value Ei of each sub-band node and its energy proportion Pi = Ei / ∑Ei as the feature vector to characterize the energy distribution characteristics of the frequency band.

[0044] (2) Hybrid distance metric strategy:

[0045] Low-frequency sub-bands (< 10 MHz): Use weighted Euclidean distance to strengthen the difference in energy proportion:

[0046]

[0047] In the formula, E j represents the total energy of the jth sub-band obtained by wavelet packet decomposition, and P j represents the proportion of the energy of the jth sub-band in the total energy of all sub-bands.

[0048] High-frequency sub-bands (≥ 10 MHz): Use cosine similarity to focus on the consistency of energy distribution patterns:

[0049]

[0050] W iis the coefficient vector of the ith subband.

[0051] Dynamic merge threshold adjustment:

[0052] Automatically adjust the inter-cluster distance threshold based on the noise intensity (the proportion of low-energy cluster nodes Rnoise):

[0053] δ=0.1×max(D)×(1+0.5×R noise );

[0054] When the noise ratio increases, increase the threshold to strictly separate the interference frequency band.

[0055] This threshold adjustment formula can effectively distinguish high-energy discharge clusters from low-energy noise clusters in complex noise environments, and is highly coordinated with core methods such as dynamic decomposition and hybrid distance measurement.

[0056] (4) Clustering result determination:

[0057] According to the hierarchical clustering results, the sub-bands were divided into high-energy-discharge clusters (energy proportion > 85%) and low-energy-noise clusters (energy proportion ≤ 15%), and the frequency bands corresponding to the latter were filtered out.

[0058] 3. Signal dynamic reconstruction and optimization

[0059] In the signal reconstruction stage, the wavelet packet nodes corresponding to the high-energy clusters are first selected based on the hierarchical clustering results, and the low-energy noise cluster nodes are eliminated; then the inverse wavelet packet transform is performed to reconstruct the signal: from the nth layer (the number of decomposition layers after dynamic adjustment) to the 0th layer, each layer is synthesized by a dual-channel filter bank (low-pass With Qualcomm ) to perform upsampling and signal superposition to ensure the reconstruction formula The strict mathematical implementation of , and the fidelity is guaranteed by energy conservation verification (the total energy deviation between the reconstructed signal and the retained node is <5%) and spectrum similarity comparison (correlation coefficient>0.9). After reconstruction, time-frequency joint filtering optimization is further adopted, where x rec The reconstructed denoised signal, After the nth layer of wavelet packet decomposition, the wavelet packet coefficient of the kth sub-band is, The reconstruction filter corresponding to the kth subband is:

[0060] (1) Frequency domain mask processing: The spectral coefficients of the noise frequency band (low energy cluster) determined by clustering are set to zero to completely eliminate residual interference;

[0061] (2) Time-domain adaptive smoothing: Dynamically select the sliding average window length according to the local signal-to-noise ratio - use a 5-point window (100 ns) to suppress high-frequency glitches in the low signal-to-noise ratio region (SNR < 10 dB), and use a 10-point window (200 ns) to retain pulse details in the high signal-to-noise ratio region (SNR ≥ 10 dB);

[0062] (3) Pulse enhancement: Apply the non-linear amplification function y(t) = sign(x(t))·|x(t)| 1.5 , enhance the discharge pulse amplitude in the time domain, improve the visual recognition of weak signals (amplitude ≥ 0.2 mV), and avoid amplifying high-frequency noise at the same time. This three-step optimization strategy suppresses residual noise while maximizing the retention and enhancement of the time-frequency characteristics of discharge pulses, and finally outputs a high-fidelity denoised signal. x(t) represents the intermediate signal after the time-frequency joint filtering optimization process.

[0063] 4. Algorithm verification and performance evaluation

[0064] Simulation data test: Generate a composite signal containing Gaussian white noise (SNR = 8 dB), power frequency harmonics, and 28 MHz radio interference in MATLAB, and compare the signal-to-noise ratio (SNR improvement ≥ 16 dB) and waveform similarity (SSIM ≥ 0.95) before and after denoising.

[0065] Example 2

[0066] 1. Signal acquisition and preprocessing

[0067] Collect partial discharge signals of UHV transformers through UHF sensors (frequency band 300 MHz - 1.5 GHz, model UHF-2000), set the sampling rate to 50 MHz (sampling interval 20 ns), and the single acquisition duration to 1 ms (50,000 points). Perform the following preprocessing on the original signal:

[0068] (1) Baseline correction: Calculate the signal mean and subtract the DC component to eliminate the zero drift of the sensor;

[0069] (2) Normalization processing: Scale the signal amplitude to the range [-1, 1];

[0070] (3) Power frequency filtering: Use a 50 Hz digital notch filter (FIR order 101) to suppress power frequency and its harmonic interference and retain high-frequency discharge pulses.

[0071] 2. Dynamic wavelet packet decomposition and energy feature extraction

[0072] (1) Decomposition parameter configuration:

[0073] Wavelet basis: Db4, filter length 8 points;

[0074] Basic decomposition layers: 3 layers (generating 8 sub-bands), increased to 4 layers (generating 16 sub-bands) when high-frequency interference is triggered.

[0075] (2) Dynamic layer adjustment: The signal FFT is calculated in every 1 ms window. If the energy proportion of the 20–30 MHz frequency band is ≥15%, the decomposition is switched to 4 layers. If the proportion is less than 10% for 5 consecutive windows, the decomposition is restored to 3 layers.

[0076] (3) Energy feature extraction: Calculate the energy value E of each sub-band i =∑|W i (t)| 2 And energy ratio With (E i ,P i ) to construct the feature vector.

[0077] It is worth explaining that, by definition: the input data of RHF is the full-band energy of the original signal after FFT, which is used to dynamically judge the intensity of high-frequency interference and trigger the adjustment of the number of wavelet packet decomposition layers (for example, from 3 layers to 4 layers). The input data of Pi is the energy value Ei of each sub-band after dynamic wavelet packet decomposition, which together with Ei constitutes a feature vector, which is used for adaptive hierarchical clustering to distinguish between discharge clusters and noise clusters. Logically related: RHF is a dynamic control parameter of the front end, which directly affects the way the sub-bands are divided, ensuring that the number of decomposition layers adapts to the real-time changes of high-frequency interference. Pi is a characteristic analysis parameter of the back end, which is used to accurately identify noise and signals in the decomposed sub-bands.

[0078] 3. Adaptive hierarchical clustering interference identification

[0079] (1) Hybrid distance metric: For sub-bands <10 MHz, the weighted Euclidean distance is used as follows:

[0080]

[0081] For sub-bands ≥ 10MHz, the cosine similarity is used as follows:

[0082]

[0083] (2) Dynamic threshold adjustment: initial threshold δ base =0.1×max(D), according to the noise cluster ratio R noise Corrected to δ = δ base ×(1+0.5R noise ).

[0084] (3) Interference determination: The clustering results are divided into high-energy clusters (energy proportion > 85%) and low-energy clusters (≤ 15%), and the latter are filtered out in the corresponding frequency band.

[0085] 4. Signal reconstruction and post-processing

[0086] (1) Inverse wavelet packet transform: Only retain the coefficients of high-energy cluster nodes, perform inverse transform to reconstruct the signal, and verify that the energy deviation is <5%.

[0087] (2) Time-frequency joint filtering:

[0088] Frequency domain zeroing: Set the spectral coefficients of low-energy cluster nodes to zero;

[0089] Time domain smoothing: Select the sliding window length according to the local SNR (5 points for SNR < 10 dB, 10 points for SNR ≥ 10 dB);

[0090] (3) Pulse enhancement: Use y(t) = sign(x(t))·|x(t)| 1.5 Amplify the pulse amplitude and output the final denoised signal.

[0091] 5. Embedded deployment and real-time optimization

[0092] (1) Hardware configuration:

[0093] Processor: ARM Cortex-M7 (main frequency 300 MHz), FPGA accelerates FFT and wavelet packet operations;

[0094] Memory allocation: Dynamic memory pool of 50 KB (stores feature vectors and model parameters).

[0095] (2) Real-time guarantee:

[0096] Segment the signal for processing (each 1 ms window), perform parallel computing for FFT and clustering analysis; Pre-generate the Db4 wavelet filter bank to reduce the real-time computing amount.

[0097] 6. Verification and output

[0098] Simulation verification: Inject Gaussian noise (SNR = 8 dB) and 28 MHz interference in MATLAB, the SNR is increased to 24.5 dB, and the pulse detection rate is 96%.

[0099] The embodiment of the present application also provides a partial discharge signal denoising system for UHV transformers, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the above method are implemented. The partial discharge signal denoising system for UHV transformers can implement the various embodiments of the above method and can achieve the same beneficial effects, which will not be elaborated here.

[0100] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations based on the concept of the present invention without creative efforts. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field based on the concept of the present invention through logical analysis, reasoning, or limited experiments on the basis of the prior art shall fall within the protection scope determined by the claims.

Claims

1. A denoising method for partial discharge signals of UHV transformers, characterized in that, Including: S1: Collect the partial discharge signals of the transformer and preprocess the electrical signals; S2: Perform dynamic wavelet packet decomposition on the preprocessed electrical signals, and extract the sub-band energy characteristics of the results of the dynamic wavelet packet decomposition; S3: Perform adaptive hierarchical clustering based on the extraction results; According to the hierarchical clustering results, divide the sub-bands into high-energy discharge clusters and low-energy noise clusters, and filter out the frequency bands corresponding to the low-energy noise clusters; S4: Reconstruct and optimize based on the wavelet packet nodes corresponding to the high-energy discharge clusters to obtain the finally output denoised signal.

2. The partial discharge signal denoising method for the UHV transformer according to claim 1, characterized in that The S1 includes: Collect the original partial discharge signals of the transformer through UHF sensors, and perform baseline correction, normalization processing, and power frequency filtering on the original partial discharge signals.

3. The method for denoising partial discharge signals of an UHV transformer according to claim 1, wherein, The S2 includes: Configure the basic decomposition parameters; Perform fast Fourier transform on the preprocessed electrical signals, calculate the signal FFT with each 1 ms as a calculation window, and calculate the energy ratio of the 20–30 MHz frequency band, satisfying the following relational expression: where X(f) represents the complex component of the signal at frequency f, and |X(f)| 2 represents the energy at frequency f; If R HF ≥ 15%, it is determined that there is significant high-frequency interference, and the decomposition layer is triggered to increase from the default 3 layers to 4 layers; If R is detected in 5 consecutive analysis windows HF ≤ 10%, then it will return to 3 layers; Calculate the energy value E of each sub - band i = ∑|W i (t)| 2 and the energy proportion Construct a feature vector with (E i , P i ), where W i (t) represents the coefficient sequence of the i - th sub - band in the time domain obtained by dynamic wavelet packet decomposition, and E k represents the total energy of the k - th sub - band.

4. The method for denoising partial discharge signals of an UHV transformer according to claim 3, characterized in that, During the process of performing dynamic wavelet packet decomposition on the preprocessed electrical signals, the method further includes: Dynamically adjust the process of dynamic wavelet packet decomposition, specifically including: Calculate the frequency band width, satisfying the following relational expression: where f s represents the sampling frequency of the signal; When the number of layers n = 3, the sub - band width is 6.25 MHz, and the computational complexity is O(Nlog N ); when the number of layers increases to n = 4, the sub - band width is reduced to 3.125 MHz, the computational complexity increases by 30%, and the control of the high - frequency interference suppression rate is increased from 68% to 92%; Adopt sliding window processing and pre-generated filter bank optimization, combined with FPGA parallel acceleration, to reduce the single processing time from 3.8 ms for a fixed 4 layers to 3.1 ms for dynamic adjustment, and the memory occupancy remains <50 KB.

5. The method for denoising partial discharge signals of an UHV transformer according to claim 1, wherein, The S3 includes: Using the energy value E of each sub - band i and its energy proportion P i as the feature vector to characterize the frequency - band energy distribution characteristics and construct the feature vector; Calculate the weighted Euclidean distance for the low-frequency sub-bands less than 10 MHz, satisfying the following relational expression: where E j represents the total energy of the j-th sub-band obtained by wavelet packet decomposition, and P j represents the proportion of the energy of the j-th sub-band in the total energy of all sub-bands; Calculate the cosine similarity for the high-frequency sub-bands exceeding 10 MHz, satisfying the following relational expression: Where, W i is the coefficient vector of the i-th sub-band, and W j is the coefficient vector of the j-th sub-band; Based on the proportion R of low-energy cluster nodes noise Automatically adjust the inter-cluster distance threshold to satisfy the following relationship: δ=0.1×max(D)×(1 + 0.5×R noise ); When the noise ratio increases, increase the threshold; Divide the interference into high-energy discharge clusters and low-energy discharge clusters according to the inter-cluster distance threshold.

6. The method for denoising partial discharge signals of an UHV transformer according to claim 1, wherein, The S4 includes: Perform wavelet packet inverse transform to reconstruct the signal for the high-energy discharge clusters: Synthesize layer by layer from the nth layer to the 0th layer of the dynamically adjusted decomposition layer, and perform upsampling and signal superposition through a two-channel filter bank for each layer to ensure that the following reconstruction formula is satisfied: where x rec represents the reconstructed denoised signal, represents the wavelet packet coefficient of the k-th sub-band after the n-th layer of wavelet packet decomposition, represents the reconstruction filter corresponding to the k-th sub-band; Verify the fidelity through energy conservation verification and spectrum similarity comparison; Adopt time-frequency joint filtering optimization after reconstruction: (1) Frequency domain mask processing: Set the spectrum coefficients of the low-energy noise clusters determined by clustering to zero; (2) Time domain adaptive smoothing: Dynamically select the sliding average window length according to the local signal-to-noise ratio, use a 5-point window to suppress high-frequency glitches in the low signal-to-noise ratio area, and use a 10-point window to retain pulse details in the high signal-to-noise ratio area; (3) Pulse enhancement: Apply the non-linear amplification function y(t) = sign(x(t))·|x(t)| 1.5 , to enhance the amplitude of the discharge pulse in the time domain, where x(t) represents the intermediate signal after the time-frequency joint filtering optimization process.

7. A partial discharge signal denoising system for an ultra-high voltage transformer, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of any one of the above claims 1 to 6.

Images